{"id":114,"date":"2026-07-30T19:32:57","date_gmt":"2026-07-30T19:32:57","guid":{"rendered":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/?post_type=part&#038;p=114"},"modified":"2026-09-22T18:27:01","modified_gmt":"2026-09-22T18:27:01","slug":"6-ethics-bias-and-critical-evaluation","status":"publish","type":"part","link":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/part\/6-ethics-bias-and-critical-evaluation\/","title":{"rendered":"6. Ethical Considerations of AI Use in Higher Education"},"content":{"raw":"Jorge Gatica, Mandi Goodsett, Xiongyi Liu, Emily Rauschert, Patrick Wachira\r\n<h2>INTRODUCTION<\/h2>\r\nGenerative AI has been adopted in many industries, and higher education faces increasing pressure to provide training and support for students to use this new technology in the workplace. However, while hype narratives often present this technology as inevitable and necessary, generative AI raises many ethical concerns that deserve careful consideration before it is used extensively. College campuses are ideal places to bring critical thinking, ethical frameworks, and evidence-based viewpoints to the development and use of this new technology, and students can practice critical thinking skills and learn more about AI through examining ethical concerns.\r\n\r\nHigher education institutions are, in some ways, directly threatened by the ethical issues that generative AI presents, especially when it comes to academic integrity and job replacement. College students\u200b are often aware of the ethical issues that generative AI presents, and <a href=\"https:\/\/sites.campbell.edu\/academictechnology\/2025\/03\/06\/ai-in-higher-education-a-summary-of-recent-surveys-of-students-and-faculty\/\">they express concern about AI\u2019s impact<\/a> on their ability to get jobs, on the environment, and on creativity more broadly. At the same time, <a href=\"https:\/\/www.insidehighered.com\/news\/students\/academics\/2025\/08\/29\/survey-college-students-views-ai\">surveys show<\/a> that college students continue to use generative AI for schoolwork in large numbers, sometimes in ways that undermine their learning. To remain viable, and to graduate students who can be responsible AI users, higher ed institutions must critically examine and effectively communicate the ethical threats that generative AI poses.\r\n<h2>FOUNDATIONS<\/h2>\r\n<h3><strong>Bias in AI in Higher Education<\/strong><\/h3>\r\n<h4>Defining Bias in AI Systems<\/h4>\r\nBefore surveying how bias manifests across higher education, it is worth considering what \"bias\" actually means in this context, since the term is used loosely both in popular discourse and in some of the scholarly literature. In machine learning, statistical bias refers to a systematic deviation between a model's predictions and the true values it is trying to estimate \u2014 a technical property of an estimator that can exist even in a perfectly \"fair\" system. Social bias, by contrast, refers to the systematic disadvantage or unfair treatment of particular groups, typically along lines of race, gender, disability, language background, or socioeconomic status. The two are related but distinct: a model can be statistically unbiased in aggregate while still producing socially biased outcomes for specific subgroups, and this distinction is important for understanding how institutions diagnose and respond to problems when they arise.\r\n\r\nIt is also useful to distinguish bias from simple errors. Random error \u2014 noise that affects all groups roughly equally \u2014 is a nuisance but not, on its own, an equity problem. Bias is what remains when error is not random: when a system is reliably worse for some groups than others. When considering AI ethics, it is critical to be aware of this, as several of the tools discussed here do not merely make mistakes \u2014 they make patterned, predictable mistakes that fall more heavily on some students than others.\r\n<h4>Causes: Where Bias Enters the Pipeline<\/h4>\r\nBias is rarely introduced at a single point. It typically accumulates across a pipeline, and understanding where it originates matters because different causes call for different remedies.\r\n<ul>\r\n \t<li><strong>Data-level causes.<\/strong> Most AI systems used in higher education are trained on historical institutional data \u2014 past admissions decisions, past grades, past enrollment and retention patterns. Because that history reflects real-world inequities (e.g., in access, in resourcing, in whose success the institution was set up to support), models trained on it tend to reproduce those inequities rather than correct for them. Underrepresentation compounds this: when a subgroup is thin in the training data, the model has less signal to learn from for that group, and its predictions for that group become noisier and less reliable.<\/li>\r\n \t<li><strong>Design-level causes.<\/strong> Even with balanced data, design choices can introduce bias. Developers must choose which variables (features) to include, and even when protected characteristics like race or gender are explicitly excluded, other variables can act as <em>proxies<\/em> \u2014 zip code standing in for race, first-generation status standing in for class background \u2014 carrying much of the same discriminatory signal under a different name. Models are also typically optimized to maximize average accuracy across the entire dataset, which can mean sacrificing accuracy for smaller subgroups in exchange for better overall performance.<\/li>\r\n \t<li><strong>Human and institutional causes.<\/strong> Someone decides what counts as the \"correct\" label during training \u2014 what counts as a \"successful\" student, a \"well-written\" essay, \"suspicious\" exam behavior \u2014 and those judgments encode the assumptions, and sometimes the unconscious biases, of the people making them.<\/li>\r\n \t<li><strong>Deployment-context causes.<\/strong> A model validated on one population can perform very differently when applied to a different one. A tool built and tested on a large, well-resourced research university's data may behave unpredictably at an under-resourced regional or community college with a different student population \u2014 yet vendor tools are frequently marketed and deployed across very different institutional contexts with little re-validation.<\/li>\r\n \t<li><strong>Feedback loops.<\/strong> Perhaps the most insidious cause is not a one-time input error but a self-reinforcing cycle: a biased prediction leads to a biased intervention (or lack of one), which shapes the student's subsequent outcomes, which then becomes new training data that confirms the model's original \u2014 biased \u2014 prediction.<\/li>\r\n \t<li><strong>Unequal access as an upstream cause.<\/strong> A related but distinct cause, discussed in more depth below, is that students and faculty do not begin with equal access to AI tools in the first place. Because that unequal access shapes the very outcomes \u2014 grades, retention, engagement \u2014 that later become training data for the predictive systems discussed throughout this section, today's access gap can become tomorrow's algorithmic bias.<\/li>\r\n<\/ul>\r\n<h4>Types of Bias<\/h4>\r\nBuilding on these causes, the technical literature identifies several recurring <em>types<\/em> of bias relevant to educational AI: <strong>representation bias<\/strong> (some groups are simply underrepresented in the data used to build the system); <strong>measurement bias<\/strong> (the proxies used to measure a construct, such as \"engagement\" or \"risk,\" do not mean the same thing across groups); <strong>aggregation bias<\/strong> (a single model is applied uniformly to a heterogeneous population when subgroup-specific models would perform better); <strong>evaluation bias<\/strong> (the benchmarks used to validate a model's fairness are themselves unrepresentative); and <strong>historical bias<\/strong> (the model faithfully learns a pattern from the past that was itself unjust). These categories recur throughout the use cases surveyed below.\r\n<h3>Survey of Use Cases<\/h3>\r\n<h4>Student-Facing and Administrative Use Cases<\/h4>\r\n<strong>Admissions algorithms.<\/strong> Several institutions now use predictive models to screen or rank applicants, estimating the likelihood that an applicant will enroll, succeed, or persist. Because these models are trained on historical admissions and outcome data, they risk encoding the same racial, socioeconomic, and geographic patterns that shaped who was admitted \u2014 and who succeeded \u2014 in the past, effectively automating and obscuring decisions that would draw far more scrutiny if made explicitly by a human admissions officer.\r\n\r\n<strong>Automated essay scoring and writing assessment.<\/strong> Automated scoring systems, increasingly used for placement testing and large-enrollment writing courses, have been shown to systematically disadvantage certain groups of writers. Research on automated scoring of English-language learners has documented what researchers term <em>bias amplification<\/em>: because high-scoring responses from English Language Learners are comparatively rare in training data, models trained through standard methods tend to favor linguistic patterns typical of non-ELL writers, and as a result systematically under-predict scores for ELL students even when their responses demonstrate comparable underlying knowledge (Wang et al., 2026). Critically, this study found that the resulting prediction gap between groups can be <em>larger<\/em> than the gap already present in the training data \u2014 the model does not merely inherit the disparity; it magnifies it.\r\n\r\n<strong>AI-text detection and plagiarism tools.<\/strong> Perhaps the most widely reported bias finding in this space concerns AI-generated-text detectors. In an influential Stanford study, detectors correctly classified essays written by native English-speaking eighth graders with near-perfect accuracy but incorrectly flagged more than half of TOEFL essays written by non-native English speakers as AI-generated (Liang, Yuksekgonul, Mao, Wu, &amp; Zou, 2023). The likely mechanism is that many detectors rely on \"perplexity\" \u2014 a measure of how predictable a text's word choices are \u2014 and non-native writing tends to be less lexically varied, which these tools mistake for the low-perplexity, formulaic patterns characteristic of AI output. Subsequent reporting found the same pattern held for widely used commercial tools deployed at scale in higher education, raising the possibility that international and multilingual students face disproportionate risk of false accusations of academic dishonesty (Garc\u00eda Mathewson, 2023).\r\n\r\n<strong>Learning analytics and at-risk prediction.<\/strong> These systems flag students believed to be at risk of failing or withdrawing so that institutions can direct advising or support resources toward them. AI systems predict student risk based on historical data, which often reflects and replicates systemic biases like lower graduation rates for marginalized or low-income groups. As a result, the algorithm can unfairly tag these students as \"high risk,\" leading to harmful stereotyping, discouraged students, or the misallocation of support resources. Ultimately, this turns past inequalities into self-fulfilling prophecies instead of objective academic measures. A real-world example can be found in the case study in this chapter.\r\n\r\n<strong>Chatbots and advising tools.<\/strong> As AI chatbots are adopted for academic advising and student services, early evidence suggests the quality, tone, and accuracy of responses can vary depending on how a student's question is phrased or how the underlying model perceives the student's background \u2014 an area that remains comparatively understudied relative to other use cases but merits ongoing attention as adoption expands.\r\n\r\n<strong>Proctoring software.<\/strong> Automated remote-proctoring tools that use facial detection to verify student identity and flag \"suspicious\" behavior during exams have produced some of the most direct and well-replicated evidence of racial bias in educational AI. A 2022 study analyzing a major proctoring platform found that students with darker skin tones, and Black students specifically, were significantly more likely to be flagged for instructor review due to lower facial-detection rates \u2014 and, examining the intersection of race and gender, that women with the darkest skin tones were flagged far more often than any other group (Yoder-Himes et al., 2022). This mirrors earlier, foundational work by Buolamwini and Gebru (2018) showing that commercial facial-recognition systems were most accurate for lighter-skinned men and least accurate for darker-skinned women. Beyond skin tone, disability advocates have documented that movement-flagging features in proctoring software routinely misclassify the involuntary movements of neurodivergent students, or interruptions from caregiving responsibilities, as evidence of cheating (Surveillance and Disability in Online Proctored Exams, 2025).\r\n<h4>Academic Writing Process<\/h4>\r\nBeyond assessment, AI tools are now embedded directly in how students and scholars <em>produce<\/em> academic writing, which introduces a related but distinct set of concerns:\r\n<ul>\r\n \t<li><strong>Grammar and writing assistants<\/strong> (Grammarly, Word's AI features, ChatGPT-based editors) are trained predominantly on standard academic English and can flag dialect features and translanguaging practices \u2014 for instance, elements of African American Vernacular English \u2014 as errors, effectively nudging all writers toward a single normative register regardless of rhetorical intent or cultural context.<\/li>\r\n \t<li><strong>AI-generated writing feedback in composition courses<\/strong> raises a related concern: automated feedback tools may work best for students already fluent in dominant academic conventions, potentially widening rather than narrowing the gap for multilingual and first-generation writers the tools are often marketed as helping.<\/li>\r\n \t<li><strong>Voice and style homogenization<\/strong> is a broader worry as AI drafting and editing tools become routine: heavy reliance on these tools risks flattening diverse rhetorical traditions \u2014 including non-Western argumentative structures \u2014 into a single expected shape. (A related but distinct form of homogenization, affecting the <em>substance<\/em> of research findings rather than writing style, is discussed separately.)<\/li>\r\n \t<li><strong>Access disparities<\/strong> in premium AI writing tools add a further, more mundane inequity: students and institutions with resources to pay for higher-tier tools gain an assistance advantage that is not available to everyone, layering a new gap on top of existing ones.<\/li>\r\n<\/ul>\r\n<h4>Academic Research Process<\/h4>\r\nA parallel set of concerns arises once AI tools move from supporting students' coursework to supporting scholars' own research process:\r\n<ul>\r\n \t<li><strong>AI literature-review and discovery tools<\/strong> (e.g., AI-powered search and summarization tools built on academic databases) tend to be biased toward English-language, Global North, and highly cited sources, systematically underrepresenting scholarship published outside those channels. While this was already a problem with research tools before AI, AI tools can exacerbate it.<\/li>\r\n \t<li><strong>Citation recommendation systems<\/strong> risk reinforcing existing citation bias, in which already-prominent scholars and institutions receive disproportionate visibility \u2014 a dynamic with documented gender and racial dimensions in the pre-AI literature that automated recommendation could easily entrench further.<\/li>\r\n \t<li><strong>AI-assisted peer review<\/strong>, an emerging practice in which reviewers or editors use AI tools to screen or summarize submissions, carries a distinct risk: non-native English phrasing could be misread by these tools as an indicator of lower quality, and unconventional methodologies or topics underrepresented in training data could be undervalued relative to mainstream approaches.<\/li>\r\n \t<li><strong>Research summarization and \"AI research assistant\" tools<\/strong> raise a subtler concern: what a tool treats as an \"authoritative\" source shapes what a novice researcher \u2014 particularly a graduate student new to a field \u2014 encounters first, potentially steering them toward an already-mainstream evidence base rather than genuinely comprehensive coverage.<\/li>\r\n \t<li><strong>Homogenization and the flattening of dissenting or extreme findings.<\/strong> A distinct and less-discussed risk runs in the opposite direction from the biases above: rather than favoring one perspective over another, AI tools trained and tuned to sound balanced, measured, and broadly acceptable can systematically smooth over genuine variation \u2014 including legitimate scientific disagreement, minority findings, and outlier results that a field may most need to reckon with. This tendency, termed <em>homogenization bias<\/em> in a recent synthesis spanning linguistics, psychology, and cognitive science, arises because models are optimized during training to favor patterns that are frequent and easily generalizable, which has the effect of smoothing over minority representations in their outputs (Cell Press\/Trends in Cognitive Sciences, 2026). Importantly, the same synthesis cautions that this narrowing does not converge on a genuinely neutral center \u2014 it converges on a center shaped by whichever populations and viewpoints are best represented in training data, meaning \"balanced-sounding\" AI output is not actually neutral, only differently biased. This has direct consequences for research synthesis specifically: a controlled study of LLM-generated summaries of scientific abstracts found a consistent pattern of overgeneralization, in which models broadened and flattened the scope of specific findings even when explicitly prompted to produce faithful, detailed summaries (PMC, 2025). A related account of AI-assisted science makes the underlying mechanism explicit: literature-summarization systems can omit negative or dissenting findings altogether and present genuinely mixed evidence as more settled and consistent than it actually is (Si et al., 2024; Chauhan, 2026). For a field that depends on preserving and engaging with disagreement \u2014 rather than averaging it away \u2014 this is a bias with few precedents in the pre-AI literature, and one that current fairness frameworks, built primarily to detect disadvantage to specific demographic groups, are not well equipped to catch.<\/li>\r\n<\/ul>\r\n<h4>AI in Data Generation and Analysis<\/h4>\r\nThis is perhaps the least examined \u2014 and most consequential \u2014 area of bias, since it touches the empirical core of the research process itself, upstream of any conclusion a scholar might draw.\r\n\r\n<strong>Synthetic data generation.<\/strong> AI tools are increasingly used to generate synthetic datasets, whether to protect privacy, augment small samples, or simulate populations that are hard to reach. The central risk here is what researchers call <em>bias inheritance<\/em>: because large language models reflect the biases present in their own training data, synthetic data generated by these models can propagate and amplify those biases in ways that materially affect the fairness of anything subsequently trained or analyzed using that data (Li et al., 2025). This is not merely a theoretical concern. A study generating 140,000 synthetic health records across seven different large language models found that larger, more capable models actually exhibited <em>greater<\/em> demographic bias than smaller ones \u2014 some systematically over-representing White or Black patients while under-representing Hispanic and Asian patients relative to real population data (Huang et al., 2025). This is a genuinely counterintuitive finding worth foregrounding: a \"better\" model, by ordinary performance benchmarks, is not automatically a fairer one.\r\n\r\n<strong>Synthetic survey respondents (\"silicon sampling\").<\/strong> A newer and fast-growing practice involves prompting large language models to simulate the responses a person with a given demographic profile would give to a survey \u2014 intended as a faster, cheaper substitute for fielding real surveys. Researchers studying this practice caution that LLM outputs in this context are model-dependent artifacts rather than measurements of the world, and that treating them as real observations risks systematically misrepresenting marginalized groups \u2014 a concern that is especially serious for subgroup analysis, since model error can concentrate unevenly across demographic groups even when overall agreement with real survey data looks acceptable (Bisbee, Imai, Rosenman, &amp; Zhou, 2024). Large-scale empirical audits bear this out: one study comparing LLM-simulated respondents to a real national arts-participation survey found the synthetic respondents showed a systematic positive bias \u2014 expressing inflated \"liking\" \u2014 relative to actual human respondents (Karell, 2026). As one review summarized the underlying problem bluntly, language models do not represent a genuine sample of any real population; they represent patterns in training data that itself systematically underrepresents marginalized groups, non-English speakers, and offline populations (Verian Group, 2026).\r\n\r\n<strong>AI-assisted qualitative coding and thematic analysis.<\/strong> Tools that use large language models to code interview transcripts or open-ended survey responses are being adopted rapidly for their efficiency gains, but the evidence on their reliability is mixed. One frequently cited study using ChatGPT and Llama 2 to code semantically complex interview data found systematic bias in the resulting outputs, with the potential to mislead subsequent interpretation, and recommended that AI coding be used to extend \u2014 never replace \u2014 analysis that begins with human researchers (Ashwin, Chhabra, &amp; Rao, 2023). A more applied account is specific about <em>how<\/em> this bias surfaces in practice: AI coding tools lack the contextual understanding needed for interpretive analysis, especially where the data involves humor, sarcasm, or coded language reflecting the cultural context or lived experience of marginalized participants \u2014 content that is easily misread by a system without that lived context (Child Trends, 2025).\r\n\r\n<strong>AI-assisted statistical analysis and coding tools.<\/strong> AI systems that select statistical models, generate analysis code, or interpret results are becoming common in research workflows, and their errors have a distinctive property: they cascade. A validation study testing ChatGPT's code-interpreter feature against real health data found that when the tool selected an incorrect statistical method, the errors it introduced propagated through every subsequent step of the analysis; conversely, when method selection was correct, the remaining steps were usually completed accurately (Ruta, Gaidici, Irwin, &amp; Lifshitz, 2025). This makes method selection the single highest-leverage point in the pipeline for bias or error to take hold \u2014 a useful concrete illustration of the \"upstream\" framing introduced in Section 1. A separate, more basic concern compounds this: empirical testing of ChatGPT's code generation found that it is non-deterministic, capable of returning meaningfully different code for the exact same prompt, which undermines both the reliability of any single analysis and the reproducibility of research that depends on it (Ouyang, Wang, Li, &amp; Zhang, 2024). Finally, an equity dimension is worth naming explicitly: a study of introductory-lab students using AI code-interpretation tools found a real risk that less experienced users would be unable to identify errors in the tool's output or know how to correct them \u2014 meaning the researchers and students least equipped to catch an AI analysis tool's mistakes are often the least statistically sophisticated, which is itself an unequal exposure to the risks these tools introduce (Low &amp; Kalender, 2023).\r\n<h4>AI-Generated Instructional Materials and Assessment<\/h4>\r\n<strong>Instructional materials.<\/strong> The clearest evidence here concerns AI-generated images used in course materials. A systematic review of 31 peer-reviewed studies on AI text-to-image tools in educational settings found that biased representation was pervasive across the literature: generated images disproportionately centered white, male, Western, thin, and non-disabled figures, while diversity in age, body type, and disability was largely absent (systematic review, <em>Computers and Education Open<\/em>, 2026 \u2014 author names could not be confirmed via available search tools and should be verified directly from the journal before citing). This is not merely an abstract pattern \u2014 a study using AI-generated illustrations drawn from a children's book excerpt in focus groups with Latine undergraduates found that students actively noticed these representational cues, negotiating the tension between the stereotypes embedded in the images and their own lived experience rather than passively accepting the images as neutral (Rho &amp; Karumbaiah, 2026). Beyond images, a broader taxonomy developed for instructors names the range of biases text-generation tools can introduce into course materials, including cultural, demographic, ideological, and linguistic bias, noting that English and a handful of other languages dominate online training data and so are disproportionately well-represented in what generative AI tools produce \u2014 a concrete illustration of the \"uneven quality across languages\" concern this subsection raises for multilingual campuses (University of Kansas Center for Teaching Excellence, n.d.). Encouragingly, at least one study measuring student and pre-service teacher awareness found that a majority of respondents already recognized that generative AI tools can reproduce gender and cultural stereotypes, suggesting critical AI literacy instruction has a meaningful foundation to build on (Ribes-Lafoz, Navarro-Colorado, &amp; Rovira-Collado, 2026).\r\n\r\n<strong>AI-generated assessment.<\/strong> Automatically generated exam questions raise a parallel but distinct concern, rooted in the psychometric concept of <em>differential item functioning<\/em> (DIF) \u2014 a pattern in which students from different groups who have the same underlying ability nonetheless have different probabilities of answering an item correctly, indicating the item itself, not the students' knowledge, is the source of the disparity. A student-centered study of an LLM-generated question set applied exactly this kind of analysis and found that some AI-generated exam items showed evidence of bias linked to unmeasured subgroup characteristics \u2014 such as differential familiarity with how a question happened to be worded \u2014 and concluded such items should be flagged for review before any high-stakes use (Student-Centered LLM Q&amp;A Study, 2025). The same study named the specific mechanism worth flagging for readers: AI-generated distractors (the incorrect answer options in multiple-choice items) can be culturally loaded, and item phrasing may inadvertently privilege students who are already comfortable with chatbot-style language over those who are not. A related methods paper aimed at instructors using AI to generate exam variants makes the same point more generally, noting that bias in AI-generated items can surface subtly \u2014 in the examples or cultural framing a tool defaults to \u2014 which is why the paper recommends pairing human-in-the-loop review with formal DIF analysis as a matter of course, rather than treating AI-generated items as ready to use as-is (Educational Sciences, 2025).\r\n<h4>Institutional Adoption and Procurement of AI Tools<\/h4>\r\nA different point of entry for bias, distinct from any single tool design, is the process by which institutions decide to adopt one AI tool over another \u2014 or adopt one at all. This matters because procurement decisions are frequently made under conditions that make it hard to catch bias before a tool reaches thousands of students: under time pressure, with limited technical expertise on the evaluating committee, and often without the leverage to demand the kind of transparency an independent audit requires.\r\n\r\n<strong>AI-detection tools as a case in point.<\/strong> The rollout of AI-text detection illustrates the problem concretely. When Turnitin added an AI-detection feature to its existing plagiarism-checking product in April 2023, the feature was enabled for existing institutional customers with less than 24 hours' notice, no option to disable it at the time, and no meaningful insight into how the underlying model worked (Vanderbilt University, 2023). In other words, thousands of institutions found themselves using a bias-prone detection tool not because they evaluated and selected it, but because it arrived bundled inside a product they had already procured for an unrelated purpose \u2014 a mode of adoption that bypasses the kind of vetting a standalone purchasing decision would normally require. Vanderbilt subsequently disabled the feature after months of testing and consultation, citing exactly these concerns (Vanderbilt University, 2023), and the University of Pittsburgh's teaching center reached a similar conclusion, stating explicitly that it had concluded current AI-detection software was not reliable enough to deploy without substantial risk of false positives, and disabled the tool campus-wide as a result (University of Pittsburgh Teaching Center, n.d.). Not every institution reached the same conclusion, however: an investigative analysis of purchasing records found institutions renewing AI-detection subscriptions year after year despite documented flaws in the technology and known privacy concerns about the vendor's growing database of student papers, suggesting that inertia and faculty demand for a bright line on academic integrity can outweigh the accumulating evidence of a tool's unreliability (CalMatters\/The Markup, 2025).\r\n\r\n<strong>Structural barriers to good procurement decisions.<\/strong> This unevenness across institutions is not simply a matter of some campuses caring more than others. A 2025 review of AI procurement practices across higher education found that technology purchasing today routinely involves IT, cybersecurity, legal, privacy, and academic stakeholders, and that the resulting review process \u2014 while thorough on security and compliance \u2014 often leaves equity and bias review as an afterthought relative to those better-established criteria (EDUCAUSE Review, 2025). The same review reported that procurement leaders most wanted external standards or frameworks to lean on, since few institutions have the in-house technical capacity to independently evaluate a vendor's bias-testing claims \u2014 a gap that tools like EDUCAUSE's Higher Education Community Vendor Assessment Tool (HECVAT) are beginning to address, though bias and fairness questions remain less standardized within it than security and privacy questions. This connects directly to the vendor-opacity pattern: an institution cannot demand evidence of fairness testing it does not know to ask for, and a vendor has limited incentive to volunteer testing that might complicate a sale.\r\n\r\n<strong>Accountability without eliminating bias.<\/strong> A further concern, raised directly by AI-governance researchers working with higher education leaders, is that automating part of a decision does not remove human bias from the process \u2014 it can instead remove the accountability that previously made a biased decision correctable. Framed as a caution for procurement committees themselves: adopting an AI tool to make an institutional process more \"objective\" can create the appearance of neutrality while quietly making it harder to identify who is responsible when the tool's output turns out to be biased after all (Changing Higher Ed, 2026).\r\n<h4>Inequality in AI Access<\/h4>\r\nA related but conceptually distinct concern is not how AI tools behave once in use, but who gets to use them at all. This is generally not bias in the technical sense but rather an access and resourcing gap that researchers have increasingly labeled the <em>AI divide<\/em>, explicitly building on the older concept of the digital divide (bioRxiv, 2024). The distinction matters for this section for a specific reason: unequal access today becomes a cause of algorithmic bias tomorrow, since the advantages or disadvantages it produces feed directly into the institutional outcome data (e.g., grades, retention, \"risk\" scores) that later trains the predictive models.\r\n\r\n<strong>Student access.<\/strong> The scale of disparity is substantial and measurable. A narrative review of recent literature found that only 27% of rural students had access to devices compatible with generative AI tools, compared to 70% of urban students, and that students with stronger digital competencies obtained up to 60% greater academic benefit from the same tools than students without those competencies (Springer, 2026a). This is not simply a matter of institutions failing to act: survey data reported by Inside Higher Ed found that half of chief technology officers said their institution does not grant students institutional access to generative AI tools at all \u2014 tools that, when institutionally licensed, are often free to the student and more capable and secure than whatever a student might access on their own (Inside Higher Ed, 2025). The same reporting noted that more than half of students said most or all of their instructors prohibit generative AI use outright, meaning that even students at well-resourced institutions may face inconsistent access depending on individual faculty policy rather than any institutional standard.\r\n\r\n<strong>Beyond access: a second-order divide in literacy and confidence.<\/strong> Even where access exists, researchers have found a second divide around who is equipped to use these tools effectively. A qualitative study of academic staff at Norwegian universities concluded that new digital divides could emerge not from unequal access to technology itself, but from unequal possession of the AI literacy, critical judgment, and prompting skill required to benefit from it \u2014 and that this divide appears not only between groups of students, but between entire academic disciplines (Springer, 2026b). A large single-institution survey similarly found that although the large majority of students were broadly familiar with generative AI concepts, only about a quarter were actually using these tools for academic work, and roughly three-quarters had received no formal classroom instruction on how to use them at all (DeStefano, Hackney, &amp; Moskal, 2026) \u2014 suggesting that awareness of a tool's existence is a poor proxy for a student's actual capacity to use it well.\r\n\r\n<strong>Faculty, not only students.<\/strong> The access gap extends to faculty as well, with direct consequences for the fairness of the tools students encounter. Faculty digital-access gaps have long been documented as a standing concern in educational technology generally (PMC, 2025), and if faculty themselves have uneven access to, or training in, AI tools, that unevenness shapes which students benefit from AI-assisted instruction, feedback, and advising in the first place \u2014 compounding, rather than independently sitting alongside, the student-facing access gap described above.\r\n\r\nA 2025 policy commentary makes the throughline explicit: without deliberate institutional and policy intervention, unequal access to generative AI risks concentrating educational advantage among students and institutions that already have the most resources, producing what the author terms a systemic divergence in educational outcomes rather than a marginal difference in classroom experience (Wong, 2025). This is precisely the mechanism described as a <em>feedback loop<\/em>: today's access gap shapes tomorrow's grades and retention data, which becomes the training data for the predictive systems \u2014 meaning that closing the access gap is not a separate equity initiative from addressing algorithmic bias, but one of the more effective upstream interventions available for preventing it.\r\n<h3>Case Study: Racial Bias in Community College Early-Alert Prediction<\/h3>\r\n<h4>Background<\/h4>\r\nPredictive analytics for identifying \"at-risk\" students has become one of the most widely adopted AI applications in American higher education, frequently implemented through partnerships with private vendors. Georgia State University's partnership with the vendor EAB is probably the most publicized example: the university reported a meaningful increase in its graduation rate after adopting predictive analytics to identify struggling students and direct advising resources toward them, a result that has been widely cited as a success story for the technology, though it coincided with several other institutional changes (Swaak, 2022, as cited in AIR, 2023). Many other institutions, including Temple University, have built similar \"early alert\" systems, most commonly used to flag students at risk of dropping out before completing their degree so that advisors can reach out proactively (Bird, Castleman, Mabel, &amp; Song, 2021).\r\n<h4>The Study<\/h4>\r\nEconomists Kelli Bird, Benjamin Castleman, and Yifeng Song conducted one of the most rigorous independent audits of this class of system, examining two prediction models used at the community-college level: one predicting course completion and one predicting degree completion \u2014 the two outcomes most commonly targeted by early-alert systems nationally (Bird, Castleman, &amp; Song, 2024).\r\n<h4>The Finding<\/h4>\r\nThe central finding is direct: if either model were used, as intended, to target additional advising or support resources toward \"at-risk\" students, the resulting algorithmic bias would mean fewer marginal Black students received those resources than a fair system would allocate to them (Bird, Castleman, &amp; Song, 2024). This is not simply a case of the model being \"less accurate\" for Black students in the abstract \u2014 it is a bias in the <em>allocation<\/em> of a scarce, real-world resource, occurring at exactly the decision threshold institutions use to determine who gets help.\r\n\r\nNotably, the bias was not evenly distributed across the risk distribution \u2014 it concentrated precisely where institutional decision-making tends to draw the line. With the degree-completion model, the magnitude of bias was found to be several times higher when institutions defined \"at-risk\" using only the bottom decile of predicted scores than when they used a broader bottom-half cutoff (Bird, Castleman, &amp; Song, 2023). This is a critical operational detail: colleges with limited advising capacity are pushed, for resource reasons, toward exactly the narrow cutoffs that maximize this disparity.\r\n<h4>Mechanism<\/h4>\r\nThe likely cause is not a maliciously chosen variable but a more structural asymmetry: the researchers' findings suggest algorithmic bias arises in part because the administrative data institutions already have on hand is less predictive of Black students' success than of White students' success, particularly for new students \u2014 implying that additional, more informative data collection could help narrow the gap (Bird, Castleman, &amp; Song, 2024).\r\n<h4>Structural Context and Mitigation<\/h4>\r\nThis case also illustrates the governance concerns raised throughout this section. Because most predictive analytics tools in higher education are proprietary, institutions adopting them typically have little visibility into how the underlying model was built or validated (Bird, Castleman, Mabel, &amp; Song, 2021), which means that most colleges using an early-alert system have no equivalent, independent bias audit of their own vendor's tool. The Bird, Castleman, and Song study is notable precisely because it was conducted by researchers with no stake in the tool's commercial success \u2014 a form of scrutiny that vendor's opacity otherwise forecloses. Without independent auditing and data-sharing requirements built into procurement, institutions have limited means of discovering whether a tool they have adopted in good faith is quietly reallocating support away from the students who need it most.\r\n<h3>Cross-Cutting Patterns of AI Biases in Higher Education<\/h3>\r\nAcross the use cases surveyed above, several patterns recur often enough to be treated as structural features of AI bias in higher education, rather than as isolated problems specific to any one tool.\r\n\r\n<strong>Proxy variables.<\/strong> Across admissions, learning analytics, and writing assessment alike, models rarely need an explicit protected variable to reproduce discrimination \u2014 geography, prior coursework, or linguistic patterns frequently function as effective proxies for race, class, or national origin.\r\n\r\n<strong>Feedback loops.<\/strong> The at-risk prediction case study illustrates a dynamic visible elsewhere as well: a biased prediction shapes a biased intervention (or its absence), which shapes real student outcomes, which then becomes the next round of training data \u2014 quietly hardening the original bias into something that looks, on the surface, like objective evidence.\r\n\r\n<strong>Opacity in vendor tools.<\/strong> From proctoring software to predictive analytics to AI-text detectors, the tools shaping high-stakes decisions in higher education are overwhelmingly proprietary. Institutions frequently cannot see the training data, validation methodology, or model architecture behind a tool they have licensed, which makes independent verification \u2014 of the kind performed in the case study above \u2014 the exception rather than the norm.\r\n\r\n<strong>Disproportionate impact on intersectional groups.<\/strong> AI bias in higher education disproportionately impacts on certain groups across various use cases: Black students in early-alert and proctoring systems; non-native English speakers and international students in AI-text detection and writing-assessment tools; first-generation and lower-income students in access to premium AI writing and research tools.\r\n\r\n<strong>Homogenization.<\/strong> Rather than disadvantaging one group relative to another, several tools discussed above (e.g., writing assistants, research-summarization tools) share a tendency to smooth over genuine variation altogether \u2014 in rhetorical style, in scientific findings, in perspective \u2014 converging on outputs that read as neutral but are shaped by whichever patterns are best represented in training data. This pattern is harder to detect with the subgroup-based auditing methods that address the other four patterns, since the harm is not unequal treatment of identifiable groups but a general narrowing of what gets represented or preserved at all.\r\n<h3>Institutional and Ethical Stakes<\/h3>\r\nThe stakes of these patterns are not only ethical but also legal and operational. In the United States, biased outcomes tied to race, national origin, or disability status can implicate Title VI of the Civil Rights Act and the Americans with Disabilities Act, and institutions that deploy opaque, vendor-controlled tools without independent validation face growing legal and regulatory exposure as scrutiny of these systems increases \u2014 a concern the U.S. Government Accountability Office has itself flagged with respect to predictive analytics in higher education (Bauman, 2022, as cited in AIR, 2023). At the same time, institutions face genuine pressure to adopt these tools: advising staff and writing-center resources are finite, and AI tools promise to extend limited institutional capacity across more students than human staff alone could serve. This creates an unresolved tension between equity and efficiency \u2014 the goal of using AI to serve more students can come into direct conflict with the goal of serving them fairly, particularly when speed of adoption outpaces the institution's capacity to audit what it has adopted. Consequently, this tension can play out in institutional procurement decisions.\r\n<h3>Mitigation Strategies<\/h3>\r\nWhile no single intervention addresses bias across every stage of the AI pipeline, we choose to summarize a few AI bias mitigation strategies that have the greatest potential in higher education.\r\n\r\n<strong>Technical approaches<\/strong> include fairness-aware machine learning methods designed to explicitly balance accuracy across subgroups rather than optimizing only for aggregate performance, and pre-procurement bias auditing, in which institutions require a vendor's tool to be tested for subgroup disparities before adoption rather than after a problem is reported. The community-college case study above illustrates the value of this kind of audit \u2014 and how rarely it happens absent independent researcher access.\r\n\r\n<strong>Policy and governance approaches<\/strong> include establishing institutional AI governance committees with the authority to review and approve (or decline) AI tools before deployment; contractual requirements for vendor transparency about training data and validation methods; and human-in-the-loop requirements that prevent any single AI output \u2014 an at-risk flag, a proctoring alert, a detector score \u2014 from triggering a consequential decision without a human reviewer able to exercise independent judgment.\r\n\r\n<strong>Pedagogical approaches<\/strong> address the problem from the other direction: building AI and data literacy among faculty and administrators so that they are equipped to question vendor claims, interpret model outputs critically, and recognize when a tool's assumptions do not fit their student population \u2014 competencies that are, at present, unevenly distributed across the institutions adopting these tools fastest.\r\n\r\n<strong>Access and infrastructure approaches<\/strong> address the upstream cause directly: institutional licensing that grants all students and faculty a baseline level of AI access, rather than leaving access to individual ability to pay; device and connectivity support for students without reliable access outside campus; and integrating AI literacy instruction into the curriculum itself rather than assuming students arrive with it.\r\n<h3>Open Questions and Future Directions<\/h3>\r\nSeveral questions remain open. First, generative AI-specific bias in AI tutors, writing assistants, and the data-generation and analysis tools is a newer and comparatively understudied frontier relative to the more established literature on predictive analytics and proctoring software, and is likely to grow more consequential as these tools become further embedded in both teaching and research workflows. Second, the same personalization that makes AI tools appealing \u2014 tailoring feedback, recommendations, or interventions to an individual student or researcher \u2014 is difficult to distinguish from stereotyping along demographic lines when the underlying model's \"personalization\" is itself learned from biased historical data. Finally, responsibility for addressing these harms remains contested and diffuse. It is not obvious whether the vendor building the tool, the institution deploying it, the instructor or advisor using its output, or the researcher relying on its analysis bears primary responsibility when bias causes harm \u2014 and until that question is resolved with greater institutional clarity, accountability is likely to continue falling through the gaps between these parties.\r\n<h3>AI and Environmental Impact<\/h3>\r\nAs generative AI models have exploded into daily life, public awareness has evolved from viewing AI as a clean, virtual \"cloud\" technology to recognizing it as a highly resource-intensive physical industry.\r\n\r\nThe prevailing public sentiment toward AI\u2019s environmental impact is one of worry rather than optimism.\r\n<h4>Fear of net harm<\/h4>\r\nThe public is twice as likely to believe that AI will ultimately do more to hurt the environment than help it. Strikingly, people express greater concern about AI\u2019s environmental consequences than about other notoriously carbon-heavy industries, such as air travel, cryptocurrency mining, and meat production (t\u00a0he AP-NORC at the University of Chicago, apnorc.org).\r\n\r\nCertainly, the environmental footprint of AI-related data centers is growing rapidly, inviting frequent comparisons to heavy industrial sectors. While both AI data centers and large petrochemical plants have massive environmental footprints, they operate on completely different paradigms: AI data centers primarily generate indirect (Scope 2) environmental pressure driven by massive electricity and water consumption (ONU <a href=\"https:\/\/unu.edu\/inweh\/collection\/environmental-cost-of-AIs-Enrgy-Use-Carbon-water-and-land-footprints\"><em>Environmental Cost of Artificial Intelligence)<\/em><\/a>. Petrochemical plants inflict direct (Scope 1) industrial degradation through the chemical processing of fossil fuels into plastics, resins, and synthetic materials (Yan et al., 2024)\r\n<h4>Carbon Footprint and Greenhouse Gas Emissions<\/h4>\r\nThe mechanism and scale of emissions differ significantly between the two sectors.\r\n<table class=\"grid aligncenter\" style=\"height: 307px;\">\r\n<thead>\r\n<tr style=\"height: 18px;\">\r\n<th style=\"height: 18px; width: 132.031px;\" scope=\"col\"><strong>Impact Category<\/strong><\/th>\r\n<th style=\"height: 18px; width: 395.766px;\" scope=\"col\"><strong>AI-Related Data Centers<\/strong><\/th>\r\n<th style=\"height: 18px; width: 264.922px;\" scope=\"col\"><strong>Large Petrochemical Plants<\/strong><\/th>\r\n<\/tr>\r\n<\/thead>\r\n<tbody>\r\n<tr style=\"height: 124px;\">\r\n<td style=\"height: 124px; width: 132.031px;\"><strong>Primary Emission Source<\/strong><\/td>\r\n<td style=\"height: 124px; width: 395.766px;\"><strong>Indirect (Scope 2)<\/strong>\r\n\r\nElectricity drawn from regional power grids to run chips and cooling infrastructure.<\/td>\r\n<td style=\"height: 124px; width: 264.922px;\"><strong>Direct (Scope 1)<\/strong>\r\n\r\nOn-site fossil fuel combustion, cracking furnaces, and chemical process venting.<\/td>\r\n<\/tr>\r\n<tr style=\"height: 92px;\">\r\n<td style=\"height: 92px; width: 132.031px;\"><strong>Global Sector Emissions<\/strong><\/td>\r\n<td style=\"height: 92px; width: 395.766px;\">Global data centers collectively emit <strong>between 190\u2013208 million metric tons of <\/strong>carbon dioxide (CO2) <strong>annually<\/strong> (AI tasks represe190 and 208 million metric tons of carbon dioxide (CO2) annually (AI tasks account for 0% of this total).<\/td>\r\n<td style=\"height: 92px; width: 264.922px;\">The chemical and petrochemical sectors worldwide emit roughly <strong>1.3 to 1.5 billion metric tons of <\/strong>carbon dioxide (CO2) <strong>annually<\/strong>.<\/td>\r\n<\/tr>\r\n<tr style=\"height: 73px;\">\r\n<td style=\"height: 73px; width: 132.031px;\"><strong>Decarbonization Path<\/strong><\/td>\r\n<td style=\"height: 73px; width: 395.766px;\"><strong>Highly adaptable:<\/strong> Can theoretically achieve net-zero if the regional electrical grid transitions entirely to renewable energy.<\/td>\r\n<td style=\"height: 73px; width: 264.922px;\"><strong>Hard to abate:<\/strong> Carbon is inherently baked into the raw materials (crude oil, natural gas) and high-heat chemical processes.<\/td>\r\n<\/tr>\r\n<\/tbody>\r\n<\/table>\r\n<strong>The Petrochemical Baseline:<\/strong> A single large petrochemical facility can emit several million tons of carbon dioxide (CO2) directly from its cracking stacks, making the baseline sector's carbon footprint significantly larger than that of the AI sector.\r\n\r\nWhen people express concern about AI, public anxiety usually zeroes in on the physical infrastructure required to keep the technology running:\r\n<ul>\r\n \t<li><strong>Strain on local power grids:<\/strong> Public awareness of massive data centers has grown, and these facilities demand large amounts of electricity, much of which is still generated by fossil fuels.<\/li>\r\n \t<li><strong>The AI Data Center Dilemma:<\/strong> While tech \u201chyperscalers\u201d contract heavily for green energy via Market-Based Carbon Certificates, their continuous 24\/7 power draw forces local grids to rely on fossil-fuel backup power. Projections indicate AI data center expansion will add up to 44 million metric tons of new carbon dioxide (CO2) annually by 2030.<\/li>\r\n \t<li><strong>Water depletion:<\/strong> Public pushback has grown in communities hosting these facilities, as people realize data centers require millions of gallons of water daily just to keep servers cool. Water supply may deplete municipal water resources when these centers are built in water insecure locations (cf. <a href=\"https:\/\/sf.tradepub.com\/?pt=adv&amp;page=Data%20Center%20World\">Data Center World).<\/a><\/li>\r\n \t<li><strong>Lack of government oversight:<\/strong> A significant majority of the public believes that the government and tech regulatory bodies are not doing enough to address or mitigate AI\u2019s environmental harms. State legislation and initiatives differ widely across the US (see below for OH references).<\/li>\r\n<\/ul>\r\n<h4>The \"Sustainability Paradox\"<\/h4>\r\nDespite these worries, public perception remains conflicted (cf. Parker, 2026). People acknowledge a \"sustainability paradox\": they are uncomfortable with the massive carbon and water footprints of training models, yet they remain hopeful that traditional machine learning can be used to model climate patterns, optimize renewable energy grids, and track deforestation.\r\n\r\nThe exact resource metrics consumed by a single AI prompt depend heavily on the complexity of the model (e.g., standard text vs. image generation), hardware efficiency, and data center cooling methods. Official data published by major AI providers\u00a0 and independent researchers break down the baseline metrics for a single text-based prompt:\r\n<h4>Core Resource Consumption Metrics<\/h4>\r\n<h4>(per AI user\u2019s prompt, O\u2019Donnell and Crownhart, 2025; Ritchie, 2025)<\/h4>\r\n<table class=\"grid aligncenter\">\r\n<thead>\r\n<tr>\r\n<th style=\"text-align: center;\" scope=\"col\"><strong>Metric<\/strong><\/th>\r\n<th style=\"text-align: center;\" scope=\"col\"><strong>Google Gemini (Median)<\/strong><\/th>\r\n<th style=\"text-align: center;\" scope=\"col\"><strong>OpenAI ChatGPT (Baseline)<\/strong><\/th>\r\n<th style=\"text-align: center;\" scope=\"col\"><strong>Independent\/Multi-Model Estimates<\/strong><\/th>\r\n<\/tr>\r\n<\/thead>\r\n<tbody>\r\n<tr>\r\n<td><strong>Electricity<\/strong><\/td>\r\n<td>0.24 watt-hours (Wh)<\/td>\r\n<td>0.34 watt-hours (Wh)<\/td>\r\n<td>0.3 to 0.5 watt-hours (Wh)<\/td>\r\n<\/tr>\r\n<tr>\r\n<td><strong>Water<\/strong><\/td>\r\n<td>0.26 milliliters (mL)<\/td>\r\n<td>0.32 milliliters (mL)<\/td>\r\n<td>10 to 40 milliliters (mL)<\/td>\r\n<\/tr>\r\n<tr>\r\n<td><strong>Carbon Emissions<\/strong><\/td>\r\n<td>0.03 grams of CO\u2082eq<\/td>\r\n<td><em>Not officially disclosed<\/em><\/td>\r\n<td>0.03 to 0.1 grams of CO\u2082eq<\/td>\r\n<\/tr>\r\n<\/tbody>\r\n<\/table>\r\n<h4>Metric Breakdown and Real-World Equivalents<\/h4>\r\n<h5>Electricity (0.24 to 0.5 Watt-hours)<\/h5>\r\n<strong>What it means:<\/strong> The actual physical computing power required by AI accelerators (like Nvidia GPUs or Google TPUs) to process your prompt. <strong>Real-world equivalent:<\/strong> Running a standard kitchen microwave for about <strong>one second<\/strong> or powering a highly efficient LED lightbulb for <strong>approximately two minutes<\/strong>.\r\n<h5>Water (0.26 to 40 Milliliters)<\/h5>\r\nThe sharp difference in water data comes down to methodology. Tech giants like Google and OpenAI disclose <em>direct operational consumption<\/em>\u2014the exact amount of clean water that evaporates on-site through cooling towers to prevent servers from overheating during that specific calculation. Independent researchers consider the total ecological footprint, factoring in the large volume of water used upstream by the power plants that generate the data center's electricity.\r\n\r\n<strong>Real-world equivalent:<\/strong> Low-end official estimates are equal to about <strong>5 drops of water<\/strong>. High-end independent tracking equates a short back-and-forth conversation (20 to 50 prompts) to about <strong>500 mL<\/strong>, or a standard plastic water bottle (Kelly et al., 2026).\r\n<h5>Carbon Emissions (0.03 to 0.1 grams of CO\u2082eq)<\/h5>\r\nIn environmental impact quantification, <strong>Carbon Dioxide Equivalents (expressed as CO\u2082e or CO\u2082eq)<\/strong> serve as the universal \"currency\" for measuring and comparing the climate impact of various greenhouse gases (GHGs). Instead of tracking multiple pollutants such as methane, nitrous oxide, and fluorinated gases, CO\u2082eq bundles them into a single, standardized metric.\r\n\r\nThe math behind CO\u2082eq relies on a multiplier factor called <strong>Global Warming Potential (GWP)<\/strong>. GWP measures how much heat a specific greenhouse gas traps in the atmosphere over a set period (usually a 100-year time horizon) relative to carbon dioxide (CO\u2082).\r\n<h5>Key Factors That Change these Metrics<\/h5>\r\n<ul>\r\n \t<li><strong>Task Complexity:<\/strong> Generating a high-resolution AI image or video requires vastly more math than text. Creating a single AI image consumes roughly <strong>2 to 5 liters of water<\/strong>, several times the footprint of a simple text interaction.<\/li>\r\n \t<li><strong>Model Size &amp; Architecture:<\/strong> Advanced \"reasoning\" models or deeply layered architectures (like DeepSeek R1) can consume up to <strong>30+ Wh per query<\/strong> because they perform billions more mathematical operations before spitting out an answer.<\/li>\r\n \t<li><strong>Geography and Climate:<\/strong> A prompt routed to a data center in a cool climate utilizing outdoor air cooling uses almost no local water. The same prompt routed to a data center in a hot, arid climate relying on evaporative cooling towers uses significantly more.<\/li>\r\n<\/ul>\r\n<h4>Corporate initiatives to achieve net-zero AI infrastructure<\/h4>\r\nTech giants are facing a major dilemma: the massive AI computing boom has sent greenhouse gas emissions soaring, throwing corporate sustainability timelines off track. Reports reveal recent year-over-year emission spikes of 25% at Google, 23% at Microsoft, and 64% at Meta.\r\n\r\nHere, \u201chyperscalers\u201d refers to massive technology companies that dominate the global cloud computing, data storage, and digital infrastructure industries. The term \"hyperscale\" refers to their ability to seamlessly and massively scale a computer network\u2014adding thousands of servers and massive amounts of storage\u2014to meet geometric increases in data demand.\r\n<h5>1. The Nuclear Power Pivot<\/h5>\r\nTech companies are rapidly moving beyond traditional solar and wind power to secure baseload clean electricity that runs 24\/7, leading a historic corporate shift toward nuclear energy:\r\n<ul>\r\n \t<li><strong>Microsoft:<\/strong> Signed a massive 20-year power purchase agreement to back the commercial restart of the 835-megawatt <a href=\"https:\/\/introl.com\/blog\/nuclear-power-ai-data-centers-microsoft-google-amazon-2025\">Three Mile Island Nuclear Plant<\/a>.<\/li>\r\n \t<li><strong>Google:<\/strong> Placed a historic order for a fleet of up to 500 megawatts of Small Modular Reactors (SMRs) from Kairos Power to power its data centers starting by the end of the decade.<\/li>\r\n \t<li><strong>Amazon:<\/strong> Invested over $20 billion to scale and convert the Susquehanna nuclear site into a completely nuclear-powered AI data center campus.<\/li>\r\n \t<li><strong>Meta:<\/strong> Teamed up with companies like Vistra, TerraPower, and Oklo to target up to 4 gigawatts of nuclear power for its <a href=\"https:\/\/www.cnbc.com\/2026\/01\/09\/meta-signs-nuclear-energy-deals-to-power-prometheus-ai-supercluster.html\">Prometheus AI Supercluster<\/a> in Ohio.<\/li>\r\n<\/ul>\r\n<h5>2. Gigawatt-Scale Battery Storage and 24\/7 Carbon-Free Energy<\/h5>\r\nBecause wind and solar are intermittent, tech companies are installing unprecedented levels of utility-scale energy storage. This allows them to capture clean energy during the day and discharge it when the grid is strained.\r\n<ul>\r\n \t<li><strong>The Goal:<\/strong> Companies are transitioning from \"annual carbon matching\" to true 24\/7 hourly matching, ensuring that every watt of power drawn by an AI accelerator is zero-carbon in real time.<\/li>\r\n \t<li><strong>The Scale:<\/strong> Massive utility-scale projects, such as Amazon's co-located 400 MWh Megapack storage facilities, have made the data center market the largest single driver of <a href=\"https:\/\/www.facebook.com\/chichubkidaps\/posts\/-americas-ai-data-center-boom-is-driving-unprecedented-battery-storage-deploymen\/1031704529384819\/\">grid battery storage deployment<\/a> in the US.<\/li>\r\n<\/ul>\r\n<h5>3. Chip and Hardware Efficiency Gains<\/h5>\r\nEngineers are trying to outpace the growing demands for computational power by significantly optimizing chip performance and life cycle.\r\n<ul>\r\n \t<li><strong>Custom AI Accelerators:<\/strong> Google\u2019s eighth-generation custom AI chips (TPU 8t and 8i) deliver up to a <a href=\"https:\/\/sustainability.google\/\">2x improvement in performance per watt<\/a> compared to previous iterations.<\/li>\r\n \t<li><strong>Lifecycle Extension:<\/strong> Hyperscalers are implementing aggressive hardware circularity initiatives. Extending GPU and server life cycles from 3 years to 5\u20137 years drastically lowers the \"embodied carbon\" generated during manufacturing and concrete-heavy facility construction.<\/li>\r\n<\/ul>\r\n<h5>4. Cross-Industry Tech Alliances<\/h5>\r\nRather than building isolated solutions, the tech sector is co-funding collaborative infrastructure innovations:\r\n<ul>\r\n \t<li><strong>Data Center Innovation Initiative:<\/strong> Amazon, Google, Meta, and Microsoft have jointly launched a funding coalition with groups like Breakthrough Energy. They are deploying up to $5 million per startup to test <a href=\"https:\/\/sustainabilitymag.com\/news\/the-four-tech-giants-funding-low-carbon-data-centre-startups\">sustainable data center innovations<\/a> inside active, working facilities.<\/li>\r\n \t<li><strong>Carbon Dioxide Removal (CDR):<\/strong> Tech companies are leveraging long-term advance market commitments to fund massive direct-air carbon capture startups. This private-sector backing helps commercialize carbon removal technologies to offset unavoidable Scope 3 supply chain emissions (<a href=\"https:\/\/carboncredits.com\/data-center-giants-enter-carbon-credit-market-as-hyperscalers-fuel-a-new-green-tech-gold-rush\/\">CarbonCredits.com<\/a>).<\/li>\r\n<\/ul>\r\n<h4>Curving the High Demand for Fresh Water or the Effects of Water Pollution<\/h4>\r\nThe rapid rise of generative AI has significantly increased the heat density within server racks. Traditional data center cooling relies on open-loop <strong>evaporative cooling towers<\/strong>. These systems reject heat by constantly evaporating millions of gallons of treated freshwater into the atmosphere every day.\r\n\r\nTo curb this massive environmental drain, the data center industry is undergoing a comprehensive infrastructure overhaul. They are shifting from cooling whole rooms of hot air to managing thermal loads directly at the chip level.\r\n<h5>1. The \"Hot Tub\" Fluid Breakthrough (Warm-Water Closed Loops)<\/h5>\r\nThe most significant industry shift involves raising the operating temperature of the liquid coolant itself. Nvidia's latest AI infrastructure architecture standardizes to <strong>100% closed-loop liquid cooling<\/strong>. The coolant (typically a mix of 75% water and 25% propylene glycol) enters the server rack at <strong>45\u00b0C (113\u00b0F)<\/strong>\u2014hotter than a residential hot tub\u2014and can absorb heat up to 55\u00b0C (131\u00b0F).\r\n\r\nBecause the coolant is so hot, it creates a massive temperature differential with the surroundings (outside air). This allows the heat to be dissipated into the environment using simple outdoor <strong>dry coolers<\/strong> (giant radiator coils) instead of evaporative chillers. The system is filled once during construction and continuously recirculated, cutting local facility cooling water usage from 2.6 million gallons per megawatt per year down to <strong>near zero<\/strong>.\r\n<h5>2. Direct-to-Chip (DTC) Cold Plate Architecture<\/h5>\r\nAs AI chips surpass 1,000+ watts each, pushing cold air across them is no longer physically viable. Direct-to-Chip (DTC) systems attach copper micro-channel <strong>cold plates<\/strong> directly to the physical surface of the GPU or TPU. The dielectric fluid or water mix flows through these microscopic channels, directly capturing up to 98% of the processor's heat at the exact source of emission.\r\n\r\nMajor tech hyperscalers are standardizing DTC designs for all new AI builds. By localizing the heat transfer, companies can bypass massive air-conditioning infrastructure entirely, allowing facilities to maintain ultra-low Water Usage Effectiveness (WUE) metrics without relying on municipal water loops.\r\n<h5>3. Two-Phase Immersion Cooling<\/h5>\r\nFor extreme AI rack densities (approaching 100kW+ per rack), immersion cooling eliminates the need for physical water plumbing inside the server. The entire server chassis is fully submerged in a bath of specially engineered, non-conductive <strong>dielectric fluid<\/strong>. In a <em>two-phase<\/em> system, the fluid has a low boiling point.\r\n\r\nAs the AI chips run hot, the fluid boils, vaporizes, rises to a condenser coil at the top of the sealed tank, condenses back into a liquid, and falls back down. This creates a completely hermetic, self-contained thermodynamic cycle. The external loop that cools the condenser coil can rely entirely on dry-air heat exchangers, resulting in zero evaporative water loss in the data center's mechanical operations.\r\n<h5>4. Eco-Chilling and Non-Potable Circularity<\/h5>\r\nWhere liquid cooling is not fully implemented, and air cooling must be augmented with water, tech giants are altering their water sourcing: Companies are investing heavily in on-site water reclamation plants. Instead of pulling clean drinking water from local aquifers, data centers use raw industrial wastewater or sewer lines. They use internal <strong>reverse osmosis filtration<\/strong> to purify the water to a level suitable for mechanical use, while keeping potable drinking water in the community.\r\n<h4>The Unseen Bottleneck: \"Indirect\" Footprints<\/h4>\r\nWhile these closed-loop facility designs effectively address on-site water issues, sustainability researchers point out a major loophole: <strong>Scope 2 indirect water consumption<\/strong>.\r\n\r\nIf a \"zero-water\" data center draws its electricity from a local power grid reliant on coal, nuclear, or natural gas, thousands of gallons of water are still being evaporated upstream at the utility power plant to generate that electricity. This is why tech companies are coupling their liquid-cooling architectures with direct solar and wind installations and advanced battery installations to ensure the entire supply chain becomes water-neutral.\r\n<h3>OH, Legislation and Regulations about large Data Centers<\/h3>\r\n<h4>The Tax Exemption Backlash and Public Revenue Loss<\/h4>\r\nSince 2011, Ohio has aggressively courted tech giants by offering <strong>100% sales and use tax exemption<\/strong> on data center server equipment, infrastructure, and construction. However, the cost of these subsidies has exploded far beyond what the state originally anticipated.\r\n<ul>\r\n \t<li><strong>The $1.5 Billion Revenue Gap:<\/strong> In 2024, the tax break cost Ohio $555 million. By 2025, that figure snowballed to <strong>nearly $1.6 billion in foregone revenue<\/strong>\u2014more than 11 times the initial estimates provided by the <a href=\"https:\/\/tax.ohio.gov\/\">Ohio Department of Taxation<\/a>.<\/li>\r\n \t<li><strong>The Impact on Local Public Services:<\/strong> An additional $166.8 million in local county sales taxes was lost. This massive shortfall removes funds that would otherwise support local public schools, infrastructure repairs, and community services, shifting the financial burden back onto Ohio taxpayers.<\/li>\r\n \t<li><strong>The Legislative Freeze:<\/strong> In response to ballooning costs, Governor Mike DeWine issued an <strong>executive order pausing all new data center sales tax exemptions<\/strong>. Meanwhile, state lawmakers introduced bills such as <a href=\"https:\/\/fox8.com\/news\/ohio-bill-would-end-data-center-tax-breaks-effective-oct-1\/\">Ohio House Bill 975<\/a> was created to permanently eliminate all sales and use tax exemptions for large-scale data centers, while <a href=\"https:\/\/www.legislature.ohio.gov\/legislation\/136\/hb646\">Ohio House Bill 646<\/a> aims to reduce the 100% exemption to 50% permanently.<\/li>\r\n<\/ul>\r\nExisting long-term tax deals locked in by giants like Amazon, Meta, and Google through statewide agreements, however, will remain active.\r\n\r\nFor real-time legislative trackers on active statehouse bills (such as HB 646, HB 784, or SB 381), you can audit the docket on the official <a href=\"https:\/\/www.occ.ohio.gov\/data-centers\">Ohio Consumers' Counsel Data Centers Page.<\/a>\r\n<h3>AI and Demand and Depletion of Minerals (Rare Earths)<\/h3>\r\nWhile everyday operations consume electricity and water, the upfront manufacturing and long-term hardware replacement cycles tie AI directly to the environmental and geopolitical crises of critical mineral extraction. (<a href=\"https:\/\/fpanalytics.foreignpolicy.com\/2025\/07\/18\/artificial-intelligence-critical-minerals-supply-chains\/\">Artificial Intelligence and the Critical Minerals Crunch<\/a>). To build the specialized hardware required for deep learning, manufacturers rely heavily on a specific suite of critical minerals:\r\n<ul>\r\n \t<li><strong>Gallium and Germanium:<\/strong> Highly integrated into high-speed semiconductor transistors and the fiber-optic networks that link data centers.<\/li>\r\n \t<li><strong>Neodymium and Dysprosium:<\/strong> Vital rare earth minerals used to manufacture high-strength permanent magnets for data center cooling fans, hard disk drives, and robotic actuators.<\/li>\r\n \t<li><strong>Lithium and Cobalt:<\/strong> The raw materials required for the massive industrial battery arrays used as backup power to protect data centers from power grid failures.<\/li>\r\n<\/ul>\r\n<h4>Long-Term Ecological Damage from \"AI Mining\"<\/h4>\r\nThe environmental cost of acquiring these minerals differs structurally from that of traditional mining, as rare earths are rarely found in concentrated veins. They require destructive, chemically intensive processes to refine.\r\n<h4>The Electronic Waste (E-Waste) Crisis<\/h4>\r\nThe short lifespan of AI infrastructure accelerates the mineral depletion crisis. Because AI models advance so quickly, the specialized GPUs that run them become obsolete in roughly <strong>three to five years<\/strong> (<a href=\"https:\/\/blog.citp.princeton.edu\/2025\/10\/15\/lifespan-of-ai-chips-the-300-billion-question\/\">Center for Information Technology Policy (CITP) at Princeton University).<\/a>\r\n<h4>Structural and Supply Chain Bottlenecks<\/h4>\r\nThe long-term outlook is further complicated by geographic concentration. The <a href=\"https:\/\/www.iea.org\/\">International Energy Agency (IEA)<\/a> notes that China accounts for approximately 60% of global rare earth extraction and over 90% of global rare earth<strong> refining capacity<\/strong>. As Western tech giants build massive server farms, the strain on this highly centralized mineral supply chain is driving a frantic rush to open new, ecologically disruptive mines in parts of Africa, South America, and the Arctic.\r\n<h3><strong>Copyright \/ Intellectual Property <\/strong><\/h3>\r\nThe intersection of Generative AI and copyright is one of the most contentious ethical issues surrounding this new technology, especially among artists and others who make a living on creating and selling their creative works.\r\n\r\nIn this chapter, we are primarily referring to copyright as it exists in the United States. According to the US Copyright Office, copyright is \u201ca type of intellectual property that protects original works of authorship as soon as an author fixes the work in a tangible form of expression\u201d (<a href=\"https:\/\/www.copyright.gov\/what-is-copyright\/\">copyright.gov, n.d.<\/a>). For a work to be protected by copyright, it must be sufficiently original (i.e. different from other, published works) and fixed in a tangible form of expression (i.e. not an idea). Although it was required in the past, copyright protection does not require the creator to register the work with the Copyright Office or take any other formal action. Just fixing the original work in a tangible form of expression is sufficient to gain copyright protection.\r\n\r\nCopyright protects the copyright holder\u2019s ability to make copies of the work, make derivatives, distribute unpublished copies, perform the work, or display the work. The length of copyright protection varies depending on whether the copyright holder is an individual or an organization, when the work was published, and whether the work was published. Cornell University Library maintains an excellent <a href=\"https:\/\/guides.library.cornell.edu\/copyright\/publicdomain\">copyright term chart<\/a>, but if in doubt about whether something is still protected by copyright, try contacting your librarian or campus legal counsel.\r\n<h4>Training LLMs and Copyright<\/h4>\r\nThere are two considerations when it comes to copyright and generative AI. One is the intellectual property that was used to train large language models (LLMs), and the other is the outputs of the LLMs. In the case of training data, there is concern that the creative works of authors, artists, graphic designers, photographers, and others were used without permission as part of the training that make LLMs possible. AI companies are profiting from the use of this content without providing any compensation to the creators that make this wealth possible.\r\n\r\nTo train a large language model, billions of source works must be ingested (this is true for LLMs that create language, as well as those that create images and\/or video). When an LLM creates an output, it isn\u2019t replicating any particular work from its training data; it\u2019s using predictive algorithms to determine what content to produce. This makes it nearly impossible to give attribution to individual works used to produce an LLM output. As of this writing, courts have not definitively determined if using copyrighted works in the training of LLMs is a copyright violation. Anthropic, for example, was able to rely on a fair use argument for some of the works that it used to train Claude models (<a href=\"https:\/\/www.insidetechlaw.com\/blog\/2025\/09\/bartz-v-anthropic-settlement-reached-after-landmark-summary-judgment-and-class-certification\">they were not able to rely on fair use for pirated works, however<\/a>). Despite the outcome of the Anthropic case, many legal cases (<a href=\"https:\/\/www.mishcon.com\/generative-ai-intellectual-property-cases-and-policy-tracker\">around 80 at the time of this writing<\/a>) about the use of copyrighted works for LLM training are still ongoing.\r\n\r\nIt is worth noting that some creative professionals have used specially designed software to make their published works unrecognizable by generative AI, as a way of undermining these tools. One such app, <a href=\"https:\/\/nightshade.cs.uchicago.edu\/whatis.html\">Nightshade<\/a>, was developed by researchers at the University of Chicago. Nightshade allows image creators to make their images look different to a generative AI tool that is undergoing training (although the change is not detectable by human viewers). Over time, the LLM begins to learn to assign incorrect metadata for the images it is training on, because the human-visible image doesn\u2019t match the AI-visible one. This both protects artists and gives them the ability to reduce the effectiveness of tools that they perceive as predatory.\r\n<h4>LLM Outputs and Copyright<\/h4>\r\nWhen it comes to LLM outputs, the concern is that the tools can be used to create works that replace the work of existing artists. In fact, some have even used LLMs to create art \u201cin the style of\u201d an existing, professional artist, effectively destroying that artist\u2019s ability to make a living. In rare cases, LLMs can be used to make almost exact replicas of existing, copyrighted works which, in other contexts, is a clear copyright violation.\r\n\r\nIn general, a work produced solely by an LLM is not protected by copyright, because only works created by humans can be copyright-protected. However, a work that has some AI content and some human content can be protected, with the extent of protection being worked out in courts on a case-by-case basis (<a href=\"https:\/\/www.copyright.gov\/newsnet\/2025\/1060.html\">Copyright Office, n.d.<\/a>). Also, if an LLM output is similar to an existing copyrighted work or character, it could be a copyright violation. Some AI companies shield users from this kind of liability (and many have taken steps to disallow intentional outputs of this kind), but individuals should check the user agreement for the AI tool they are employing.\r\n<h3>Privacy &amp; Security<\/h3>\r\nThe primary concern when it comes to generative AI and privacy is the fact that AI companies use personal data (among other data) to train LLMs, exposing that data to risk of data breaches and also allowing LLMs to use the data in chatbot outputs to other users. In the initial training of LLMs, AI companies use huge amounts of scraped data from the internet that often includes personal data collected without consent. The LLM training process is, overall, not very transparent and, some have argued, not sufficiently regulated (<a href=\"https:\/\/www.bbc.com\/news\/technology-65139406\">McCallum, 2023<\/a>). In Europe, there is some regulation of generative AI as it relates to privacy, but regulation is much laxer in the United States. There is new research showing that, even if a user is careful not to input a lot of data into an AI chatbot, <a href=\"https:\/\/cyberscoop.com\/ai-deanonymization-risks-online-anonymity-study\/\">LLMs are able to use existing information on the internet to identify individuals<\/a>, sometimes in great detail (<a href=\"https:\/\/arxiv.org\/pdf\/2602.16800\">Lerman et al, 2026<\/a>).\r\n\r\nWhile any use of generative AI tools involves some privacy risk, there are ways to improve the protection of your private information while using the tools. For CSU employees, using Microsoft Copilot available through our enterprise license does provide some assurances of privacy and security. There are also generative AI tools that don\u2019t require a login, therefore preserving some of your anonymity, such as <a href=\"https:\/\/duck.ai\/\">Duck.ai<\/a> and <a href=\"https:\/\/chat.mistral.ai\/chat\">Mistral AI (Le Chat)<\/a>. In general, the best way to preserve your privacy with these tools is to refrain from entering any sensitive data into them in the first place.\r\n<h3>Human Labor<\/h3>\r\nWhile AI has emerged as a transformative force in the labor market, it has also raised profound ethical challenges which arise mainly from the way the systems are developed, designed, deployed, and used. In the realm of human labor, these ethical challenges include: displacement of jobs, algorithmic bias in hiring, low-wage and traumatic work environments, and intrusive workplace surveillance.\r\n<h4>Automation and jobs displacement<\/h4>\r\nWith its ability to automate repetitive tasks, there is concern that AI could eventually replace millions of jobs, especially white-collar entry-level positions. A report by McKinsey Global Institute suggests that AI could automate up to 30% of hours currently worked across the US economy by 2030. Further, a study by Stanford and the Digital Economy Lab found a 16% decline in early-career employment within the most AI-exposed fields.\r\n\r\nAI\u2019s impact on the job market may disproportionately affect certain demographics such as minority workers who are overly represented in positions that are at a higher risk of automation, potentially exacerbating existing inequalities. Other demographic groups with slower adaptive capacity, such as older workers who may have fewer transferable skills, will also be affected. AI also is expected to impact the labor market across genders differently. The International Labor Organization (ILO) predicts that 7.8% of women's occupations in high income countries could be automated, totaling around 21 million jobs, compared to only 2.9% of jobs held by men.\r\n\r\nThese impacts on the labor market are expected to create significant challenges, including loss of income for displaced workers and exacerbating unemployment numbers. It is important to note, however, that the effect of AI on employment is not even across different sectors and skill sets. Certain industries will see minimal disruption, whereas others will experience substantial workforce displacement.\r\n<h4>Algorithmic bias in hiring<\/h4>\r\nEmployers are increasingly using AI for recruiting, hiring, and performance evaluation. There is concern related to algorithmic bias in using AI to screen, evaluate, and make decisions on potential employees during the hiring process.\u00a0 Ethical challenges arise in that AI algorithms have the potential to inherit and reinforce existing biases based on the training data it receives. The automated tools used to screen applicants' resumes could unfairly filter out qualified candidates based on historical data. This will potentially lead to an exacerbation of existing discriminatory practices.\r\n<h4>Low wage and traumatic work environments<\/h4>\r\nFor its underlying functionality such as data labeling\/annotation and content moderation and reviewing, AI depends on a large amount of human labor. These labor roles are often outsourced to low-wage contract workers, mostly in developing countries, who feed data to and train the AI algorithms. Companies hire from poor and underserved communities, refugee populations, incarcerated people, and others with few job options. \u00a0A <em>CBS 60 Minutes <\/em>investigation titled <em>Humans in the Loop, for<\/em> instance, showed that the AI data labeling work takes a toll on these workers as they often work long hours with low pay (as little as $2 per hour gross). To minimize their costs, the giant tech companies do not hire directly; instead, they subcontract these workers via digital platforms allowing companies to circumvent labor laws and benefits. There is little transparency, with the workers often not knowing which companies they are working for or which systems they are training, which raises the concern that they may be training, unknowingly, systems that may be used for surveillance and subsequent repression.\r\n\r\nThe workers in content moderation are responsible for finding and flagging content that is deemed inappropriate, such as texts and imagery that contain hate speech, violence, abuse, sexually explicit or other types of harmful content. This work is critical to ensure that the AI can detect this harmful content. The workers exposed to these traumatic tasks often have little or no emotional or mental health support.\r\n<h4>Workplace Surveillance<\/h4>\r\nAnother ethical concern relates to workplace surveillance and privacy. Many companies are now using AI to monitor employee performance and productivity. In the content moderation of work, speed and efficiency are prioritized, and workers are pressured to make decisions within seconds, measured against the pre-determined time every task should take. This means that they work under close surveillance and are punished if they deviate from their assigned tasks. These surveillance practices are also seen in other industries including warehouses work where workers fulfill online orders and delivery work. Delivery drivers, for instance, are monitored through automated surveillance systems. Delivery time \u00a0expectations \u00a0are often unrealistic forcing many drivers to take risks to ensure that they deliver all the packages assigned to them within a specific time. Surveillance, tracking and productivity monitoring thus creates a high-pressure working environment which put workers at the risk of\u00a0 \u00a0injury or death. These practices of monitoring also can overstep into worker privacy violations.\r\n\r\n<em>Human distillation<\/em>\r\n\r\nAnother area of ethical concern regarding human labor is the process often referred to as \u201chuman distillation.\u201d This practice involves using human intelligence, decision-making strategies, and feedback as data for training AI models, effectively transferring aspects of human expertise into artificial systems. The process of human distillation creates several concerns when it comes to human labor. One of these is that it raises fears that employees can be replaced after their operational knowledge is extracted in what might be termed as the \u201ctrain your replacement\u201d paradox. This means that workers are essentially building the software that might make their roles redundant. There are also concerns over intellectual property rights regarding the professional skills used in training AI models.\r\n\r\nKey methods in human distillation include Human-in-the-Loop (HITL), RLHF (Reinforcement Learning from Human Feedback), and synthetic data generation. These approaches map human preferences, steps, or rules into model training data to ensure AI aligns with human logic and values. To translate human knowledge into AI training, the following techniques are primarily used: Human-in-the-Loop (HITL) distillation: This involves human experts reviewing, correcting, and refining AI-generated responses in real-time. The corrected, higher-quality data is then used to train and fine-tune smaller, more efficient AI models. Reinforcement Learning from Human Feedback (RLHF): Humans rate and rank different AI responses. This data is used to create a reward model that teaches the AI which types of answers are preferred, aligning it with human ethics and conversational preferences. Distilling step-by-step: Humans write out rationales or step-by-step instructions for completing a task. These step-by-step human paths serve as supervised data, allowing smaller models to perform complex reasoning without needing massive datasets. Synthesizing from Human Rules: AI models can be trained to follow specific rule-based logic systems originally created by humans, distilling human judgment and uncertainty parameters directly into mathematical scoring\r\n<h3><strong>Power<\/strong><\/h3>\r\nThe combination of the abovementioned ethical concerns can lead to considerable disparity in access to the benefits of AI, as well as inequalities in experiencing negative impacts of AI. \u00a0For example, who has quality access to AI, and what content is represented, can exacerbate existing power dynamics (Furze 2023).\r\n<h4>Who benefits from AI?<\/h4>\r\nAI developments in recent years have involved the concentration of power in terms of who controls how AI exists and how it might exist in the future. The companies leading AI\u2019s growth are some of the wealthiest companies in the world. The large AI companies that are newer were developed with funding from companies that were already among the wealthiest; for example, OpenAI was mainly funded by Microsoft, while Anthropic received large investments from Google and Amazon. These investments have so far led to massive increases in the value of all of these companies.\r\n\r\nBremmer and Suleyman (2025) argue that Big Tech companies can now wield power more like nation-states, but without the regulation, such as democratic accountability, that nation-states are typically subject to. This can undermine democratic norms that are widely accepted, such as the power of people in a democracy to self-regulate. In some cases, the amount of money spent on AI projects, such as the \u201cStargate\u201d project, can be similar to the annual spending even of wealthy nations like Australia (Furze 2026).\r\n\r\nAt a more individual level, while AI is to some extent available to all, many users find their access to higher quality AI models be quite limited unless they purchase subscriptions. To some extent, power is held by those with the economic resources to purchase better AI access, and these benefits can lead to positive feedback loops further concentrating power.\r\n<h4>Who experiences the negative impact of AI?<\/h4>\r\nEnvironmental issues from the high carbon footprint of AI are likely to be unequally distributed. Bender et al. (2021) note that it is critical to consider the environmental costs of AI, which may hit poorer countries that are simultaneously less likely to benefit from AI.\r\n\r\nWhen AI is presented as inevitable, this can result in a gloomy outlook and negative thoughts. The perception of AI as unavoidable is further amplified by the AIs themselves and the companies producing them, producing a new kind of hegemony leading to further disparities (Furze 2023). The prospect of an AI-driven future can be quite disempowering, particularly for younger adults seeking to find careers in this rapidly changing landscape.\r\n<h4>Educational considerations for power in AI<\/h4>\r\nWhile the internet itself has already disrupted traditional information sources such as printed books and newspapers, AI will likely accelerate this process. Unfortunately, as mentioned above, the biases inherent in AI may limit the quality of this information for educational purposes. Many textbooks are offering AI as a component, and this can be a way to limit AI access to information that is (mostly) correct, as students use AIs as study tools.\r\n\r\nTeaching students about AI ethics will likely involve a conversation about the power dynamic resulting from or reinforced by AI. In some ways, AI's connection to power is similar to other technologies, such as the internet or smartphones, which also have environmental and equitability issues, so it may be possible to find some ways to mitigate power concerns by examining what has helped mitigate problems with those technologies. For example, to some extent, support for public access to the internet in public libraries has helped with power issues with internet access. Policies around smartphone waste, including opportunities to recycle and reuse components, have also mitigated some environmental impacts.\r\n<h2>SCENARIOS &amp; CONSIDERATIONS<\/h2>\r\nDifferent disciplines may have stronger links to some components of AI ethical concerns.\u00a0 Here we suggest some case studies that might be explored with students in a variety of contexts.\r\n<h3>Bias<\/h3>\r\n<h4>Scenario<\/h4>\r\nDr. Priya Chandrasekaran had four days before finals week to decide what to recommend, and she was fairly sure that whatever she recommended, someone would be angry.\r\n\r\nAs coordinator of Halloway University's \"Statistics for the Social Sciences,\" a required course taught in eleven sections each fall to roughly 900 students, Chandrasekaran had spent two years trying to solve a problem every large-enrollment course eventually runs into: consistency. With eleven sections taught by a mix of two tenure-track faculty, one lecturer, and eight adjunct instructors \u2014 several of them teaching at two other institutions to make ends meet \u2014 exam difficulty and grading standards had drifted noticeably across sections. Students compared notes; some sections' averages ran fifteen points higher than others on ostensibly the same material. The associate dean had flagged it as an equity issue in its own right: which section you happened to land in was affecting your GPA.\r\n\r\nThe exam-generation tool. Last spring, Chandrasekaran piloted a solution: an AI-assisted exam-bank platform, ItemForge, that could generate large pools of statistically calibrated questions from a shared set of learning objectives, which instructors could then pull from to build individualized exams that were nominally equivalent in difficulty. Adjunct instructors, several of whom told Chandrasekaran privately that they simply did not have the unpaid hours to write high-quality exam questions from scratch every term, were enthusiastic. The two tenure-track faculty were more skeptical but agreed to a trial.\r\n\r\nA pattern in the data. After the midterm, one of the tenure-track faculty, Dr. Marcus Whitfield, noticed something while reviewing item-level statistics the platform provides: several ItemForge-generated word problems used names, contexts, and idioms (references to unfamiliar consumer brands, U.S. sports statistics, colloquial phrasing) that seemed, anecdotally, to trip up international and multilingual students more than domestic ones \u2014 not because the statistics were harder, but because parsing the scenario took longer or was ambiguous. When he cross-referenced item-level scores against the registrar's (limited, self-reported) international-student flag, students in that group scored on average nine points lower specifically on ItemForge-generated items, but showed no significant gap on the traditionally-written items covering the same statistical concepts, which had been retained on the exam as a comparison set.\r\n\r\nA second, separate pattern. Independently, the course also uses an AI-text-detection tool, VerifyWrite, on a short take-home written-interpretation component, where students explain a statistical result in their own words. Several adjunct instructors have been referring unusually high numbers of multilingual and international students for academic-integrity review based on VerifyWrite's flags. One adjunct, Dana Okafor, brought a specific case to Chandrasekaran: a student from Vietnam whose written English was, in Okafor's judgment, clearly her own \u2014 grammatically imperfect but conceptually strong \u2014 but who had been flagged at \"87% likely AI-generated\" by the tool. Okafor did not refer the student for a hearing, trusting her own judgment over the tool's, but admitted she wasn't sure that was the \"correct\" or consistent thing to do, and that other instructors in the course might be handling identical situations differently.\r\n\r\nWhy nobody adopted these tools carelessly. It would be easy to tell this case as a story about instructors blindly trusting software, but that isn't quite what happened. Chandrasekaran had asked ItemForge's sales representative directly, before the pilot, whether the platform's items had been checked for cultural or linguistic bias; she was told the platform used \"diverse item-writing guidelines\" but was not shown any actual bias-testing data. She hadn't pushed further, in part because the university's procurement office had already approved ItemForge as a licensed vendor for an unrelated purpose (adaptive homework problem sets) the year before, which made it feel like a known quantity rather than a new risk. VerifyWrite, similarly, had been adopted university-wide two years earlier by the Provost's office as part of a broader academic-integrity initiative; Chandrasekaran's course didn't choose it so much as inherit it, and individual instructors have no formal mechanism to opt out of using it, only informal discretion in how much weight to give a flag once it appears.\r\n\r\nThe stakeholders, now. With finals in four days:\r\n<ul>\r\n \t<li>The adjunct instructors, as a group, want to keep ItemForge \u2014 without it, several say they will have to write finals from scratch with almost no paid time to do it, and some worry that abandoning it now, days before finals, would be logistically chaotic.<\/li>\r\n \t<li>Dr. Whitfield wants ItemForge pulled immediately from the final exam, arguing that knowingly using items with a demonstrated score gap, even one term after finding it, is indefensible regardless of the inconvenience.<\/li>\r\n \t<li>Dana Okafor wants VerifyWrite flags removed from instructor view entirely for the written component, arguing that if instructors like her are already learning to disregard the tool's judgment in favor of their own, the tool isn't adding anything except risk and stress for the flagged students, many of whom hear about being \"flagged\" before any human review even happens.<\/li>\r\n \t<li>The Registrar's office, asked about better demographic data to investigate the ItemForge pattern further, notes that the \"international student\" flag Whitfield used is a rough proxy \u2014 it doesn't capture multilingual domestic students, who may face similar issues, and pulling more granular demographic data before finals would require a formal data request that typically takes three to four weeks.<\/li>\r\n \t<li>The Provost's office, which owns the VerifyWrite contract university-wide, is not part of this course's decision-making at all, and Chandrasekaran does not have the authority to disable VerifyWrite even if she wanted to \u2014 only to advise her instructors on how much weight to give its flags.<\/li>\r\n<\/ul>\r\nChandrasekaran has to send guidance to all eleven sections by Friday.\r\n<table class=\"grid aligncenter\" style=\"border-collapse: collapse; width: 100%;\" border=\"0\"><caption>Exhibit A: Midterm Item-Type Comparison (Fictional Data)<\/caption>\r\n<thead>\r\n<tr class=\"shaded\">\r\n<th style=\"width: 26.3139%; text-align: center;\" scope=\"col\">Student Group<\/th>\r\n<th style=\"width: 32.4452%; text-align: center;\" scope=\"col\">Avg. Score, Item Forge-Generated Items<\/th>\r\n<th style=\"width: 30.2555%; text-align: center;\" scope=\"col\">Avg. Score, Traditionally-Written Items<\/th>\r\n<th style=\"width: 10.9854%; text-align: center;\" scope=\"col\">Gap<\/th>\r\n<\/tr>\r\n<\/thead>\r\n<tbody>\r\n<tr>\r\n<td style=\"width: 26.3139%;\">Domestic Students<\/td>\r\n<td style=\"width: 32.4452%;\">78%<\/td>\r\n<td style=\"width: 30.2555%;\">77%<\/td>\r\n<td style=\"width: 10.9854%;\">+1<\/td>\r\n<\/tr>\r\n<tr>\r\n<td style=\"width: 26.3139%;\">International Students (registrar flag)<\/td>\r\n<td style=\"width: 32.4452%;\">69%<\/td>\r\n<td style=\"width: 30.2555%;\">76%<\/td>\r\n<td style=\"width: 10.9854%;\">-7<\/td>\r\n<\/tr>\r\n<tr>\r\n<td style=\"width: 26.3139%;\">All students, combined<\/td>\r\n<td style=\"width: 32.4452%;\">77%<\/td>\r\n<td style=\"width: 30.2555%;\">77%<\/td>\r\n<td style=\"width: 10.9854%;\">0<\/td>\r\n<\/tr>\r\n<\/tbody>\r\n<\/table>\r\n<table class=\"grid aligncenter\" style=\"border-collapse: collapse; width: 100%; height: 127px;\" border=\"0\"><caption>Exhibit B:Verify Write Flag Rates by Section (Fictional Data, This Term)<\/caption>\r\n<thead>\r\n<tr class=\"shaded\" style=\"height: 55px;\">\r\n<th style=\"width: 36.9708%; text-align: center; height: 55px;\" scope=\"col\">Instructor<\/th>\r\n<th style=\"width: 21.7883%; text-align: center; height: 55px;\" scope=\"col\">% of Written Components Flagged<\/th>\r\n<th style=\"width: 30.2555%; text-align: center; height: 55px;\" scope=\"col\">Instructor's Self-Reported Override Rate<\/th>\r\n<\/tr>\r\n<\/thead>\r\n<tbody>\r\n<tr style=\"height: 18px;\">\r\n<td style=\"width: 36.9708%; height: 18px;\">Dr. Whitfield (tenure-track)<\/td>\r\n<td style=\"width: 21.7883%; height: 18px;\">4%<\/td>\r\n<td style=\"width: 30.2555%; height: 18px;\">Reviews every flag personally; overrides \u223c 60%<\/td>\r\n<\/tr>\r\n<tr style=\"height: 18px;\">\r\n<td style=\"width: 36.9708%; height: 18px;\">Dana Okafor (adjunct)<\/td>\r\n<td style=\"width: 21.7883%; height: 18px;\">11%<\/td>\r\n<td style=\"width: 30.2555%; height: 18px;\">Reviews every flag personally; overrides \u223c 70%<\/td>\r\n<\/tr>\r\n<tr style=\"height: 18px;\">\r\n<td style=\"width: 36.9708%; height: 18px;\">Adjunct Instructor C<\/td>\r\n<td style=\"width: 21.7883%; height: 18px;\">9 %<\/td>\r\n<td style=\"width: 30.2555%; height: 18px;\">Refers most flags to formal integrity review without independent assessment<\/td>\r\n<\/tr>\r\n<tr style=\"height: 18px;\">\r\n<td style=\"width: 36.9708%; height: 18px;\">Adjunct Instructor D<\/td>\r\n<td style=\"width: 21.7883%; height: 18px;\">3%<\/td>\r\n<td style=\"width: 30.2555%; height: 18px;\">Rarely checks the dashboard; flags mostly go unreviewed<\/td>\r\n<\/tr>\r\n<\/tbody>\r\n<\/table>\r\n<h4>Considerations<\/h4>\r\n<h5>Diagnosing the two problems<\/h5>\r\n<ol>\r\n \t<li>The ItemForge gap and the VerifyWrite gap are both examples of AI bias, but they arise from different parts of the AI pipeline. What is the difference between a tool that generates biased content and a tool that evaluates students in a biased way? Does that difference matter for how each problem should be fixed?<\/li>\r\n \t<li>Exhibit A shows no gap in the combined average \u2014 only once the data is disaggregated does the pattern appear. What does this suggest about the limits of the kind of monitoring most instructors realistically have time to do?<\/li>\r\n<\/ol>\r\n<h5>Weighing the stakeholders<\/h5>\r\n<ol>\r\n \t<li>The adjunct instructors' concerns are logistical and economic (unpaid labor, time pressure), not indifference to bias. How should Chandrasekaran weigh a genuine resource constraint against a documented equity problem? Is there a version of \"pull ItemForge from finals\" that doesn't simply transfer the cost onto the instructors with the least power in the course?<\/li>\r\n \t<li>Dana Okafor already overrides most flags using her own judgment. Is an AI-detection tool that experienced instructors mostly learn to override still doing useful work, or has it just become a source of stress and risk for flagged students with no compensating benefit?<\/li>\r\n<\/ol>\r\n<h5>Deciding what to do next, with four days left<\/h5>\r\n<ol>\r\n \t<li>What should Chandrasekaran's Friday guidance say about ItemForge for the final exam \u2014 keep it, drop it entirely, or something in between (e.g., keep it but require human review of any item involving cultural\/linguistic context)? What are you trading off with your answer?<\/li>\r\n \t<li>The Registrar's better demographic data would take three to four weeks \u2014 well past finals. Should Chandrasekaran wait for better data before acting, or act now on the imperfect data she already has? What does your answer suggest about how much evidence an institution should require before treating a pattern as a real problem?<\/li>\r\n<\/ol>\r\n<h3>Environmental Impact<\/h3>\r\n<h4>Scenario: The Municipal Zoning and Local Utility Decision<\/h4>\r\nA city council in a rural, developing region must vote on whether to approve a zoning permit for a tech giant to build a new 500-megawatt hyperscale AI data center campus. The tech giant promises a $5 billion local investment, hundreds of high-paying construction jobs, and a massive boost to the municipal tax base to fund local schools and roads.\r\n<h4>Considerations<\/h4>\r\n<ul>\r\n \t<li>The proposed data center requires massive amounts of power, forcing the regional energy utility to build a new natural gas plant to meet demand.<\/li>\r\n \t<li>The facility\u2019s cooling system will draw heavily from the local aquifer, which supplies the town's drinking water and agricultural irrigation.<\/li>\r\n \t<li>The council must decide if immediate economic growth justifies long-term ecological risks. Approving the project could strain the local power grid\u2014causing residential electricity bills to spike\u2014and permanently deplete the local water supply, threatening the region's agricultural economy.<\/li>\r\n<\/ul>\r\n<h3>Copyright \/ Intellectual Property<\/h3>\r\n<h4>Scenario<\/h4>\r\nEliza is working on creating flyers for a fundraiser for the local food bank, but she\u2019s really struggling. Eliza knows she has very poor graphic design skills, but she doesn\u2019t know any graphic designers, and she needs to get these flyers up as soon as possible. She wants to support local graphic designers and their work, but she doesn\u2019t feel that she has the funds or time to hire someone. She decides to use generative AI to create the flyer, reasoning that her only alternative is the ugly flyer she has tried to make.\r\n<h4>Considerations<\/h4>\r\n<ul>\r\n \t<li>When is it appropriate for someone to forgo paying a professional for their creative labor and instead use generative AI?<\/li>\r\n \t<li>What might be the societal outcomes if creative workers are frequently undervalued in favor of easy-to-access generative AI tools?<\/li>\r\n \t<li>What if Eliza had some money to pay for a professional graphic designer to help her, but would prefer to donate the money to the food bank's drive? Where does her ethical obligation lie in this situation?<\/li>\r\n<\/ul>\r\n<h3>Privacy &amp; Security<\/h3>\r\n<h4>Scenario:<\/h4>\r\nGina is having some health issues and is embarrassed to talk about them with her doctor. She turns to a generative AI chatbot instead and is relieved to get some reasonable-sounding advice. Gina\u2019s mom warns her that generative AI tools train on the conversations that she has and might now have sensitive data about Gina in its training data. Gina is not worried about this and doubts anyone would be able to find the data and trace it back to her.\r\n<h4>Considerations:<\/h4>\r\n<ul>\r\n \t<li>Does Gina\u2019s age matter in this scenario? How would knowing her age change your response to the scenario?<\/li>\r\n \t<li>What risks is Gina taking by entering personal data? What is she gaining in exchange? Is this exchange worth it, in your opinion?<\/li>\r\n \t<li>How could system-level change remove the incentive for Gina to use a generative AI chatbot for this purpose?<\/li>\r\n<\/ul>\r\n<h3>Human Labor<\/h3>\r\n<h4>Scenario:<\/h4>\r\nAlani, a computer science student, learns that the generative AI model they use for class projects was trained on data labeled by low\u2011paid workers overseas. These workers spent hours tagging violent or explicit content to make the model \u201csafe.\u201d However, she reasons, now that the technology exists, there is no harm in making use of it \u2013 in fact, if people don\u2019t use the tool, these workers\u2019 suffering would be for nothing.\r\n<h4>Considerations:<\/h4>\r\n<ul>\r\n \t<li>In contexts like this one, to what extent should individual people change their behaviors, as opposed to putting the blame\/responsibility on large corporations? What actions are most likely to make positive change and how does acting in alignment with values play a role?<\/li>\r\n \t<li>Much of our technology is only possible because of the exploitation of people in less advantaged parts of the world. Should we stop using technology for this reason?<\/li>\r\n<\/ul>\r\n<h4>Scenario:<\/h4>\r\nA design major named Betsy uses an AI art generator to complete freelance logo commissions. The client is thrilled with the quick turnaround, but Betsy\u2019s classmate Aarathi argues that Betsy is undercutting human artists by relying on automation. Betsy continues to do this freelance work, insisting that she is simply using available tools efficiently, and that generative AI is the way of the future graphic design.\r\n<h4>Considerations:<\/h4>\r\n<ul>\r\n \t<li>Is generative AI use inevitable in every field? What might be AI\u2019s role in professions that currently rely heavily on human creativity?<\/li>\r\n \t<li>Do individuals have an obligation to pay and employ creative professionals, even though generative AI tools can sometimes do a good-enough job? What uniquely human value do artists bring to our world?<\/li>\r\n<\/ul>\r\n<h3>Power<\/h3>\r\n<h4>Scenario:<\/h4>\r\nSarah has been trying to research potential career options for herself. Sarah\u2019s only electronic device is her phone, which is older, and she is limited in the amount of AI time available to her. Usually, after asking one or two more complicated questions, the AI she is using becomes very slow. She then has to keep going back and forth between other tasks she needs to complete at home and the AI to see if it has responded. Most of the careers the AI is suggesting for her interests require graduate study, but due to her low household income and need to work while studying, getting through an undergraduate degree is going to be quite challenging. AI doesn\u2019t seem to have information about lower barrier-to-entry or part-time jobs that exist in those fields that she could potentially start off with to gain valuable experience. \u00a0When she ask the AI to fix this issue, it thinks for a long time and then just repeats the same information again without listening to her updated request. Some of her classmates have paid for premium AI subscriptions and are not having any of these problems, as AI is quickly suggesting career paths that are within their reach.\r\n<h4>Considerations:<\/h4>\r\n<ul>\r\n \t<li>Who has more power in this situation?<\/li>\r\n \t<li>Would having better AI access through libraries help with some of the power issues?<\/li>\r\n \t<li>What else could be done to mitigate this?<\/li>\r\n<\/ul>\r\n<h2>CONCLUSION<\/h2>\r\nThis chapter scratches the surface of ethical issues related to generative AI, but we hope it provides a helpful overview to spur further discussion and exploration. The content and case studies described here could be a useful way to explore the ethics of AI in the classroom. To ensure AI literacy, it is important to directly address these issues with students.\r\n\r\nGenerative AI is a rapidly changing technology, and there are likely to be new considerations in the future. We did our best to address ethical issues related to generative AI as they were discussed at the time of writing. As these technologies continue to evolve, some problems may be resolved but also new issues may arise. We recommend re-examining ethical considerations for AI usage on a regular basis.\r\n<h2>HELPFUL RESOURCES<\/h2>\r\n<h3>Bias<\/h3>\r\n<ul>\r\n \t<li><a href=\"https:\/\/cee.ucdavis.edu\/ai-student-writing\">AI &amp; Student Writing<\/a> \u2013 UC Davis Center for Educational Effectiveness<\/li>\r\n \t<li><a href=\"https:\/\/post.parliament.uk\/research-briefings\/post-pn-0712\/\">Use of AI in education delivery and assessment<\/a> \u2013 UK Parliament<\/li>\r\n \t<li><a href=\"https:\/\/fas.org\/publication\/modernizing-ai-fairness-analysis-in-education-contexts\/\">Modernizing AI Fairness Analysis in Education Contexts<\/a> \u2013 Federation of American Scientists<\/li>\r\n \t<li><a href=\"https:\/\/digital-strategy.ec.europa.eu\/en\/policies\/regulatory-framework-ai\">The The AI ActAct<\/a> \u2013\u2013 European Comm Commission<\/li>\r\n \t<li><a href=\"https:\/\/er.educause.edu\/articles\/2025\/3\/ai-procurement-in-higher-education-benefits-and-risks-of-emerging-tools\">AI Procurement in Higher Education<\/a> \u2013 EDUCAUSE Review<\/li>\r\n \t<li><a href=\"https:\/\/files.eric.ed.gov\/fulltext\/ED661949.pdf\">Designing for Education with Artificial Intelligence: An Essential Guide for Developers<\/a> \u2013 U.S. Department of Education<\/li>\r\n \t<li><a href=\"https:\/\/www.nist.gov\/itl\/ai-risk-management-framework\">AI Risk Management Framework<\/a> \u2013 National Institute of Standards and Technology<\/li>\r\n \t<li><a href=\"https:\/\/standards.ieee.org\/ieee\/7003\/11357\/\">Standard for Algorithmic Bias Considerations<\/a> \u2013 IEEE<\/li>\r\n \t<li>Standard for Algorithmic Bias Considerations \u2013<\/li>\r\n \t<li><a href=\"https:\/\/libraryfreedom.org\/wp-content\/uploads\/2026\/07\/LFP-Critical-AI-in-Higher-Education-Toolkit.pdf\">Critical AI in Higher Education Toolkit<\/a> \u2013 Library Freedom Project<\/li>\r\n<\/ul>\r\n<h3>Environmental Impact<\/h3>\r\n<ul>\r\n \t<li><a href=\"https:\/\/www.youtube.com\/@ClimateEmergencyForum\">Climate Emergency Forum<\/a><\/li>\r\n \t<li><a href=\"https:\/\/youtu.be\/eOkz6KqqakA?si=eAFN-8HkekkKk8X1\">How Green Is Your Prompt?<\/a><\/li>\r\n \t<li><a href=\"https:\/\/www.datacenterwatch.org\/\">Data Center Watch<\/a><\/li>\r\n \t<li><a href=\"https:\/\/spectrumnews1.com\/oh\/columbus\/news\/2026\/02\/22\/data-centers--impact-on-the-environment-\">Data centers in Ohio: Economic boost or environmental burden?<\/a><\/li>\r\n \t<li><a href=\"https:\/\/www.ceres.org\/resources\/reports\/drained-by-data-the-cumulative-impact-of-data-centers-on-regional-water-stress?\">Drained by Data: The Cumulative Impact of Data Centers on Regional Water Stress<\/a><\/li>\r\n \t<li><a href=\"https:\/\/theoec.org\/\">Ohio Environmental Council.<\/a><\/li>\r\n \t<li>Oh<a href=\"https:\/\/ohiohouse.gov\/news\/republican\/ohio-house-passes-bill-establishing-the-ohio-data-center-study-commission-142643\">io Data Center Study Commission<\/a>,<\/li>\r\n<\/ul>\r\n<h3>Copyright \/ Intellectual Property<\/h3>\r\n<ul>\r\n \t<li><a href=\"https:\/\/copyright.gov\/ai\/\">Report on Copyright and Artificial Intelligence<\/a> \u2013 U.S. Copyright Office<\/li>\r\n \t<li><a href=\"https:\/\/www.wired.com\/story\/ai-copyright-case-tracker\/\">AI Copyright Case Tracker<\/a> \u2013 Wired Magazine<\/li>\r\n \t<li><a href=\"https:\/\/blogs.gwu.edu\/law-eti\/ai-litigation-database\/\">AI Litigation Database<\/a> \u2013 George Washington University<\/li>\r\n \t<li><a href=\"https:\/\/blogs.microsoft.com\/on-the-issues\/2023\/09\/07\/copilot-copyright-commitment-ai-legal-concerns\/\">Microsoft Copilot Copyright Commitment<\/a><\/li>\r\n \t<li><a href=\"https:\/\/www.ropesgray.com\/en\/insights\/alerts\/2025\/07\/a-tale-of-three-cases-how-fair-use-is-playing-out-in-ai-copyright-lawsuits\">A Tale of Three Cases: How Fair Use Is Playing Out in AI Copyright Lawsuits<\/a> \u2013 Ropes &amp; Gray Legal Firm<\/li>\r\n \t<li><a href=\"https:\/\/nightshade.cs.uchicago.edu\/whatis.html\">Nightshade App<\/a><\/li>\r\n \t<li><a href=\"https:\/\/guides.library.cornell.edu\/copyright\/publicdomain\">Copyright Term and Public Domain Chart<\/a> \u2013 Cornell University<\/li>\r\n<\/ul>\r\n<h3>Privacy &amp; Security<\/h3>\r\n<ul>\r\n \t<li><a href=\"https:\/\/hai.stanford.edu\/news\/privacy-ai-era-how-do-we-protect-our-personal-information\">Privacy in an AI Era: How Do We Protect Our Personal Information?<\/a> - Stanford University Human-Centered Artificial Intelligence<\/li>\r\n \t<li><a href=\"https:\/\/www.sciencedirect.com\/org\/science\/article\/pii\/S1062737525000605\">Artificial Intelligence (AI) and the Future of Information Privacy<\/a> \u2013 Journal of Global Information Management<\/li>\r\n \t<li><a href=\"https:\/\/www.forbes.com\/sites\/federicoguerrini\/2024\/11\/17\/ai-driven-dark-patterns-how-artificial-intelligence-is-supercharging-digital-manipulation\/\">AI-Driven Dark Patterns: How Artificial Intelligence Is Supercharging Digital Manipulation<\/a> - Forbes<\/li>\r\n \t<li><a href=\"https:\/\/www.ed.gov\/sites\/ed\/files\/documents\/ai-report\/ai-report.pdf\">Artificial Intelligence and the Future of Teaching and Learning Insights and Recommendations<\/a> \u2013 Office of Educational Technology (see pg. 32)<\/li>\r\n \t<li><a href=\"https:\/\/chat.mistral.ai\/chat\">Mistral AI<\/a> and <a href=\"https:\/\/duck.ai\/\">Duck.ai<\/a><\/li>\r\n<\/ul>\r\n<h3>Human Labor<\/h3>\r\n<ul>\r\n \t<li>Humans in the loop; Training AI takes heavy toll on Kenyans working for $2 an hour<\/li>\r\n \t<li><a href=\"https:\/\/www.cbsnews.com\/video\/60minutes-2025-06-29\/?intcid=CNM-00-10abd1h\">https:\/\/www.cbsnews.com\/video\/60minutes-2025-06-29\/?intcid=CNM-00-10abd1h<\/a><\/li>\r\n \t<li>How big AI companies exploit data workers in Kenya-<a href=\"https:\/\/www.youtube.com\/watch?v=ehkECk2KJjY&amp;t=149s\">https:\/\/www.youtube.com\/watch?v=ehkECk2KJjY&amp;t=149s<\/a><\/li>\r\n \t<li>Artificial Intelligence recommendations - <a href=\"https:\/\/www.unesco.org\/en\/artificial-intelligence\/recommendation-ethics\">https:\/\/www.unesco.org\/en\/artificial-intelligence\/recommendation-ethics<\/a><\/li>\r\n \t<li>Generative AI and the Future of Work in America | McKinsey,\u201d accessed January 5, 2025,<\/li>\r\n \t<li><a href=\"https:\/\/www.mckinsey.com\/mgi\/our-research\/generative-ai-and-the-future-of-work-in-america\">https:\/\/www.mckinsey.com\/mgi\/our-research\/generative-ai-and-the-future-of-work-in-america<\/a><\/li>\r\n<\/ul>\r\n<h3>Power<\/h3>\r\n<ul>\r\n \t<li><a href=\"https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/\">Teaching AI Ethics: Power<\/a>. Leon Furze, (2023). <a href=\"https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/\">https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/<\/a><\/li>\r\n \t<li><a href=\"https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf\">Teaching AI Ethics: A Guide for Educators<\/a>. Leon Furze, 2026. <a href=\"https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf\">https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf<\/a><\/li>\r\n \t<li><a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3442188.3445922\">On the dangers of stochastic parrots: Can language models be too big?<\/a>\ud83e\udd9c. Bender, E. M., Gebru, T., McMillan-Major, A., &amp; Shmitchell, S. (2021). <em>Proceedings of the 2021 ACM conference on fairness, accountability, and transparency<\/em> (pp. 610-623). <a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3442188.3445922\">https:\/\/dl.acm.org\/doi\/10.1145\/3442188.3445922<\/a><\/li>\r\n<\/ul>\r\n<h2>REFERENCES<\/h2>\r\nAI-Assisted Exam Variant Generation: A Human-in-the-Loop Framework for Automatic Item Creation. (2025). <em>Education Sciences, 15<\/em>(8).\r\n\r\nThe AP-NORC Center for Public Affairs Research. (October 2025). <a href=\"https:\/\/apnorc.org\/?post_type=project&amp;p=11263\">https:\/\/apnorc.org\/projects\/epic-climate-change-2025\/<\/a>\r\n\r\nAshwin, J., Chhabra, A., &amp; Rao, V. (2023). Using large language models for qualitative analysis can introduce serious bias. <em>arXiv preprint<\/em>.\r\n\r\nBias and representation in AI generated text-to-image in education: A systematic review. (2026). [Journal not fully specified in source; verify prior to submission].\r\n\r\nBender, E. M., Gebru, T., McMillan-Major, A., &amp; Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In *<em>Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency<\/em>* (pp. 610-623). Association for Computing Machinery. <a href=\"https:\/\/doi.org\/10.1145\/3442188.3445922\" target=\"_blank\" rel=\"noopener\">https:\/\/doi.org\/10.1145\/3442188.3445922<\/a>\r\n\r\nBird, K. A., Castleman, B. L., Mabel, Z., &amp; Song, Y. (2021). Bringing transparency to predictive analytics: A systematic comparison of predictive modeling methods in higher education. <em>AERA Open<\/em>, 7.\r\n\r\nBird, K. A., Castleman, B. L., &amp; Song, Y. (2023). Are algorithms biased in education? Exploring racial bias in predicting community college student success (EdWorkingPaper No. 23-717). <em>Annenberg Institute at Brown University<\/em>.\r\n\r\nBird, K. A., Castleman, B. L., &amp; Song, Y. (2024). Are algorithms biased in education? Exploring racial bias in predicting community college student success. <em>Journal of Policy Analysis and Management<\/em>.\r\n\r\nBisbee, J., Imai, K., Rosenman, E., &amp; Zhou, X. (2024). Synthetic replacements for human survey data? The perils of large language models. <em>Political Analysis<\/em>.\r\n\r\nbioRxiv. (2024, December 3). Generative AI divide: How college students' backgrounds affect their Gen AI literacy.\r\n\r\nBremmer, I., &amp; Suleyman, M. (2023). The AI power paradox: Can states learn to govern artificial intelligence before it is too late? <em>Foreign Affairs, 102<\/em>(5), 26-43. <a href=\"https:\/\/www.foreignaffairs.com\/world\/artificial-intelligence-power-paradox\" target=\"_blank\" rel=\"noopener\">https:\/\/www.foreignaffairs.com\/world\/artificial-intelligence-power-paradox<\/a>\r\n\r\nBuolamwini, J., &amp; Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. <em>Proceedings of Machine Learning Research, 81,<\/em> 1\u201315.\r\n\r\nCalMatters\/The Markup. (2025, July 15). Turnitin helps colleges detect AI writing \u2014 but its cost is more than just a price tag. CalMatters.\r\n\r\nChanging Higher Ed. (2026, June 9). AI procurement in higher education and the risk of bias at scale [Podcast episode featuring K. Panetta].\r\n\r\nChauhan, K. (2026). <em>Dead science walking: Publication bias and the AI scientist pipeline<\/em>. arXiv preprint arXiv:2606.04220<em>.<\/em>\r\n\r\nDeStefano, C., Hackney, J., &amp; Moskal, P. (2026). Understanding the generative AI divide: Faculty and student perspectives in higher education.<em> Online Learning, 30<\/em>(2), 31\u201348.\r\n\r\nEDUCAUSE Review. (2025, March 11). AI procurement in higher education: Benefits and risks of emerging tools.\r\n\r\nFurze, L. (2023, June 19). <em>Teaching AI ethics: Power<\/em>.\u00a0 <a href=\"https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/\">https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/<\/a>\r\n\r\nFurze, L. (2026). Teaching AI ethics: A guide for educators. <a href=\"https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf\">https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf<\/a>\r\n\r\nGarc\u00eda Mathewson, T. (2023, August 14). AI detection tools falsely accuse international students of cheating. <em>The Markup<\/em>.\r\n\r\nG\u00e1ndara, D., Anahideh, H., Ison, M. P., &amp; Picchiarini, L. (2024). Inside the black box: Detecting and mitigating algorithmic bias across racialized groups in college student-success prediction. <em>AERA Open<\/em>.\r\n\r\nGeneralization bias in large language model summarization of scientific research. (2025). [Journal not fully specified in source, PMC-indexed; verify prior to submission].\r\n\r\nHuang, R., Wu, H., Yuan, Y., Xu, Y., Qian, H., Zhang, C., Wei, X., Lu, S., Zhang, X., Kan, J., Wan, C., &amp; Liu, Y. (2025). Evaluation and bias analysis of large language models in generating synthetic electronic health records: Comparative study. <em>Journal of Medical Internet Research, 27<\/em>, e65317.\r\n\r\nInside Higher Ed. (2025, April 21). The digital divide: Student generative AI access.\r\n\r\nJohnson, D. (2026, March 4). LLMs are getting better at unmasking people online. <em>CyberScoop<\/em>. <a href=\"https:\/\/cyberscoop.com\/ai-deanonymization-risks-online-anonymity-study\/\">https:\/\/cyberscoop.com\/ai-deanonymization-risks-online-anonymity-study\/<\/a>\r\n\r\nKarell, D. (2026). Not-quite-human tastes: The stylized omnivorousness of LLM survey surrogates. <em>arXiv preprint<\/em>.\r\n\r\nKelly, J, S. Pham, and S. Reilly \u201cHow much water does AI really use?\u201d CBS News Data, June 9, 2026 (<a href=\"https:\/\/www.cbsnews.com\/projects\/2026\/how-much-water-ai-uses\/\" target=\"_blank\" rel=\"noopener\">https:\/\/www.cbsnews.com\/projects\/2026\/how-much-water-ai-uses\/<\/a>)\r\n\r\nLermen, S., Paleka, D., Swanson, J., Aerni, M., Carlini, N., &amp; Tram\u00e8r, F. (2026). Large-scale online deanonymization with LLMs. <em>arXiv preprint<\/em> arXiv:2602.16800.\r\n\r\nLi, M., et al. (2025). <em>Understanding and mitigating the bias inheritance in LLM-based data augmentation on downstream tasks<\/em>. arXiv preprint arXiv:2502.04419.\r\n\r\nLiang, W., Yuksekgonul, M., Mao, Y., Wu, E., &amp; Zou, J. (2023). GPT detectors are biased against non-native English writers. <em>Patterns, 4<\/em>(7).\r\n\r\nLow, A., &amp; Kalender, Z. Y. (2023). Data dialogue with ChatGPT: Using Code Interpreter to simulate and analyse experimental data. <em>arXiv preprint <\/em>arXiv:2311.12415.\r\n\r\nMcCallum, S. (2023, March 31). ChatGPT banned in Italy over privacy concerns. <em>BBC News.<\/em> <a href=\"https:\/\/www.bbc.com\/news\/technology-65139406\">https:\/\/www.bbc.com\/news\/technology-65139406<\/a>\r\n\r\nMishcon de Reya. (n.d.). Generative AI \u2013 IP cases and policy tracker. <em>Mishcon de Reya LLP<\/em>. Retrieved July 24, 2026, from <a href=\"https:\/\/www.mishcon.com\/generative-ai-intellectual-property-cases-and-policy-tracker\">https:\/\/www.mishcon.com\/generative-ai-intellectual-property-cases-and-policy-tracker<\/a>\r\n\r\nO\u2019Donnell, J and C Crownhart \u201cWe did the math on AI\u2019s energy footprint. Here\u2019s the story you haven\u2019t heard,\u201d MIT Technology Review, May 20, 2025 (<a href=\"https:\/\/www.technologyreview.com\/2025\/05\/20\/1116327\/ai-energy-usage-climate-footprint-big-tech\/\" target=\"_blank\" rel=\"noopener\">https:\/\/www.technologyreview.com\/2025\/05\/20\/1116327\/ai-energy-usage-climate-footprint-big-tech\/<\/a>).\r\n\r\nOuyang, S., Wang, C., Li, Y., &amp; Zhang, T. (2024). An empirical study of the non-determinism of ChatGPT in code generation. ACM Transactions on Software Engineering and Methodology.\r\n\r\nWong, W. (2025, September 19). The new digital divide in higher education: AI as public capability or private advantage? <em>PA Times Online.<\/em> <a href=\"https:\/\/patimes.org\/the-new-digital-divide-in-higher-education-ai-as-public-capability-or-private-advantage\/\" target=\"_blank\" rel=\"noopener\">https:\/\/patimes.org\/the-new-digital-divide-in-higher-education-ai-as-public-capability-or-private-advantage\/<\/a>\r\n\r\nParker, Will \u201cThe Americans Striking It Rich in the Data-Center Buildout.\u201d The Wall Street Journal, July 13, 2026 (<a href=\"https:\/\/www.wsj.com\/real-estate\/commercial\/the-americans-striking-it-rich-in-the-data-center-buildout-71c6a8a8\" target=\"_blank\" rel=\"noopener\">https:\/\/www.wsj.com\/real-estate\/commercial\/the-americans-striking-it-rich-in-the-data-center-buildout-71c6a8a8<\/a>).\r\n\r\nSoomro, K.A., Kale, U., Curtis, R. et al. (2020). Digital divide among higher education faculty. <em>International Journal of Educational Technology in Higher Education,<\/em> <em>17<\/em>(21). <a href=\"https:\/\/doi.org\/10.1186\/s41239-020-00191-5\" target=\"_blank\" rel=\"noopener\">https:\/\/doi.org\/10.1186\/s41239-020-00191-5<\/a>\r\n\r\nRho, J., &amp; Karumbaiah, S. (2026). How students interpret representational cues in AI-generated educational illustrations. In E. G. Blanchard, G. Chen, M. Chi, &amp; S. Isotani (Eds.), <em>Proceedings of the 27th International Conference on Artificial Intelligence in Education (AIED)<\/em>. Springer.\r\n\r\nRibes-Lafoz, M., Navarro-Colorado, B., &amp; Rovira-Collado, J. (2026). Exploring strategies to detect and mitigate bias in AI in education: Students' perceptions and didactic approaches. Trends in Higher Education, 5(2), 33.\r\n\r\nRitchie, H. (2025, Aug 21). <em>What's the carbon footprint of using ChatGPT or Gemini?<\/em> <a href=\"https:\/\/hannahritchie.substack.com\/p\/ai-footprint-august-2025\" target=\"_blank\" rel=\"noopener\">https:\/\/hannahritchie.substack.com\/p\/ai-footprint-august-2025<\/a>\r\n\r\nRuta, M. R., Gaidici, T., Irwin, C., &amp; Lifshitz, J. (2025). ChatGPT for univariate statistics: Validation of AI-assisted data analysis in healthcare research. Journal of Medical Internet Research, 27, e63550.\r\n\r\nSchmidt, S. (2025, Sep 5). Bartz v. Anthropic: Settlement reached after landmark summary judgment and class certification. <em>Inside Tech Law. <\/em><a href=\"https:\/\/www.insidetechlaw.com\/blog\/2025\/09\/bartz-v-anthropic-settlement-reached-after-landmark-summary-judgment-and-class-certification\">https:\/\/www.insidetechlaw.com\/blog\/2025\/09\/bartz-v-anthropic-settlement-reached-after-landmark-summary-judgment-and-class-certification<\/a>\r\n\r\nSi, C., Yang, D., &amp; Hashimoto, T. (2024). <em>Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers<\/em>. ArXiv Reprint arXiv:2409.04109\r\n\r\nSpringer. (2026a). Academic performance and inequality: Can generative AI widen the digital divide in higher education?\r\n\r\nSpringer. (2026b). From access to ability: Educator's perspectives on AI and new digital divides in higher education. Digital Society.\r\n\r\nStudent-Centered LLM Q&amp;A Study. (2025). Evaluating LLM-generated Q&amp;A test: A student-centered study. In Artificial Intelligence in Education (Springer).\r\n\r\nSurveillance and disability in online proctored exams: Student perspectives and design implications. (2025). arXiv preprint.\r\n\r\nThe homogenizing effect of large language models on human expression and thought. (2026). Trends in Cognitive Sciences.\r\n\r\nUniversity of Kansas Center for Teaching Excellence. (n.d.). Addressing bias in AI.\r\n\r\nUniversity of Pittsburgh Teaching Center. (n.d.). Encouraging academic integrity.\r\n\r\nU.S. Copyright Office (n.d.). Copyright Office Releases Part 2 of Artificial Intelligence Report. <em>Newsnet issue 1060<\/em>. Retrieved July 24, 2026, from <a href=\"https:\/\/www.copyright.gov\/newsnet\/2025\/1060.html\">https:\/\/www.copyright.gov\/newsnet\/2025\/1060.html<\/a>\r\n\r\nU.S. Copyright Office (n.d.). <em>What is Copyright?<\/em> Retrieved July 24, 2026, from <a href=\"https:\/\/www.copyright.gov\/what-is-copyright\/\">https:\/\/www.copyright.gov\/what-is-copyright\/<\/a>\r\n\r\nVanderbilt University Brightspace\/Instructional Technology. (2023, August 16). Guidance on AI detection and why we're disabling Turnitin's AI detector.\r\n\r\nWang, Y., et al. (2026). <em>BRIDGE the gap: Mitigating bias amplification in automated scoring of English language learners via inter-group data augmentation<\/em>. arXiv preprint arXiv:2602.23580.\r\n\r\nYan, Y., Y. X. Pang, X. Luo, Q. Lin, C. H. Pang, H, Zhang, X. Gao, &amp; T. Wu (2024). Carbon dioxide-focused greenhouse gas emissions from petrochemical plants and associated industries: Critical overview, recent advances and future prospects of mitigation strategies. <em>Process Safety and Environmental Protection, 188, <\/em>406\u2013421.\r\n\r\nYoder-Himes, D. R., et al. (2022). Racial, skin tone, and sex disparities in automated proctoring software. <em>Frontiers in Education, 7<\/em>.\r\n<h2>AI STATEMENT<\/h2>\r\n<table class=\"grid aligncenter\">\r\n<thead>\r\n<tr>\r\n<th style=\"text-align: center;\" scope=\"col\"><strong>Category<\/strong><\/th>\r\n<th style=\"text-align: center;\" scope=\"col\"><strong>Label<\/strong><\/th>\r\n<th style=\"text-align: center;\" scope=\"col\"><strong>Description of AI Use<\/strong><\/th>\r\n<\/tr>\r\n<\/thead>\r\n<tbody>\r\n<tr>\r\n<td>Content Research<\/td>\r\n<td><img src=\"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-content\/uploads\/sites\/216\/2026\/08\/Picture1.png\" alt=\"4 squares within a square\" width=\"55\" height=\"55\" class=\"aligncenter size-full wp-image-160\" \/>\r\n\r\nCyborg<\/td>\r\n<td>Some sections used generative AI to find sources to cite.<\/td>\r\n<\/tr>\r\n<tr>\r\n<td>Writing\u2014Content Generation<\/td>\r\n<td><img src=\"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-content\/uploads\/sites\/216\/2026\/08\/Picture2.png\" alt=\"4 squares within a square\" width=\"55\" height=\"55\" class=\"aligncenter size-full wp-image-161\" \/>\r\n\r\nHandyperson<\/td>\r\n<td>Some section authors used generative AI to write content, with significant human review. A few case studies were created by generative AI and then edited.<\/td>\r\n<\/tr>\r\n<tr>\r\n<td>Writing\u2014Review &amp; Editing<\/td>\r\n<td><img src=\"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-content\/uploads\/sites\/216\/2026\/08\/Picture2.png\" alt=\"4 squares within a square\" width=\"55\" height=\"55\" class=\"aligncenter size-full wp-image-161\" \/>\r\n\r\nHandyperson<\/td>\r\n<td>Some section authors used generative AI to review and edit human-written content.<\/td>\r\n<\/tr>\r\n<\/tbody>\r\n<\/table>\r\n&nbsp;","rendered":"<p>Jorge Gatica, Mandi Goodsett, Xiongyi Liu, Emily Rauschert, Patrick Wachira<\/p>\n<h2>INTRODUCTION<\/h2>\n<p>Generative AI has been adopted in many industries, and higher education faces increasing pressure to provide training and support for students to use this new technology in the workplace. However, while hype narratives often present this technology as inevitable and necessary, generative AI raises many ethical concerns that deserve careful consideration before it is used extensively. College campuses are ideal places to bring critical thinking, ethical frameworks, and evidence-based viewpoints to the development and use of this new technology, and students can practice critical thinking skills and learn more about AI through examining ethical concerns.<\/p>\n<p>Higher education institutions are, in some ways, directly threatened by the ethical issues that generative AI presents, especially when it comes to academic integrity and job replacement. College students\u200b are often aware of the ethical issues that generative AI presents, and <a href=\"https:\/\/sites.campbell.edu\/academictechnology\/2025\/03\/06\/ai-in-higher-education-a-summary-of-recent-surveys-of-students-and-faculty\/\">they express concern about AI\u2019s impact<\/a> on their ability to get jobs, on the environment, and on creativity more broadly. At the same time, <a href=\"https:\/\/www.insidehighered.com\/news\/students\/academics\/2025\/08\/29\/survey-college-students-views-ai\">surveys show<\/a> that college students continue to use generative AI for schoolwork in large numbers, sometimes in ways that undermine their learning. To remain viable, and to graduate students who can be responsible AI users, higher ed institutions must critically examine and effectively communicate the ethical threats that generative AI poses.<\/p>\n<h2>FOUNDATIONS<\/h2>\n<h3><strong>Bias in AI in Higher Education<\/strong><\/h3>\n<h4>Defining Bias in AI Systems<\/h4>\n<p>Before surveying how bias manifests across higher education, it is worth considering what &#8220;bias&#8221; actually means in this context, since the term is used loosely both in popular discourse and in some of the scholarly literature. In machine learning, statistical bias refers to a systematic deviation between a model&#8217;s predictions and the true values it is trying to estimate \u2014 a technical property of an estimator that can exist even in a perfectly &#8220;fair&#8221; system. Social bias, by contrast, refers to the systematic disadvantage or unfair treatment of particular groups, typically along lines of race, gender, disability, language background, or socioeconomic status. The two are related but distinct: a model can be statistically unbiased in aggregate while still producing socially biased outcomes for specific subgroups, and this distinction is important for understanding how institutions diagnose and respond to problems when they arise.<\/p>\n<p>It is also useful to distinguish bias from simple errors. Random error \u2014 noise that affects all groups roughly equally \u2014 is a nuisance but not, on its own, an equity problem. Bias is what remains when error is not random: when a system is reliably worse for some groups than others. When considering AI ethics, it is critical to be aware of this, as several of the tools discussed here do not merely make mistakes \u2014 they make patterned, predictable mistakes that fall more heavily on some students than others.<\/p>\n<h4>Causes: Where Bias Enters the Pipeline<\/h4>\n<p>Bias is rarely introduced at a single point. It typically accumulates across a pipeline, and understanding where it originates matters because different causes call for different remedies.<\/p>\n<ul>\n<li><strong>Data-level causes.<\/strong> Most AI systems used in higher education are trained on historical institutional data \u2014 past admissions decisions, past grades, past enrollment and retention patterns. Because that history reflects real-world inequities (e.g., in access, in resourcing, in whose success the institution was set up to support), models trained on it tend to reproduce those inequities rather than correct for them. Underrepresentation compounds this: when a subgroup is thin in the training data, the model has less signal to learn from for that group, and its predictions for that group become noisier and less reliable.<\/li>\n<li><strong>Design-level causes.<\/strong> Even with balanced data, design choices can introduce bias. Developers must choose which variables (features) to include, and even when protected characteristics like race or gender are explicitly excluded, other variables can act as <em>proxies<\/em> \u2014 zip code standing in for race, first-generation status standing in for class background \u2014 carrying much of the same discriminatory signal under a different name. Models are also typically optimized to maximize average accuracy across the entire dataset, which can mean sacrificing accuracy for smaller subgroups in exchange for better overall performance.<\/li>\n<li><strong>Human and institutional causes.<\/strong> Someone decides what counts as the &#8220;correct&#8221; label during training \u2014 what counts as a &#8220;successful&#8221; student, a &#8220;well-written&#8221; essay, &#8220;suspicious&#8221; exam behavior \u2014 and those judgments encode the assumptions, and sometimes the unconscious biases, of the people making them.<\/li>\n<li><strong>Deployment-context causes.<\/strong> A model validated on one population can perform very differently when applied to a different one. A tool built and tested on a large, well-resourced research university&#8217;s data may behave unpredictably at an under-resourced regional or community college with a different student population \u2014 yet vendor tools are frequently marketed and deployed across very different institutional contexts with little re-validation.<\/li>\n<li><strong>Feedback loops.<\/strong> Perhaps the most insidious cause is not a one-time input error but a self-reinforcing cycle: a biased prediction leads to a biased intervention (or lack of one), which shapes the student&#8217;s subsequent outcomes, which then becomes new training data that confirms the model&#8217;s original \u2014 biased \u2014 prediction.<\/li>\n<li><strong>Unequal access as an upstream cause.<\/strong> A related but distinct cause, discussed in more depth below, is that students and faculty do not begin with equal access to AI tools in the first place. Because that unequal access shapes the very outcomes \u2014 grades, retention, engagement \u2014 that later become training data for the predictive systems discussed throughout this section, today&#8217;s access gap can become tomorrow&#8217;s algorithmic bias.<\/li>\n<\/ul>\n<h4>Types of Bias<\/h4>\n<p>Building on these causes, the technical literature identifies several recurring <em>types<\/em> of bias relevant to educational AI: <strong>representation bias<\/strong> (some groups are simply underrepresented in the data used to build the system); <strong>measurement bias<\/strong> (the proxies used to measure a construct, such as &#8220;engagement&#8221; or &#8220;risk,&#8221; do not mean the same thing across groups); <strong>aggregation bias<\/strong> (a single model is applied uniformly to a heterogeneous population when subgroup-specific models would perform better); <strong>evaluation bias<\/strong> (the benchmarks used to validate a model&#8217;s fairness are themselves unrepresentative); and <strong>historical bias<\/strong> (the model faithfully learns a pattern from the past that was itself unjust). These categories recur throughout the use cases surveyed below.<\/p>\n<h3>Survey of Use Cases<\/h3>\n<h4>Student-Facing and Administrative Use Cases<\/h4>\n<p><strong>Admissions algorithms.<\/strong> Several institutions now use predictive models to screen or rank applicants, estimating the likelihood that an applicant will enroll, succeed, or persist. Because these models are trained on historical admissions and outcome data, they risk encoding the same racial, socioeconomic, and geographic patterns that shaped who was admitted \u2014 and who succeeded \u2014 in the past, effectively automating and obscuring decisions that would draw far more scrutiny if made explicitly by a human admissions officer.<\/p>\n<p><strong>Automated essay scoring and writing assessment.<\/strong> Automated scoring systems, increasingly used for placement testing and large-enrollment writing courses, have been shown to systematically disadvantage certain groups of writers. Research on automated scoring of English-language learners has documented what researchers term <em>bias amplification<\/em>: because high-scoring responses from English Language Learners are comparatively rare in training data, models trained through standard methods tend to favor linguistic patterns typical of non-ELL writers, and as a result systematically under-predict scores for ELL students even when their responses demonstrate comparable underlying knowledge (Wang et al., 2026). Critically, this study found that the resulting prediction gap between groups can be <em>larger<\/em> than the gap already present in the training data \u2014 the model does not merely inherit the disparity; it magnifies it.<\/p>\n<p><strong>AI-text detection and plagiarism tools.<\/strong> Perhaps the most widely reported bias finding in this space concerns AI-generated-text detectors. In an influential Stanford study, detectors correctly classified essays written by native English-speaking eighth graders with near-perfect accuracy but incorrectly flagged more than half of TOEFL essays written by non-native English speakers as AI-generated (Liang, Yuksekgonul, Mao, Wu, &amp; Zou, 2023). The likely mechanism is that many detectors rely on &#8220;perplexity&#8221; \u2014 a measure of how predictable a text&#8217;s word choices are \u2014 and non-native writing tends to be less lexically varied, which these tools mistake for the low-perplexity, formulaic patterns characteristic of AI output. Subsequent reporting found the same pattern held for widely used commercial tools deployed at scale in higher education, raising the possibility that international and multilingual students face disproportionate risk of false accusations of academic dishonesty (Garc\u00eda Mathewson, 2023).<\/p>\n<p><strong>Learning analytics and at-risk prediction.<\/strong> These systems flag students believed to be at risk of failing or withdrawing so that institutions can direct advising or support resources toward them. AI systems predict student risk based on historical data, which often reflects and replicates systemic biases like lower graduation rates for marginalized or low-income groups. As a result, the algorithm can unfairly tag these students as &#8220;high risk,&#8221; leading to harmful stereotyping, discouraged students, or the misallocation of support resources. Ultimately, this turns past inequalities into self-fulfilling prophecies instead of objective academic measures. A real-world example can be found in the case study in this chapter.<\/p>\n<p><strong>Chatbots and advising tools.<\/strong> As AI chatbots are adopted for academic advising and student services, early evidence suggests the quality, tone, and accuracy of responses can vary depending on how a student&#8217;s question is phrased or how the underlying model perceives the student&#8217;s background \u2014 an area that remains comparatively understudied relative to other use cases but merits ongoing attention as adoption expands.<\/p>\n<p><strong>Proctoring software.<\/strong> Automated remote-proctoring tools that use facial detection to verify student identity and flag &#8220;suspicious&#8221; behavior during exams have produced some of the most direct and well-replicated evidence of racial bias in educational AI. A 2022 study analyzing a major proctoring platform found that students with darker skin tones, and Black students specifically, were significantly more likely to be flagged for instructor review due to lower facial-detection rates \u2014 and, examining the intersection of race and gender, that women with the darkest skin tones were flagged far more often than any other group (Yoder-Himes et al., 2022). This mirrors earlier, foundational work by Buolamwini and Gebru (2018) showing that commercial facial-recognition systems were most accurate for lighter-skinned men and least accurate for darker-skinned women. Beyond skin tone, disability advocates have documented that movement-flagging features in proctoring software routinely misclassify the involuntary movements of neurodivergent students, or interruptions from caregiving responsibilities, as evidence of cheating (Surveillance and Disability in Online Proctored Exams, 2025).<\/p>\n<h4>Academic Writing Process<\/h4>\n<p>Beyond assessment, AI tools are now embedded directly in how students and scholars <em>produce<\/em> academic writing, which introduces a related but distinct set of concerns:<\/p>\n<ul>\n<li><strong>Grammar and writing assistants<\/strong> (Grammarly, Word&#8217;s AI features, ChatGPT-based editors) are trained predominantly on standard academic English and can flag dialect features and translanguaging practices \u2014 for instance, elements of African American Vernacular English \u2014 as errors, effectively nudging all writers toward a single normative register regardless of rhetorical intent or cultural context.<\/li>\n<li><strong>AI-generated writing feedback in composition courses<\/strong> raises a related concern: automated feedback tools may work best for students already fluent in dominant academic conventions, potentially widening rather than narrowing the gap for multilingual and first-generation writers the tools are often marketed as helping.<\/li>\n<li><strong>Voice and style homogenization<\/strong> is a broader worry as AI drafting and editing tools become routine: heavy reliance on these tools risks flattening diverse rhetorical traditions \u2014 including non-Western argumentative structures \u2014 into a single expected shape. (A related but distinct form of homogenization, affecting the <em>substance<\/em> of research findings rather than writing style, is discussed separately.)<\/li>\n<li><strong>Access disparities<\/strong> in premium AI writing tools add a further, more mundane inequity: students and institutions with resources to pay for higher-tier tools gain an assistance advantage that is not available to everyone, layering a new gap on top of existing ones.<\/li>\n<\/ul>\n<h4>Academic Research Process<\/h4>\n<p>A parallel set of concerns arises once AI tools move from supporting students&#8217; coursework to supporting scholars&#8217; own research process:<\/p>\n<ul>\n<li><strong>AI literature-review and discovery tools<\/strong> (e.g., AI-powered search and summarization tools built on academic databases) tend to be biased toward English-language, Global North, and highly cited sources, systematically underrepresenting scholarship published outside those channels. While this was already a problem with research tools before AI, AI tools can exacerbate it.<\/li>\n<li><strong>Citation recommendation systems<\/strong> risk reinforcing existing citation bias, in which already-prominent scholars and institutions receive disproportionate visibility \u2014 a dynamic with documented gender and racial dimensions in the pre-AI literature that automated recommendation could easily entrench further.<\/li>\n<li><strong>AI-assisted peer review<\/strong>, an emerging practice in which reviewers or editors use AI tools to screen or summarize submissions, carries a distinct risk: non-native English phrasing could be misread by these tools as an indicator of lower quality, and unconventional methodologies or topics underrepresented in training data could be undervalued relative to mainstream approaches.<\/li>\n<li><strong>Research summarization and &#8220;AI research assistant&#8221; tools<\/strong> raise a subtler concern: what a tool treats as an &#8220;authoritative&#8221; source shapes what a novice researcher \u2014 particularly a graduate student new to a field \u2014 encounters first, potentially steering them toward an already-mainstream evidence base rather than genuinely comprehensive coverage.<\/li>\n<li><strong>Homogenization and the flattening of dissenting or extreme findings.<\/strong> A distinct and less-discussed risk runs in the opposite direction from the biases above: rather than favoring one perspective over another, AI tools trained and tuned to sound balanced, measured, and broadly acceptable can systematically smooth over genuine variation \u2014 including legitimate scientific disagreement, minority findings, and outlier results that a field may most need to reckon with. This tendency, termed <em>homogenization bias<\/em> in a recent synthesis spanning linguistics, psychology, and cognitive science, arises because models are optimized during training to favor patterns that are frequent and easily generalizable, which has the effect of smoothing over minority representations in their outputs (Cell Press\/Trends in Cognitive Sciences, 2026). Importantly, the same synthesis cautions that this narrowing does not converge on a genuinely neutral center \u2014 it converges on a center shaped by whichever populations and viewpoints are best represented in training data, meaning &#8220;balanced-sounding&#8221; AI output is not actually neutral, only differently biased. This has direct consequences for research synthesis specifically: a controlled study of LLM-generated summaries of scientific abstracts found a consistent pattern of overgeneralization, in which models broadened and flattened the scope of specific findings even when explicitly prompted to produce faithful, detailed summaries (PMC, 2025). A related account of AI-assisted science makes the underlying mechanism explicit: literature-summarization systems can omit negative or dissenting findings altogether and present genuinely mixed evidence as more settled and consistent than it actually is (Si et al., 2024; Chauhan, 2026). For a field that depends on preserving and engaging with disagreement \u2014 rather than averaging it away \u2014 this is a bias with few precedents in the pre-AI literature, and one that current fairness frameworks, built primarily to detect disadvantage to specific demographic groups, are not well equipped to catch.<\/li>\n<\/ul>\n<h4>AI in Data Generation and Analysis<\/h4>\n<p>This is perhaps the least examined \u2014 and most consequential \u2014 area of bias, since it touches the empirical core of the research process itself, upstream of any conclusion a scholar might draw.<\/p>\n<p><strong>Synthetic data generation.<\/strong> AI tools are increasingly used to generate synthetic datasets, whether to protect privacy, augment small samples, or simulate populations that are hard to reach. The central risk here is what researchers call <em>bias inheritance<\/em>: because large language models reflect the biases present in their own training data, synthetic data generated by these models can propagate and amplify those biases in ways that materially affect the fairness of anything subsequently trained or analyzed using that data (Li et al., 2025). This is not merely a theoretical concern. A study generating 140,000 synthetic health records across seven different large language models found that larger, more capable models actually exhibited <em>greater<\/em> demographic bias than smaller ones \u2014 some systematically over-representing White or Black patients while under-representing Hispanic and Asian patients relative to real population data (Huang et al., 2025). This is a genuinely counterintuitive finding worth foregrounding: a &#8220;better&#8221; model, by ordinary performance benchmarks, is not automatically a fairer one.<\/p>\n<p><strong>Synthetic survey respondents (&#8220;silicon sampling&#8221;).<\/strong> A newer and fast-growing practice involves prompting large language models to simulate the responses a person with a given demographic profile would give to a survey \u2014 intended as a faster, cheaper substitute for fielding real surveys. Researchers studying this practice caution that LLM outputs in this context are model-dependent artifacts rather than measurements of the world, and that treating them as real observations risks systematically misrepresenting marginalized groups \u2014 a concern that is especially serious for subgroup analysis, since model error can concentrate unevenly across demographic groups even when overall agreement with real survey data looks acceptable (Bisbee, Imai, Rosenman, &amp; Zhou, 2024). Large-scale empirical audits bear this out: one study comparing LLM-simulated respondents to a real national arts-participation survey found the synthetic respondents showed a systematic positive bias \u2014 expressing inflated &#8220;liking&#8221; \u2014 relative to actual human respondents (Karell, 2026). As one review summarized the underlying problem bluntly, language models do not represent a genuine sample of any real population; they represent patterns in training data that itself systematically underrepresents marginalized groups, non-English speakers, and offline populations (Verian Group, 2026).<\/p>\n<p><strong>AI-assisted qualitative coding and thematic analysis.<\/strong> Tools that use large language models to code interview transcripts or open-ended survey responses are being adopted rapidly for their efficiency gains, but the evidence on their reliability is mixed. One frequently cited study using ChatGPT and Llama 2 to code semantically complex interview data found systematic bias in the resulting outputs, with the potential to mislead subsequent interpretation, and recommended that AI coding be used to extend \u2014 never replace \u2014 analysis that begins with human researchers (Ashwin, Chhabra, &amp; Rao, 2023). A more applied account is specific about <em>how<\/em> this bias surfaces in practice: AI coding tools lack the contextual understanding needed for interpretive analysis, especially where the data involves humor, sarcasm, or coded language reflecting the cultural context or lived experience of marginalized participants \u2014 content that is easily misread by a system without that lived context (Child Trends, 2025).<\/p>\n<p><strong>AI-assisted statistical analysis and coding tools.<\/strong> AI systems that select statistical models, generate analysis code, or interpret results are becoming common in research workflows, and their errors have a distinctive property: they cascade. A validation study testing ChatGPT&#8217;s code-interpreter feature against real health data found that when the tool selected an incorrect statistical method, the errors it introduced propagated through every subsequent step of the analysis; conversely, when method selection was correct, the remaining steps were usually completed accurately (Ruta, Gaidici, Irwin, &amp; Lifshitz, 2025). This makes method selection the single highest-leverage point in the pipeline for bias or error to take hold \u2014 a useful concrete illustration of the &#8220;upstream&#8221; framing introduced in Section 1. A separate, more basic concern compounds this: empirical testing of ChatGPT&#8217;s code generation found that it is non-deterministic, capable of returning meaningfully different code for the exact same prompt, which undermines both the reliability of any single analysis and the reproducibility of research that depends on it (Ouyang, Wang, Li, &amp; Zhang, 2024). Finally, an equity dimension is worth naming explicitly: a study of introductory-lab students using AI code-interpretation tools found a real risk that less experienced users would be unable to identify errors in the tool&#8217;s output or know how to correct them \u2014 meaning the researchers and students least equipped to catch an AI analysis tool&#8217;s mistakes are often the least statistically sophisticated, which is itself an unequal exposure to the risks these tools introduce (Low &amp; Kalender, 2023).<\/p>\n<h4>AI-Generated Instructional Materials and Assessment<\/h4>\n<p><strong>Instructional materials.<\/strong> The clearest evidence here concerns AI-generated images used in course materials. A systematic review of 31 peer-reviewed studies on AI text-to-image tools in educational settings found that biased representation was pervasive across the literature: generated images disproportionately centered white, male, Western, thin, and non-disabled figures, while diversity in age, body type, and disability was largely absent (systematic review, <em>Computers and Education Open<\/em>, 2026 \u2014 author names could not be confirmed via available search tools and should be verified directly from the journal before citing). This is not merely an abstract pattern \u2014 a study using AI-generated illustrations drawn from a children&#8217;s book excerpt in focus groups with Latine undergraduates found that students actively noticed these representational cues, negotiating the tension between the stereotypes embedded in the images and their own lived experience rather than passively accepting the images as neutral (Rho &amp; Karumbaiah, 2026). Beyond images, a broader taxonomy developed for instructors names the range of biases text-generation tools can introduce into course materials, including cultural, demographic, ideological, and linguistic bias, noting that English and a handful of other languages dominate online training data and so are disproportionately well-represented in what generative AI tools produce \u2014 a concrete illustration of the &#8220;uneven quality across languages&#8221; concern this subsection raises for multilingual campuses (University of Kansas Center for Teaching Excellence, n.d.). Encouragingly, at least one study measuring student and pre-service teacher awareness found that a majority of respondents already recognized that generative AI tools can reproduce gender and cultural stereotypes, suggesting critical AI literacy instruction has a meaningful foundation to build on (Ribes-Lafoz, Navarro-Colorado, &amp; Rovira-Collado, 2026).<\/p>\n<p><strong>AI-generated assessment.<\/strong> Automatically generated exam questions raise a parallel but distinct concern, rooted in the psychometric concept of <em>differential item functioning<\/em> (DIF) \u2014 a pattern in which students from different groups who have the same underlying ability nonetheless have different probabilities of answering an item correctly, indicating the item itself, not the students&#8217; knowledge, is the source of the disparity. A student-centered study of an LLM-generated question set applied exactly this kind of analysis and found that some AI-generated exam items showed evidence of bias linked to unmeasured subgroup characteristics \u2014 such as differential familiarity with how a question happened to be worded \u2014 and concluded such items should be flagged for review before any high-stakes use (Student-Centered LLM Q&amp;A Study, 2025). The same study named the specific mechanism worth flagging for readers: AI-generated distractors (the incorrect answer options in multiple-choice items) can be culturally loaded, and item phrasing may inadvertently privilege students who are already comfortable with chatbot-style language over those who are not. A related methods paper aimed at instructors using AI to generate exam variants makes the same point more generally, noting that bias in AI-generated items can surface subtly \u2014 in the examples or cultural framing a tool defaults to \u2014 which is why the paper recommends pairing human-in-the-loop review with formal DIF analysis as a matter of course, rather than treating AI-generated items as ready to use as-is (Educational Sciences, 2025).<\/p>\n<h4>Institutional Adoption and Procurement of AI Tools<\/h4>\n<p>A different point of entry for bias, distinct from any single tool design, is the process by which institutions decide to adopt one AI tool over another \u2014 or adopt one at all. This matters because procurement decisions are frequently made under conditions that make it hard to catch bias before a tool reaches thousands of students: under time pressure, with limited technical expertise on the evaluating committee, and often without the leverage to demand the kind of transparency an independent audit requires.<\/p>\n<p><strong>AI-detection tools as a case in point.<\/strong> The rollout of AI-text detection illustrates the problem concretely. When Turnitin added an AI-detection feature to its existing plagiarism-checking product in April 2023, the feature was enabled for existing institutional customers with less than 24 hours&#8217; notice, no option to disable it at the time, and no meaningful insight into how the underlying model worked (Vanderbilt University, 2023). In other words, thousands of institutions found themselves using a bias-prone detection tool not because they evaluated and selected it, but because it arrived bundled inside a product they had already procured for an unrelated purpose \u2014 a mode of adoption that bypasses the kind of vetting a standalone purchasing decision would normally require. Vanderbilt subsequently disabled the feature after months of testing and consultation, citing exactly these concerns (Vanderbilt University, 2023), and the University of Pittsburgh&#8217;s teaching center reached a similar conclusion, stating explicitly that it had concluded current AI-detection software was not reliable enough to deploy without substantial risk of false positives, and disabled the tool campus-wide as a result (University of Pittsburgh Teaching Center, n.d.). Not every institution reached the same conclusion, however: an investigative analysis of purchasing records found institutions renewing AI-detection subscriptions year after year despite documented flaws in the technology and known privacy concerns about the vendor&#8217;s growing database of student papers, suggesting that inertia and faculty demand for a bright line on academic integrity can outweigh the accumulating evidence of a tool&#8217;s unreliability (CalMatters\/The Markup, 2025).<\/p>\n<p><strong>Structural barriers to good procurement decisions.<\/strong> This unevenness across institutions is not simply a matter of some campuses caring more than others. A 2025 review of AI procurement practices across higher education found that technology purchasing today routinely involves IT, cybersecurity, legal, privacy, and academic stakeholders, and that the resulting review process \u2014 while thorough on security and compliance \u2014 often leaves equity and bias review as an afterthought relative to those better-established criteria (EDUCAUSE Review, 2025). The same review reported that procurement leaders most wanted external standards or frameworks to lean on, since few institutions have the in-house technical capacity to independently evaluate a vendor&#8217;s bias-testing claims \u2014 a gap that tools like EDUCAUSE&#8217;s Higher Education Community Vendor Assessment Tool (HECVAT) are beginning to address, though bias and fairness questions remain less standardized within it than security and privacy questions. This connects directly to the vendor-opacity pattern: an institution cannot demand evidence of fairness testing it does not know to ask for, and a vendor has limited incentive to volunteer testing that might complicate a sale.<\/p>\n<p><strong>Accountability without eliminating bias.<\/strong> A further concern, raised directly by AI-governance researchers working with higher education leaders, is that automating part of a decision does not remove human bias from the process \u2014 it can instead remove the accountability that previously made a biased decision correctable. Framed as a caution for procurement committees themselves: adopting an AI tool to make an institutional process more &#8220;objective&#8221; can create the appearance of neutrality while quietly making it harder to identify who is responsible when the tool&#8217;s output turns out to be biased after all (Changing Higher Ed, 2026).<\/p>\n<h4>Inequality in AI Access<\/h4>\n<p>A related but conceptually distinct concern is not how AI tools behave once in use, but who gets to use them at all. This is generally not bias in the technical sense but rather an access and resourcing gap that researchers have increasingly labeled the <em>AI divide<\/em>, explicitly building on the older concept of the digital divide (bioRxiv, 2024). The distinction matters for this section for a specific reason: unequal access today becomes a cause of algorithmic bias tomorrow, since the advantages or disadvantages it produces feed directly into the institutional outcome data (e.g., grades, retention, &#8220;risk&#8221; scores) that later trains the predictive models.<\/p>\n<p><strong>Student access.<\/strong> The scale of disparity is substantial and measurable. A narrative review of recent literature found that only 27% of rural students had access to devices compatible with generative AI tools, compared to 70% of urban students, and that students with stronger digital competencies obtained up to 60% greater academic benefit from the same tools than students without those competencies (Springer, 2026a). This is not simply a matter of institutions failing to act: survey data reported by Inside Higher Ed found that half of chief technology officers said their institution does not grant students institutional access to generative AI tools at all \u2014 tools that, when institutionally licensed, are often free to the student and more capable and secure than whatever a student might access on their own (Inside Higher Ed, 2025). The same reporting noted that more than half of students said most or all of their instructors prohibit generative AI use outright, meaning that even students at well-resourced institutions may face inconsistent access depending on individual faculty policy rather than any institutional standard.<\/p>\n<p><strong>Beyond access: a second-order divide in literacy and confidence.<\/strong> Even where access exists, researchers have found a second divide around who is equipped to use these tools effectively. A qualitative study of academic staff at Norwegian universities concluded that new digital divides could emerge not from unequal access to technology itself, but from unequal possession of the AI literacy, critical judgment, and prompting skill required to benefit from it \u2014 and that this divide appears not only between groups of students, but between entire academic disciplines (Springer, 2026b). A large single-institution survey similarly found that although the large majority of students were broadly familiar with generative AI concepts, only about a quarter were actually using these tools for academic work, and roughly three-quarters had received no formal classroom instruction on how to use them at all (DeStefano, Hackney, &amp; Moskal, 2026) \u2014 suggesting that awareness of a tool&#8217;s existence is a poor proxy for a student&#8217;s actual capacity to use it well.<\/p>\n<p><strong>Faculty, not only students.<\/strong> The access gap extends to faculty as well, with direct consequences for the fairness of the tools students encounter. Faculty digital-access gaps have long been documented as a standing concern in educational technology generally (PMC, 2025), and if faculty themselves have uneven access to, or training in, AI tools, that unevenness shapes which students benefit from AI-assisted instruction, feedback, and advising in the first place \u2014 compounding, rather than independently sitting alongside, the student-facing access gap described above.<\/p>\n<p>A 2025 policy commentary makes the throughline explicit: without deliberate institutional and policy intervention, unequal access to generative AI risks concentrating educational advantage among students and institutions that already have the most resources, producing what the author terms a systemic divergence in educational outcomes rather than a marginal difference in classroom experience (Wong, 2025). This is precisely the mechanism described as a <em>feedback loop<\/em>: today&#8217;s access gap shapes tomorrow&#8217;s grades and retention data, which becomes the training data for the predictive systems \u2014 meaning that closing the access gap is not a separate equity initiative from addressing algorithmic bias, but one of the more effective upstream interventions available for preventing it.<\/p>\n<h3>Case Study: Racial Bias in Community College Early-Alert Prediction<\/h3>\n<h4>Background<\/h4>\n<p>Predictive analytics for identifying &#8220;at-risk&#8221; students has become one of the most widely adopted AI applications in American higher education, frequently implemented through partnerships with private vendors. Georgia State University&#8217;s partnership with the vendor EAB is probably the most publicized example: the university reported a meaningful increase in its graduation rate after adopting predictive analytics to identify struggling students and direct advising resources toward them, a result that has been widely cited as a success story for the technology, though it coincided with several other institutional changes (Swaak, 2022, as cited in AIR, 2023). Many other institutions, including Temple University, have built similar &#8220;early alert&#8221; systems, most commonly used to flag students at risk of dropping out before completing their degree so that advisors can reach out proactively (Bird, Castleman, Mabel, &amp; Song, 2021).<\/p>\n<h4>The Study<\/h4>\n<p>Economists Kelli Bird, Benjamin Castleman, and Yifeng Song conducted one of the most rigorous independent audits of this class of system, examining two prediction models used at the community-college level: one predicting course completion and one predicting degree completion \u2014 the two outcomes most commonly targeted by early-alert systems nationally (Bird, Castleman, &amp; Song, 2024).<\/p>\n<h4>The Finding<\/h4>\n<p>The central finding is direct: if either model were used, as intended, to target additional advising or support resources toward &#8220;at-risk&#8221; students, the resulting algorithmic bias would mean fewer marginal Black students received those resources than a fair system would allocate to them (Bird, Castleman, &amp; Song, 2024). This is not simply a case of the model being &#8220;less accurate&#8221; for Black students in the abstract \u2014 it is a bias in the <em>allocation<\/em> of a scarce, real-world resource, occurring at exactly the decision threshold institutions use to determine who gets help.<\/p>\n<p>Notably, the bias was not evenly distributed across the risk distribution \u2014 it concentrated precisely where institutional decision-making tends to draw the line. With the degree-completion model, the magnitude of bias was found to be several times higher when institutions defined &#8220;at-risk&#8221; using only the bottom decile of predicted scores than when they used a broader bottom-half cutoff (Bird, Castleman, &amp; Song, 2023). This is a critical operational detail: colleges with limited advising capacity are pushed, for resource reasons, toward exactly the narrow cutoffs that maximize this disparity.<\/p>\n<h4>Mechanism<\/h4>\n<p>The likely cause is not a maliciously chosen variable but a more structural asymmetry: the researchers&#8217; findings suggest algorithmic bias arises in part because the administrative data institutions already have on hand is less predictive of Black students&#8217; success than of White students&#8217; success, particularly for new students \u2014 implying that additional, more informative data collection could help narrow the gap (Bird, Castleman, &amp; Song, 2024).<\/p>\n<h4>Structural Context and Mitigation<\/h4>\n<p>This case also illustrates the governance concerns raised throughout this section. Because most predictive analytics tools in higher education are proprietary, institutions adopting them typically have little visibility into how the underlying model was built or validated (Bird, Castleman, Mabel, &amp; Song, 2021), which means that most colleges using an early-alert system have no equivalent, independent bias audit of their own vendor&#8217;s tool. The Bird, Castleman, and Song study is notable precisely because it was conducted by researchers with no stake in the tool&#8217;s commercial success \u2014 a form of scrutiny that vendor&#8217;s opacity otherwise forecloses. Without independent auditing and data-sharing requirements built into procurement, institutions have limited means of discovering whether a tool they have adopted in good faith is quietly reallocating support away from the students who need it most.<\/p>\n<h3>Cross-Cutting Patterns of AI Biases in Higher Education<\/h3>\n<p>Across the use cases surveyed above, several patterns recur often enough to be treated as structural features of AI bias in higher education, rather than as isolated problems specific to any one tool.<\/p>\n<p><strong>Proxy variables.<\/strong> Across admissions, learning analytics, and writing assessment alike, models rarely need an explicit protected variable to reproduce discrimination \u2014 geography, prior coursework, or linguistic patterns frequently function as effective proxies for race, class, or national origin.<\/p>\n<p><strong>Feedback loops.<\/strong> The at-risk prediction case study illustrates a dynamic visible elsewhere as well: a biased prediction shapes a biased intervention (or its absence), which shapes real student outcomes, which then becomes the next round of training data \u2014 quietly hardening the original bias into something that looks, on the surface, like objective evidence.<\/p>\n<p><strong>Opacity in vendor tools.<\/strong> From proctoring software to predictive analytics to AI-text detectors, the tools shaping high-stakes decisions in higher education are overwhelmingly proprietary. Institutions frequently cannot see the training data, validation methodology, or model architecture behind a tool they have licensed, which makes independent verification \u2014 of the kind performed in the case study above \u2014 the exception rather than the norm.<\/p>\n<p><strong>Disproportionate impact on intersectional groups.<\/strong> AI bias in higher education disproportionately impacts on certain groups across various use cases: Black students in early-alert and proctoring systems; non-native English speakers and international students in AI-text detection and writing-assessment tools; first-generation and lower-income students in access to premium AI writing and research tools.<\/p>\n<p><strong>Homogenization.<\/strong> Rather than disadvantaging one group relative to another, several tools discussed above (e.g., writing assistants, research-summarization tools) share a tendency to smooth over genuine variation altogether \u2014 in rhetorical style, in scientific findings, in perspective \u2014 converging on outputs that read as neutral but are shaped by whichever patterns are best represented in training data. This pattern is harder to detect with the subgroup-based auditing methods that address the other four patterns, since the harm is not unequal treatment of identifiable groups but a general narrowing of what gets represented or preserved at all.<\/p>\n<h3>Institutional and Ethical Stakes<\/h3>\n<p>The stakes of these patterns are not only ethical but also legal and operational. In the United States, biased outcomes tied to race, national origin, or disability status can implicate Title VI of the Civil Rights Act and the Americans with Disabilities Act, and institutions that deploy opaque, vendor-controlled tools without independent validation face growing legal and regulatory exposure as scrutiny of these systems increases \u2014 a concern the U.S. Government Accountability Office has itself flagged with respect to predictive analytics in higher education (Bauman, 2022, as cited in AIR, 2023). At the same time, institutions face genuine pressure to adopt these tools: advising staff and writing-center resources are finite, and AI tools promise to extend limited institutional capacity across more students than human staff alone could serve. This creates an unresolved tension between equity and efficiency \u2014 the goal of using AI to serve more students can come into direct conflict with the goal of serving them fairly, particularly when speed of adoption outpaces the institution&#8217;s capacity to audit what it has adopted. Consequently, this tension can play out in institutional procurement decisions.<\/p>\n<h3>Mitigation Strategies<\/h3>\n<p>While no single intervention addresses bias across every stage of the AI pipeline, we choose to summarize a few AI bias mitigation strategies that have the greatest potential in higher education.<\/p>\n<p><strong>Technical approaches<\/strong> include fairness-aware machine learning methods designed to explicitly balance accuracy across subgroups rather than optimizing only for aggregate performance, and pre-procurement bias auditing, in which institutions require a vendor&#8217;s tool to be tested for subgroup disparities before adoption rather than after a problem is reported. The community-college case study above illustrates the value of this kind of audit \u2014 and how rarely it happens absent independent researcher access.<\/p>\n<p><strong>Policy and governance approaches<\/strong> include establishing institutional AI governance committees with the authority to review and approve (or decline) AI tools before deployment; contractual requirements for vendor transparency about training data and validation methods; and human-in-the-loop requirements that prevent any single AI output \u2014 an at-risk flag, a proctoring alert, a detector score \u2014 from triggering a consequential decision without a human reviewer able to exercise independent judgment.<\/p>\n<p><strong>Pedagogical approaches<\/strong> address the problem from the other direction: building AI and data literacy among faculty and administrators so that they are equipped to question vendor claims, interpret model outputs critically, and recognize when a tool&#8217;s assumptions do not fit their student population \u2014 competencies that are, at present, unevenly distributed across the institutions adopting these tools fastest.<\/p>\n<p><strong>Access and infrastructure approaches<\/strong> address the upstream cause directly: institutional licensing that grants all students and faculty a baseline level of AI access, rather than leaving access to individual ability to pay; device and connectivity support for students without reliable access outside campus; and integrating AI literacy instruction into the curriculum itself rather than assuming students arrive with it.<\/p>\n<h3>Open Questions and Future Directions<\/h3>\n<p>Several questions remain open. First, generative AI-specific bias in AI tutors, writing assistants, and the data-generation and analysis tools is a newer and comparatively understudied frontier relative to the more established literature on predictive analytics and proctoring software, and is likely to grow more consequential as these tools become further embedded in both teaching and research workflows. Second, the same personalization that makes AI tools appealing \u2014 tailoring feedback, recommendations, or interventions to an individual student or researcher \u2014 is difficult to distinguish from stereotyping along demographic lines when the underlying model&#8217;s &#8220;personalization&#8221; is itself learned from biased historical data. Finally, responsibility for addressing these harms remains contested and diffuse. It is not obvious whether the vendor building the tool, the institution deploying it, the instructor or advisor using its output, or the researcher relying on its analysis bears primary responsibility when bias causes harm \u2014 and until that question is resolved with greater institutional clarity, accountability is likely to continue falling through the gaps between these parties.<\/p>\n<h3>AI and Environmental Impact<\/h3>\n<p>As generative AI models have exploded into daily life, public awareness has evolved from viewing AI as a clean, virtual &#8220;cloud&#8221; technology to recognizing it as a highly resource-intensive physical industry.<\/p>\n<p>The prevailing public sentiment toward AI\u2019s environmental impact is one of worry rather than optimism.<\/p>\n<h4>Fear of net harm<\/h4>\n<p>The public is twice as likely to believe that AI will ultimately do more to hurt the environment than help it. Strikingly, people express greater concern about AI\u2019s environmental consequences than about other notoriously carbon-heavy industries, such as air travel, cryptocurrency mining, and meat production (t\u00a0he AP-NORC at the University of Chicago, apnorc.org).<\/p>\n<p>Certainly, the environmental footprint of AI-related data centers is growing rapidly, inviting frequent comparisons to heavy industrial sectors. While both AI data centers and large petrochemical plants have massive environmental footprints, they operate on completely different paradigms: AI data centers primarily generate indirect (Scope 2) environmental pressure driven by massive electricity and water consumption (ONU <a href=\"https:\/\/unu.edu\/inweh\/collection\/environmental-cost-of-AIs-Enrgy-Use-Carbon-water-and-land-footprints\"><em>Environmental Cost of Artificial Intelligence)<\/em><\/a>. Petrochemical plants inflict direct (Scope 1) industrial degradation through the chemical processing of fossil fuels into plastics, resins, and synthetic materials (Yan et al., 2024)<\/p>\n<h4>Carbon Footprint and Greenhouse Gas Emissions<\/h4>\n<p>The mechanism and scale of emissions differ significantly between the two sectors.<\/p>\n<table class=\"grid aligncenter\" style=\"height: 307px;\">\n<thead>\n<tr style=\"height: 18px;\">\n<th style=\"height: 18px; width: 132.031px;\" scope=\"col\"><strong>Impact Category<\/strong><\/th>\n<th style=\"height: 18px; width: 395.766px;\" scope=\"col\"><strong>AI-Related Data Centers<\/strong><\/th>\n<th style=\"height: 18px; width: 264.922px;\" scope=\"col\"><strong>Large Petrochemical Plants<\/strong><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr style=\"height: 124px;\">\n<td style=\"height: 124px; width: 132.031px;\"><strong>Primary Emission Source<\/strong><\/td>\n<td style=\"height: 124px; width: 395.766px;\"><strong>Indirect (Scope 2)<\/strong><\/p>\n<p>Electricity drawn from regional power grids to run chips and cooling infrastructure.<\/td>\n<td style=\"height: 124px; width: 264.922px;\"><strong>Direct (Scope 1)<\/strong><\/p>\n<p>On-site fossil fuel combustion, cracking furnaces, and chemical process venting.<\/td>\n<\/tr>\n<tr style=\"height: 92px;\">\n<td style=\"height: 92px; width: 132.031px;\"><strong>Global Sector Emissions<\/strong><\/td>\n<td style=\"height: 92px; width: 395.766px;\">Global data centers collectively emit <strong>between 190\u2013208 million metric tons of <\/strong>carbon dioxide (CO2) <strong>annually<\/strong> (AI tasks represe190 and 208 million metric tons of carbon dioxide (CO2) annually (AI tasks account for 0% of this total).<\/td>\n<td style=\"height: 92px; width: 264.922px;\">The chemical and petrochemical sectors worldwide emit roughly <strong>1.3 to 1.5 billion metric tons of <\/strong>carbon dioxide (CO2) <strong>annually<\/strong>.<\/td>\n<\/tr>\n<tr style=\"height: 73px;\">\n<td style=\"height: 73px; width: 132.031px;\"><strong>Decarbonization Path<\/strong><\/td>\n<td style=\"height: 73px; width: 395.766px;\"><strong>Highly adaptable:<\/strong> Can theoretically achieve net-zero if the regional electrical grid transitions entirely to renewable energy.<\/td>\n<td style=\"height: 73px; width: 264.922px;\"><strong>Hard to abate:<\/strong> Carbon is inherently baked into the raw materials (crude oil, natural gas) and high-heat chemical processes.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>The Petrochemical Baseline:<\/strong> A single large petrochemical facility can emit several million tons of carbon dioxide (CO2) directly from its cracking stacks, making the baseline sector&#8217;s carbon footprint significantly larger than that of the AI sector.<\/p>\n<p>When people express concern about AI, public anxiety usually zeroes in on the physical infrastructure required to keep the technology running:<\/p>\n<ul>\n<li><strong>Strain on local power grids:<\/strong> Public awareness of massive data centers has grown, and these facilities demand large amounts of electricity, much of which is still generated by fossil fuels.<\/li>\n<li><strong>The AI Data Center Dilemma:<\/strong> While tech \u201chyperscalers\u201d contract heavily for green energy via Market-Based Carbon Certificates, their continuous 24\/7 power draw forces local grids to rely on fossil-fuel backup power. Projections indicate AI data center expansion will add up to 44 million metric tons of new carbon dioxide (CO2) annually by 2030.<\/li>\n<li><strong>Water depletion:<\/strong> Public pushback has grown in communities hosting these facilities, as people realize data centers require millions of gallons of water daily just to keep servers cool. Water supply may deplete municipal water resources when these centers are built in water insecure locations (cf. <a href=\"https:\/\/sf.tradepub.com\/?pt=adv&amp;page=Data%20Center%20World\">Data Center World).<\/a><\/li>\n<li><strong>Lack of government oversight:<\/strong> A significant majority of the public believes that the government and tech regulatory bodies are not doing enough to address or mitigate AI\u2019s environmental harms. State legislation and initiatives differ widely across the US (see below for OH references).<\/li>\n<\/ul>\n<h4>The &#8220;Sustainability Paradox&#8221;<\/h4>\n<p>Despite these worries, public perception remains conflicted (cf. Parker, 2026). People acknowledge a &#8220;sustainability paradox&#8221;: they are uncomfortable with the massive carbon and water footprints of training models, yet they remain hopeful that traditional machine learning can be used to model climate patterns, optimize renewable energy grids, and track deforestation.<\/p>\n<p>The exact resource metrics consumed by a single AI prompt depend heavily on the complexity of the model (e.g., standard text vs. image generation), hardware efficiency, and data center cooling methods. Official data published by major AI providers\u00a0 and independent researchers break down the baseline metrics for a single text-based prompt:<\/p>\n<h4>Core Resource Consumption Metrics<\/h4>\n<h4>(per AI user\u2019s prompt, O\u2019Donnell and Crownhart, 2025; Ritchie, 2025)<\/h4>\n<table class=\"grid aligncenter\">\n<thead>\n<tr>\n<th style=\"text-align: center;\" scope=\"col\"><strong>Metric<\/strong><\/th>\n<th style=\"text-align: center;\" scope=\"col\"><strong>Google Gemini (Median)<\/strong><\/th>\n<th style=\"text-align: center;\" scope=\"col\"><strong>OpenAI ChatGPT (Baseline)<\/strong><\/th>\n<th style=\"text-align: center;\" scope=\"col\"><strong>Independent\/Multi-Model Estimates<\/strong><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>Electricity<\/strong><\/td>\n<td>0.24 watt-hours (Wh)<\/td>\n<td>0.34 watt-hours (Wh)<\/td>\n<td>0.3 to 0.5 watt-hours (Wh)<\/td>\n<\/tr>\n<tr>\n<td><strong>Water<\/strong><\/td>\n<td>0.26 milliliters (mL)<\/td>\n<td>0.32 milliliters (mL)<\/td>\n<td>10 to 40 milliliters (mL)<\/td>\n<\/tr>\n<tr>\n<td><strong>Carbon Emissions<\/strong><\/td>\n<td>0.03 grams of CO\u2082eq<\/td>\n<td><em>Not officially disclosed<\/em><\/td>\n<td>0.03 to 0.1 grams of CO\u2082eq<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>Metric Breakdown and Real-World Equivalents<\/h4>\n<h5>Electricity (0.24 to 0.5 Watt-hours)<\/h5>\n<p><strong>What it means:<\/strong> The actual physical computing power required by AI accelerators (like Nvidia GPUs or Google TPUs) to process your prompt. <strong>Real-world equivalent:<\/strong> Running a standard kitchen microwave for about <strong>one second<\/strong> or powering a highly efficient LED lightbulb for <strong>approximately two minutes<\/strong>.<\/p>\n<h5>Water (0.26 to 40 Milliliters)<\/h5>\n<p>The sharp difference in water data comes down to methodology. Tech giants like Google and OpenAI disclose <em>direct operational consumption<\/em>\u2014the exact amount of clean water that evaporates on-site through cooling towers to prevent servers from overheating during that specific calculation. Independent researchers consider the total ecological footprint, factoring in the large volume of water used upstream by the power plants that generate the data center&#8217;s electricity.<\/p>\n<p><strong>Real-world equivalent:<\/strong> Low-end official estimates are equal to about <strong>5 drops of water<\/strong>. High-end independent tracking equates a short back-and-forth conversation (20 to 50 prompts) to about <strong>500 mL<\/strong>, or a standard plastic water bottle (Kelly et al., 2026).<\/p>\n<h5>Carbon Emissions (0.03 to 0.1 grams of CO\u2082eq)<\/h5>\n<p>In environmental impact quantification, <strong>Carbon Dioxide Equivalents (expressed as CO\u2082e or CO\u2082eq)<\/strong> serve as the universal &#8220;currency&#8221; for measuring and comparing the climate impact of various greenhouse gases (GHGs). Instead of tracking multiple pollutants such as methane, nitrous oxide, and fluorinated gases, CO\u2082eq bundles them into a single, standardized metric.<\/p>\n<p>The math behind CO\u2082eq relies on a multiplier factor called <strong>Global Warming Potential (GWP)<\/strong>. GWP measures how much heat a specific greenhouse gas traps in the atmosphere over a set period (usually a 100-year time horizon) relative to carbon dioxide (CO\u2082).<\/p>\n<h5>Key Factors That Change these Metrics<\/h5>\n<ul>\n<li><strong>Task Complexity:<\/strong> Generating a high-resolution AI image or video requires vastly more math than text. Creating a single AI image consumes roughly <strong>2 to 5 liters of water<\/strong>, several times the footprint of a simple text interaction.<\/li>\n<li><strong>Model Size &amp; Architecture:<\/strong> Advanced &#8220;reasoning&#8221; models or deeply layered architectures (like DeepSeek R1) can consume up to <strong>30+ Wh per query<\/strong> because they perform billions more mathematical operations before spitting out an answer.<\/li>\n<li><strong>Geography and Climate:<\/strong> A prompt routed to a data center in a cool climate utilizing outdoor air cooling uses almost no local water. The same prompt routed to a data center in a hot, arid climate relying on evaporative cooling towers uses significantly more.<\/li>\n<\/ul>\n<h4>Corporate initiatives to achieve net-zero AI infrastructure<\/h4>\n<p>Tech giants are facing a major dilemma: the massive AI computing boom has sent greenhouse gas emissions soaring, throwing corporate sustainability timelines off track. Reports reveal recent year-over-year emission spikes of 25% at Google, 23% at Microsoft, and 64% at Meta.<\/p>\n<p>Here, \u201chyperscalers\u201d refers to massive technology companies that dominate the global cloud computing, data storage, and digital infrastructure industries. The term &#8220;hyperscale&#8221; refers to their ability to seamlessly and massively scale a computer network\u2014adding thousands of servers and massive amounts of storage\u2014to meet geometric increases in data demand.<\/p>\n<h5>1. The Nuclear Power Pivot<\/h5>\n<p>Tech companies are rapidly moving beyond traditional solar and wind power to secure baseload clean electricity that runs 24\/7, leading a historic corporate shift toward nuclear energy:<\/p>\n<ul>\n<li><strong>Microsoft:<\/strong> Signed a massive 20-year power purchase agreement to back the commercial restart of the 835-megawatt <a href=\"https:\/\/introl.com\/blog\/nuclear-power-ai-data-centers-microsoft-google-amazon-2025\">Three Mile Island Nuclear Plant<\/a>.<\/li>\n<li><strong>Google:<\/strong> Placed a historic order for a fleet of up to 500 megawatts of Small Modular Reactors (SMRs) from Kairos Power to power its data centers starting by the end of the decade.<\/li>\n<li><strong>Amazon:<\/strong> Invested over $20 billion to scale and convert the Susquehanna nuclear site into a completely nuclear-powered AI data center campus.<\/li>\n<li><strong>Meta:<\/strong> Teamed up with companies like Vistra, TerraPower, and Oklo to target up to 4 gigawatts of nuclear power for its <a href=\"https:\/\/www.cnbc.com\/2026\/01\/09\/meta-signs-nuclear-energy-deals-to-power-prometheus-ai-supercluster.html\">Prometheus AI Supercluster<\/a> in Ohio.<\/li>\n<\/ul>\n<h5>2. Gigawatt-Scale Battery Storage and 24\/7 Carbon-Free Energy<\/h5>\n<p>Because wind and solar are intermittent, tech companies are installing unprecedented levels of utility-scale energy storage. This allows them to capture clean energy during the day and discharge it when the grid is strained.<\/p>\n<ul>\n<li><strong>The Goal:<\/strong> Companies are transitioning from &#8220;annual carbon matching&#8221; to true 24\/7 hourly matching, ensuring that every watt of power drawn by an AI accelerator is zero-carbon in real time.<\/li>\n<li><strong>The Scale:<\/strong> Massive utility-scale projects, such as Amazon&#8217;s co-located 400 MWh Megapack storage facilities, have made the data center market the largest single driver of <a href=\"https:\/\/www.facebook.com\/chichubkidaps\/posts\/-americas-ai-data-center-boom-is-driving-unprecedented-battery-storage-deploymen\/1031704529384819\/\">grid battery storage deployment<\/a> in the US.<\/li>\n<\/ul>\n<h5>3. Chip and Hardware Efficiency Gains<\/h5>\n<p>Engineers are trying to outpace the growing demands for computational power by significantly optimizing chip performance and life cycle.<\/p>\n<ul>\n<li><strong>Custom AI Accelerators:<\/strong> Google\u2019s eighth-generation custom AI chips (TPU 8t and 8i) deliver up to a <a href=\"https:\/\/sustainability.google\/\">2x improvement in performance per watt<\/a> compared to previous iterations.<\/li>\n<li><strong>Lifecycle Extension:<\/strong> Hyperscalers are implementing aggressive hardware circularity initiatives. Extending GPU and server life cycles from 3 years to 5\u20137 years drastically lowers the &#8220;embodied carbon&#8221; generated during manufacturing and concrete-heavy facility construction.<\/li>\n<\/ul>\n<h5>4. Cross-Industry Tech Alliances<\/h5>\n<p>Rather than building isolated solutions, the tech sector is co-funding collaborative infrastructure innovations:<\/p>\n<ul>\n<li><strong>Data Center Innovation Initiative:<\/strong> Amazon, Google, Meta, and Microsoft have jointly launched a funding coalition with groups like Breakthrough Energy. They are deploying up to $5 million per startup to test <a href=\"https:\/\/sustainabilitymag.com\/news\/the-four-tech-giants-funding-low-carbon-data-centre-startups\">sustainable data center innovations<\/a> inside active, working facilities.<\/li>\n<li><strong>Carbon Dioxide Removal (CDR):<\/strong> Tech companies are leveraging long-term advance market commitments to fund massive direct-air carbon capture startups. This private-sector backing helps commercialize carbon removal technologies to offset unavoidable Scope 3 supply chain emissions (<a href=\"https:\/\/carboncredits.com\/data-center-giants-enter-carbon-credit-market-as-hyperscalers-fuel-a-new-green-tech-gold-rush\/\">CarbonCredits.com<\/a>).<\/li>\n<\/ul>\n<h4>Curving the High Demand for Fresh Water or the Effects of Water Pollution<\/h4>\n<p>The rapid rise of generative AI has significantly increased the heat density within server racks. Traditional data center cooling relies on open-loop <strong>evaporative cooling towers<\/strong>. These systems reject heat by constantly evaporating millions of gallons of treated freshwater into the atmosphere every day.<\/p>\n<p>To curb this massive environmental drain, the data center industry is undergoing a comprehensive infrastructure overhaul. They are shifting from cooling whole rooms of hot air to managing thermal loads directly at the chip level.<\/p>\n<h5>1. The &#8220;Hot Tub&#8221; Fluid Breakthrough (Warm-Water Closed Loops)<\/h5>\n<p>The most significant industry shift involves raising the operating temperature of the liquid coolant itself. Nvidia&#8217;s latest AI infrastructure architecture standardizes to <strong>100% closed-loop liquid cooling<\/strong>. The coolant (typically a mix of 75% water and 25% propylene glycol) enters the server rack at <strong>45\u00b0C (113\u00b0F)<\/strong>\u2014hotter than a residential hot tub\u2014and can absorb heat up to 55\u00b0C (131\u00b0F).<\/p>\n<p>Because the coolant is so hot, it creates a massive temperature differential with the surroundings (outside air). This allows the heat to be dissipated into the environment using simple outdoor <strong>dry coolers<\/strong> (giant radiator coils) instead of evaporative chillers. The system is filled once during construction and continuously recirculated, cutting local facility cooling water usage from 2.6 million gallons per megawatt per year down to <strong>near zero<\/strong>.<\/p>\n<h5>2. Direct-to-Chip (DTC) Cold Plate Architecture<\/h5>\n<p>As AI chips surpass 1,000+ watts each, pushing cold air across them is no longer physically viable. Direct-to-Chip (DTC) systems attach copper micro-channel <strong>cold plates<\/strong> directly to the physical surface of the GPU or TPU. The dielectric fluid or water mix flows through these microscopic channels, directly capturing up to 98% of the processor&#8217;s heat at the exact source of emission.<\/p>\n<p>Major tech hyperscalers are standardizing DTC designs for all new AI builds. By localizing the heat transfer, companies can bypass massive air-conditioning infrastructure entirely, allowing facilities to maintain ultra-low Water Usage Effectiveness (WUE) metrics without relying on municipal water loops.<\/p>\n<h5>3. Two-Phase Immersion Cooling<\/h5>\n<p>For extreme AI rack densities (approaching 100kW+ per rack), immersion cooling eliminates the need for physical water plumbing inside the server. The entire server chassis is fully submerged in a bath of specially engineered, non-conductive <strong>dielectric fluid<\/strong>. In a <em>two-phase<\/em> system, the fluid has a low boiling point.<\/p>\n<p>As the AI chips run hot, the fluid boils, vaporizes, rises to a condenser coil at the top of the sealed tank, condenses back into a liquid, and falls back down. This creates a completely hermetic, self-contained thermodynamic cycle. The external loop that cools the condenser coil can rely entirely on dry-air heat exchangers, resulting in zero evaporative water loss in the data center&#8217;s mechanical operations.<\/p>\n<h5>4. Eco-Chilling and Non-Potable Circularity<\/h5>\n<p>Where liquid cooling is not fully implemented, and air cooling must be augmented with water, tech giants are altering their water sourcing: Companies are investing heavily in on-site water reclamation plants. Instead of pulling clean drinking water from local aquifers, data centers use raw industrial wastewater or sewer lines. They use internal <strong>reverse osmosis filtration<\/strong> to purify the water to a level suitable for mechanical use, while keeping potable drinking water in the community.<\/p>\n<h4>The Unseen Bottleneck: &#8220;Indirect&#8221; Footprints<\/h4>\n<p>While these closed-loop facility designs effectively address on-site water issues, sustainability researchers point out a major loophole: <strong>Scope 2 indirect water consumption<\/strong>.<\/p>\n<p>If a &#8220;zero-water&#8221; data center draws its electricity from a local power grid reliant on coal, nuclear, or natural gas, thousands of gallons of water are still being evaporated upstream at the utility power plant to generate that electricity. This is why tech companies are coupling their liquid-cooling architectures with direct solar and wind installations and advanced battery installations to ensure the entire supply chain becomes water-neutral.<\/p>\n<h3>OH, Legislation and Regulations about large Data Centers<\/h3>\n<h4>The Tax Exemption Backlash and Public Revenue Loss<\/h4>\n<p>Since 2011, Ohio has aggressively courted tech giants by offering <strong>100% sales and use tax exemption<\/strong> on data center server equipment, infrastructure, and construction. However, the cost of these subsidies has exploded far beyond what the state originally anticipated.<\/p>\n<ul>\n<li><strong>The $1.5 Billion Revenue Gap:<\/strong> In 2024, the tax break cost Ohio $555 million. By 2025, that figure snowballed to <strong>nearly $1.6 billion in foregone revenue<\/strong>\u2014more than 11 times the initial estimates provided by the <a href=\"https:\/\/tax.ohio.gov\/\">Ohio Department of Taxation<\/a>.<\/li>\n<li><strong>The Impact on Local Public Services:<\/strong> An additional $166.8 million in local county sales taxes was lost. This massive shortfall removes funds that would otherwise support local public schools, infrastructure repairs, and community services, shifting the financial burden back onto Ohio taxpayers.<\/li>\n<li><strong>The Legislative Freeze:<\/strong> In response to ballooning costs, Governor Mike DeWine issued an <strong>executive order pausing all new data center sales tax exemptions<\/strong>. Meanwhile, state lawmakers introduced bills such as <a href=\"https:\/\/fox8.com\/news\/ohio-bill-would-end-data-center-tax-breaks-effective-oct-1\/\">Ohio House Bill 975<\/a> was created to permanently eliminate all sales and use tax exemptions for large-scale data centers, while <a href=\"https:\/\/www.legislature.ohio.gov\/legislation\/136\/hb646\">Ohio House Bill 646<\/a> aims to reduce the 100% exemption to 50% permanently.<\/li>\n<\/ul>\n<p>Existing long-term tax deals locked in by giants like Amazon, Meta, and Google through statewide agreements, however, will remain active.<\/p>\n<p>For real-time legislative trackers on active statehouse bills (such as HB 646, HB 784, or SB 381), you can audit the docket on the official <a href=\"https:\/\/www.occ.ohio.gov\/data-centers\">Ohio Consumers&#8217; Counsel Data Centers Page.<\/a><\/p>\n<h3>AI and Demand and Depletion of Minerals (Rare Earths)<\/h3>\n<p>While everyday operations consume electricity and water, the upfront manufacturing and long-term hardware replacement cycles tie AI directly to the environmental and geopolitical crises of critical mineral extraction. (<a href=\"https:\/\/fpanalytics.foreignpolicy.com\/2025\/07\/18\/artificial-intelligence-critical-minerals-supply-chains\/\">Artificial Intelligence and the Critical Minerals Crunch<\/a>). To build the specialized hardware required for deep learning, manufacturers rely heavily on a specific suite of critical minerals:<\/p>\n<ul>\n<li><strong>Gallium and Germanium:<\/strong> Highly integrated into high-speed semiconductor transistors and the fiber-optic networks that link data centers.<\/li>\n<li><strong>Neodymium and Dysprosium:<\/strong> Vital rare earth minerals used to manufacture high-strength permanent magnets for data center cooling fans, hard disk drives, and robotic actuators.<\/li>\n<li><strong>Lithium and Cobalt:<\/strong> The raw materials required for the massive industrial battery arrays used as backup power to protect data centers from power grid failures.<\/li>\n<\/ul>\n<h4>Long-Term Ecological Damage from &#8220;AI Mining&#8221;<\/h4>\n<p>The environmental cost of acquiring these minerals differs structurally from that of traditional mining, as rare earths are rarely found in concentrated veins. They require destructive, chemically intensive processes to refine.<\/p>\n<h4>The Electronic Waste (E-Waste) Crisis<\/h4>\n<p>The short lifespan of AI infrastructure accelerates the mineral depletion crisis. Because AI models advance so quickly, the specialized GPUs that run them become obsolete in roughly <strong>three to five years<\/strong> (<a href=\"https:\/\/blog.citp.princeton.edu\/2025\/10\/15\/lifespan-of-ai-chips-the-300-billion-question\/\">Center for Information Technology Policy (CITP) at Princeton University).<\/a><\/p>\n<h4>Structural and Supply Chain Bottlenecks<\/h4>\n<p>The long-term outlook is further complicated by geographic concentration. The <a href=\"https:\/\/www.iea.org\/\">International Energy Agency (IEA)<\/a> notes that China accounts for approximately 60% of global rare earth extraction and over 90% of global rare earth<strong> refining capacity<\/strong>. As Western tech giants build massive server farms, the strain on this highly centralized mineral supply chain is driving a frantic rush to open new, ecologically disruptive mines in parts of Africa, South America, and the Arctic.<\/p>\n<h3><strong>Copyright \/ Intellectual Property <\/strong><\/h3>\n<p>The intersection of Generative AI and copyright is one of the most contentious ethical issues surrounding this new technology, especially among artists and others who make a living on creating and selling their creative works.<\/p>\n<p>In this chapter, we are primarily referring to copyright as it exists in the United States. According to the US Copyright Office, copyright is \u201ca type of intellectual property that protects original works of authorship as soon as an author fixes the work in a tangible form of expression\u201d (<a href=\"https:\/\/www.copyright.gov\/what-is-copyright\/\">copyright.gov, n.d.<\/a>). For a work to be protected by copyright, it must be sufficiently original (i.e. different from other, published works) and fixed in a tangible form of expression (i.e. not an idea). Although it was required in the past, copyright protection does not require the creator to register the work with the Copyright Office or take any other formal action. Just fixing the original work in a tangible form of expression is sufficient to gain copyright protection.<\/p>\n<p>Copyright protects the copyright holder\u2019s ability to make copies of the work, make derivatives, distribute unpublished copies, perform the work, or display the work. The length of copyright protection varies depending on whether the copyright holder is an individual or an organization, when the work was published, and whether the work was published. Cornell University Library maintains an excellent <a href=\"https:\/\/guides.library.cornell.edu\/copyright\/publicdomain\">copyright term chart<\/a>, but if in doubt about whether something is still protected by copyright, try contacting your librarian or campus legal counsel.<\/p>\n<h4>Training LLMs and Copyright<\/h4>\n<p>There are two considerations when it comes to copyright and generative AI. One is the intellectual property that was used to train large language models (LLMs), and the other is the outputs of the LLMs. In the case of training data, there is concern that the creative works of authors, artists, graphic designers, photographers, and others were used without permission as part of the training that make LLMs possible. AI companies are profiting from the use of this content without providing any compensation to the creators that make this wealth possible.<\/p>\n<p>To train a large language model, billions of source works must be ingested (this is true for LLMs that create language, as well as those that create images and\/or video). When an LLM creates an output, it isn\u2019t replicating any particular work from its training data; it\u2019s using predictive algorithms to determine what content to produce. This makes it nearly impossible to give attribution to individual works used to produce an LLM output. As of this writing, courts have not definitively determined if using copyrighted works in the training of LLMs is a copyright violation. Anthropic, for example, was able to rely on a fair use argument for some of the works that it used to train Claude models (<a href=\"https:\/\/www.insidetechlaw.com\/blog\/2025\/09\/bartz-v-anthropic-settlement-reached-after-landmark-summary-judgment-and-class-certification\">they were not able to rely on fair use for pirated works, however<\/a>). Despite the outcome of the Anthropic case, many legal cases (<a href=\"https:\/\/www.mishcon.com\/generative-ai-intellectual-property-cases-and-policy-tracker\">around 80 at the time of this writing<\/a>) about the use of copyrighted works for LLM training are still ongoing.<\/p>\n<p>It is worth noting that some creative professionals have used specially designed software to make their published works unrecognizable by generative AI, as a way of undermining these tools. One such app, <a href=\"https:\/\/nightshade.cs.uchicago.edu\/whatis.html\">Nightshade<\/a>, was developed by researchers at the University of Chicago. Nightshade allows image creators to make their images look different to a generative AI tool that is undergoing training (although the change is not detectable by human viewers). Over time, the LLM begins to learn to assign incorrect metadata for the images it is training on, because the human-visible image doesn\u2019t match the AI-visible one. This both protects artists and gives them the ability to reduce the effectiveness of tools that they perceive as predatory.<\/p>\n<h4>LLM Outputs and Copyright<\/h4>\n<p>When it comes to LLM outputs, the concern is that the tools can be used to create works that replace the work of existing artists. In fact, some have even used LLMs to create art \u201cin the style of\u201d an existing, professional artist, effectively destroying that artist\u2019s ability to make a living. In rare cases, LLMs can be used to make almost exact replicas of existing, copyrighted works which, in other contexts, is a clear copyright violation.<\/p>\n<p>In general, a work produced solely by an LLM is not protected by copyright, because only works created by humans can be copyright-protected. However, a work that has some AI content and some human content can be protected, with the extent of protection being worked out in courts on a case-by-case basis (<a href=\"https:\/\/www.copyright.gov\/newsnet\/2025\/1060.html\">Copyright Office, n.d.<\/a>). Also, if an LLM output is similar to an existing copyrighted work or character, it could be a copyright violation. Some AI companies shield users from this kind of liability (and many have taken steps to disallow intentional outputs of this kind), but individuals should check the user agreement for the AI tool they are employing.<\/p>\n<h3>Privacy &amp; Security<\/h3>\n<p>The primary concern when it comes to generative AI and privacy is the fact that AI companies use personal data (among other data) to train LLMs, exposing that data to risk of data breaches and also allowing LLMs to use the data in chatbot outputs to other users. In the initial training of LLMs, AI companies use huge amounts of scraped data from the internet that often includes personal data collected without consent. The LLM training process is, overall, not very transparent and, some have argued, not sufficiently regulated (<a href=\"https:\/\/www.bbc.com\/news\/technology-65139406\">McCallum, 2023<\/a>). In Europe, there is some regulation of generative AI as it relates to privacy, but regulation is much laxer in the United States. There is new research showing that, even if a user is careful not to input a lot of data into an AI chatbot, <a href=\"https:\/\/cyberscoop.com\/ai-deanonymization-risks-online-anonymity-study\/\">LLMs are able to use existing information on the internet to identify individuals<\/a>, sometimes in great detail (<a href=\"https:\/\/arxiv.org\/pdf\/2602.16800\">Lerman et al, 2026<\/a>).<\/p>\n<p>While any use of generative AI tools involves some privacy risk, there are ways to improve the protection of your private information while using the tools. For CSU employees, using Microsoft Copilot available through our enterprise license does provide some assurances of privacy and security. There are also generative AI tools that don\u2019t require a login, therefore preserving some of your anonymity, such as <a href=\"https:\/\/duck.ai\/\">Duck.ai<\/a> and <a href=\"https:\/\/chat.mistral.ai\/chat\">Mistral AI (Le Chat)<\/a>. In general, the best way to preserve your privacy with these tools is to refrain from entering any sensitive data into them in the first place.<\/p>\n<h3>Human Labor<\/h3>\n<p>While AI has emerged as a transformative force in the labor market, it has also raised profound ethical challenges which arise mainly from the way the systems are developed, designed, deployed, and used. In the realm of human labor, these ethical challenges include: displacement of jobs, algorithmic bias in hiring, low-wage and traumatic work environments, and intrusive workplace surveillance.<\/p>\n<h4>Automation and jobs displacement<\/h4>\n<p>With its ability to automate repetitive tasks, there is concern that AI could eventually replace millions of jobs, especially white-collar entry-level positions. A report by McKinsey Global Institute suggests that AI could automate up to 30% of hours currently worked across the US economy by 2030. Further, a study by Stanford and the Digital Economy Lab found a 16% decline in early-career employment within the most AI-exposed fields.<\/p>\n<p>AI\u2019s impact on the job market may disproportionately affect certain demographics such as minority workers who are overly represented in positions that are at a higher risk of automation, potentially exacerbating existing inequalities. Other demographic groups with slower adaptive capacity, such as older workers who may have fewer transferable skills, will also be affected. AI also is expected to impact the labor market across genders differently. The International Labor Organization (ILO) predicts that 7.8% of women&#8217;s occupations in high income countries could be automated, totaling around 21 million jobs, compared to only 2.9% of jobs held by men.<\/p>\n<p>These impacts on the labor market are expected to create significant challenges, including loss of income for displaced workers and exacerbating unemployment numbers. It is important to note, however, that the effect of AI on employment is not even across different sectors and skill sets. Certain industries will see minimal disruption, whereas others will experience substantial workforce displacement.<\/p>\n<h4>Algorithmic bias in hiring<\/h4>\n<p>Employers are increasingly using AI for recruiting, hiring, and performance evaluation. There is concern related to algorithmic bias in using AI to screen, evaluate, and make decisions on potential employees during the hiring process.\u00a0 Ethical challenges arise in that AI algorithms have the potential to inherit and reinforce existing biases based on the training data it receives. The automated tools used to screen applicants&#8217; resumes could unfairly filter out qualified candidates based on historical data. This will potentially lead to an exacerbation of existing discriminatory practices.<\/p>\n<h4>Low wage and traumatic work environments<\/h4>\n<p>For its underlying functionality such as data labeling\/annotation and content moderation and reviewing, AI depends on a large amount of human labor. These labor roles are often outsourced to low-wage contract workers, mostly in developing countries, who feed data to and train the AI algorithms. Companies hire from poor and underserved communities, refugee populations, incarcerated people, and others with few job options. \u00a0A <em>CBS 60 Minutes <\/em>investigation titled <em>Humans in the Loop, for<\/em> instance, showed that the AI data labeling work takes a toll on these workers as they often work long hours with low pay (as little as $2 per hour gross). To minimize their costs, the giant tech companies do not hire directly; instead, they subcontract these workers via digital platforms allowing companies to circumvent labor laws and benefits. There is little transparency, with the workers often not knowing which companies they are working for or which systems they are training, which raises the concern that they may be training, unknowingly, systems that may be used for surveillance and subsequent repression.<\/p>\n<p>The workers in content moderation are responsible for finding and flagging content that is deemed inappropriate, such as texts and imagery that contain hate speech, violence, abuse, sexually explicit or other types of harmful content. This work is critical to ensure that the AI can detect this harmful content. The workers exposed to these traumatic tasks often have little or no emotional or mental health support.<\/p>\n<h4>Workplace Surveillance<\/h4>\n<p>Another ethical concern relates to workplace surveillance and privacy. Many companies are now using AI to monitor employee performance and productivity. In the content moderation of work, speed and efficiency are prioritized, and workers are pressured to make decisions within seconds, measured against the pre-determined time every task should take. This means that they work under close surveillance and are punished if they deviate from their assigned tasks. These surveillance practices are also seen in other industries including warehouses work where workers fulfill online orders and delivery work. Delivery drivers, for instance, are monitored through automated surveillance systems. Delivery time \u00a0expectations \u00a0are often unrealistic forcing many drivers to take risks to ensure that they deliver all the packages assigned to them within a specific time. Surveillance, tracking and productivity monitoring thus creates a high-pressure working environment which put workers at the risk of\u00a0 \u00a0injury or death. These practices of monitoring also can overstep into worker privacy violations.<\/p>\n<p><em>Human distillation<\/em><\/p>\n<p>Another area of ethical concern regarding human labor is the process often referred to as \u201chuman distillation.\u201d This practice involves using human intelligence, decision-making strategies, and feedback as data for training AI models, effectively transferring aspects of human expertise into artificial systems. The process of human distillation creates several concerns when it comes to human labor. One of these is that it raises fears that employees can be replaced after their operational knowledge is extracted in what might be termed as the \u201ctrain your replacement\u201d paradox. This means that workers are essentially building the software that might make their roles redundant. There are also concerns over intellectual property rights regarding the professional skills used in training AI models.<\/p>\n<p>Key methods in human distillation include Human-in-the-Loop (HITL), RLHF (Reinforcement Learning from Human Feedback), and synthetic data generation. These approaches map human preferences, steps, or rules into model training data to ensure AI aligns with human logic and values. To translate human knowledge into AI training, the following techniques are primarily used: Human-in-the-Loop (HITL) distillation: This involves human experts reviewing, correcting, and refining AI-generated responses in real-time. The corrected, higher-quality data is then used to train and fine-tune smaller, more efficient AI models. Reinforcement Learning from Human Feedback (RLHF): Humans rate and rank different AI responses. This data is used to create a reward model that teaches the AI which types of answers are preferred, aligning it with human ethics and conversational preferences. Distilling step-by-step: Humans write out rationales or step-by-step instructions for completing a task. These step-by-step human paths serve as supervised data, allowing smaller models to perform complex reasoning without needing massive datasets. Synthesizing from Human Rules: AI models can be trained to follow specific rule-based logic systems originally created by humans, distilling human judgment and uncertainty parameters directly into mathematical scoring<\/p>\n<h3><strong>Power<\/strong><\/h3>\n<p>The combination of the abovementioned ethical concerns can lead to considerable disparity in access to the benefits of AI, as well as inequalities in experiencing negative impacts of AI. \u00a0For example, who has quality access to AI, and what content is represented, can exacerbate existing power dynamics (Furze 2023).<\/p>\n<h4>Who benefits from AI?<\/h4>\n<p>AI developments in recent years have involved the concentration of power in terms of who controls how AI exists and how it might exist in the future. The companies leading AI\u2019s growth are some of the wealthiest companies in the world. The large AI companies that are newer were developed with funding from companies that were already among the wealthiest; for example, OpenAI was mainly funded by Microsoft, while Anthropic received large investments from Google and Amazon. These investments have so far led to massive increases in the value of all of these companies.<\/p>\n<p>Bremmer and Suleyman (2025) argue that Big Tech companies can now wield power more like nation-states, but without the regulation, such as democratic accountability, that nation-states are typically subject to. This can undermine democratic norms that are widely accepted, such as the power of people in a democracy to self-regulate. In some cases, the amount of money spent on AI projects, such as the \u201cStargate\u201d project, can be similar to the annual spending even of wealthy nations like Australia (Furze 2026).<\/p>\n<p>At a more individual level, while AI is to some extent available to all, many users find their access to higher quality AI models be quite limited unless they purchase subscriptions. To some extent, power is held by those with the economic resources to purchase better AI access, and these benefits can lead to positive feedback loops further concentrating power.<\/p>\n<h4>Who experiences the negative impact of AI?<\/h4>\n<p>Environmental issues from the high carbon footprint of AI are likely to be unequally distributed. Bender et al. (2021) note that it is critical to consider the environmental costs of AI, which may hit poorer countries that are simultaneously less likely to benefit from AI.<\/p>\n<p>When AI is presented as inevitable, this can result in a gloomy outlook and negative thoughts. The perception of AI as unavoidable is further amplified by the AIs themselves and the companies producing them, producing a new kind of hegemony leading to further disparities (Furze 2023). The prospect of an AI-driven future can be quite disempowering, particularly for younger adults seeking to find careers in this rapidly changing landscape.<\/p>\n<h4>Educational considerations for power in AI<\/h4>\n<p>While the internet itself has already disrupted traditional information sources such as printed books and newspapers, AI will likely accelerate this process. Unfortunately, as mentioned above, the biases inherent in AI may limit the quality of this information for educational purposes. Many textbooks are offering AI as a component, and this can be a way to limit AI access to information that is (mostly) correct, as students use AIs as study tools.<\/p>\n<p>Teaching students about AI ethics will likely involve a conversation about the power dynamic resulting from or reinforced by AI. In some ways, AI&#8217;s connection to power is similar to other technologies, such as the internet or smartphones, which also have environmental and equitability issues, so it may be possible to find some ways to mitigate power concerns by examining what has helped mitigate problems with those technologies. For example, to some extent, support for public access to the internet in public libraries has helped with power issues with internet access. Policies around smartphone waste, including opportunities to recycle and reuse components, have also mitigated some environmental impacts.<\/p>\n<h2>SCENARIOS &amp; CONSIDERATIONS<\/h2>\n<p>Different disciplines may have stronger links to some components of AI ethical concerns.\u00a0 Here we suggest some case studies that might be explored with students in a variety of contexts.<\/p>\n<h3>Bias<\/h3>\n<h4>Scenario<\/h4>\n<p>Dr. Priya Chandrasekaran had four days before finals week to decide what to recommend, and she was fairly sure that whatever she recommended, someone would be angry.<\/p>\n<p>As coordinator of Halloway University&#8217;s &#8220;Statistics for the Social Sciences,&#8221; a required course taught in eleven sections each fall to roughly 900 students, Chandrasekaran had spent two years trying to solve a problem every large-enrollment course eventually runs into: consistency. With eleven sections taught by a mix of two tenure-track faculty, one lecturer, and eight adjunct instructors \u2014 several of them teaching at two other institutions to make ends meet \u2014 exam difficulty and grading standards had drifted noticeably across sections. Students compared notes; some sections&#8217; averages ran fifteen points higher than others on ostensibly the same material. The associate dean had flagged it as an equity issue in its own right: which section you happened to land in was affecting your GPA.<\/p>\n<p>The exam-generation tool. Last spring, Chandrasekaran piloted a solution: an AI-assisted exam-bank platform, ItemForge, that could generate large pools of statistically calibrated questions from a shared set of learning objectives, which instructors could then pull from to build individualized exams that were nominally equivalent in difficulty. Adjunct instructors, several of whom told Chandrasekaran privately that they simply did not have the unpaid hours to write high-quality exam questions from scratch every term, were enthusiastic. The two tenure-track faculty were more skeptical but agreed to a trial.<\/p>\n<p>A pattern in the data. After the midterm, one of the tenure-track faculty, Dr. Marcus Whitfield, noticed something while reviewing item-level statistics the platform provides: several ItemForge-generated word problems used names, contexts, and idioms (references to unfamiliar consumer brands, U.S. sports statistics, colloquial phrasing) that seemed, anecdotally, to trip up international and multilingual students more than domestic ones \u2014 not because the statistics were harder, but because parsing the scenario took longer or was ambiguous. When he cross-referenced item-level scores against the registrar&#8217;s (limited, self-reported) international-student flag, students in that group scored on average nine points lower specifically on ItemForge-generated items, but showed no significant gap on the traditionally-written items covering the same statistical concepts, which had been retained on the exam as a comparison set.<\/p>\n<p>A second, separate pattern. Independently, the course also uses an AI-text-detection tool, VerifyWrite, on a short take-home written-interpretation component, where students explain a statistical result in their own words. Several adjunct instructors have been referring unusually high numbers of multilingual and international students for academic-integrity review based on VerifyWrite&#8217;s flags. One adjunct, Dana Okafor, brought a specific case to Chandrasekaran: a student from Vietnam whose written English was, in Okafor&#8217;s judgment, clearly her own \u2014 grammatically imperfect but conceptually strong \u2014 but who had been flagged at &#8220;87% likely AI-generated&#8221; by the tool. Okafor did not refer the student for a hearing, trusting her own judgment over the tool&#8217;s, but admitted she wasn&#8217;t sure that was the &#8220;correct&#8221; or consistent thing to do, and that other instructors in the course might be handling identical situations differently.<\/p>\n<p>Why nobody adopted these tools carelessly. It would be easy to tell this case as a story about instructors blindly trusting software, but that isn&#8217;t quite what happened. Chandrasekaran had asked ItemForge&#8217;s sales representative directly, before the pilot, whether the platform&#8217;s items had been checked for cultural or linguistic bias; she was told the platform used &#8220;diverse item-writing guidelines&#8221; but was not shown any actual bias-testing data. She hadn&#8217;t pushed further, in part because the university&#8217;s procurement office had already approved ItemForge as a licensed vendor for an unrelated purpose (adaptive homework problem sets) the year before, which made it feel like a known quantity rather than a new risk. VerifyWrite, similarly, had been adopted university-wide two years earlier by the Provost&#8217;s office as part of a broader academic-integrity initiative; Chandrasekaran&#8217;s course didn&#8217;t choose it so much as inherit it, and individual instructors have no formal mechanism to opt out of using it, only informal discretion in how much weight to give a flag once it appears.<\/p>\n<p>The stakeholders, now. With finals in four days:<\/p>\n<ul>\n<li>The adjunct instructors, as a group, want to keep ItemForge \u2014 without it, several say they will have to write finals from scratch with almost no paid time to do it, and some worry that abandoning it now, days before finals, would be logistically chaotic.<\/li>\n<li>Dr. Whitfield wants ItemForge pulled immediately from the final exam, arguing that knowingly using items with a demonstrated score gap, even one term after finding it, is indefensible regardless of the inconvenience.<\/li>\n<li>Dana Okafor wants VerifyWrite flags removed from instructor view entirely for the written component, arguing that if instructors like her are already learning to disregard the tool&#8217;s judgment in favor of their own, the tool isn&#8217;t adding anything except risk and stress for the flagged students, many of whom hear about being &#8220;flagged&#8221; before any human review even happens.<\/li>\n<li>The Registrar&#8217;s office, asked about better demographic data to investigate the ItemForge pattern further, notes that the &#8220;international student&#8221; flag Whitfield used is a rough proxy \u2014 it doesn&#8217;t capture multilingual domestic students, who may face similar issues, and pulling more granular demographic data before finals would require a formal data request that typically takes three to four weeks.<\/li>\n<li>The Provost&#8217;s office, which owns the VerifyWrite contract university-wide, is not part of this course&#8217;s decision-making at all, and Chandrasekaran does not have the authority to disable VerifyWrite even if she wanted to \u2014 only to advise her instructors on how much weight to give its flags.<\/li>\n<\/ul>\n<p>Chandrasekaran has to send guidance to all eleven sections by Friday.<\/p>\n<table class=\"grid aligncenter\" style=\"border-collapse: collapse; width: 100%;\">\n<caption>Exhibit A: Midterm Item-Type Comparison (Fictional Data)<\/caption>\n<thead>\n<tr class=\"shaded\">\n<th style=\"width: 26.3139%; text-align: center;\" scope=\"col\">Student Group<\/th>\n<th style=\"width: 32.4452%; text-align: center;\" scope=\"col\">Avg. Score, Item Forge-Generated Items<\/th>\n<th style=\"width: 30.2555%; text-align: center;\" scope=\"col\">Avg. Score, Traditionally-Written Items<\/th>\n<th style=\"width: 10.9854%; text-align: center;\" scope=\"col\">Gap<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"width: 26.3139%;\">Domestic Students<\/td>\n<td style=\"width: 32.4452%;\">78%<\/td>\n<td style=\"width: 30.2555%;\">77%<\/td>\n<td style=\"width: 10.9854%;\">+1<\/td>\n<\/tr>\n<tr>\n<td style=\"width: 26.3139%;\">International Students (registrar flag)<\/td>\n<td style=\"width: 32.4452%;\">69%<\/td>\n<td style=\"width: 30.2555%;\">76%<\/td>\n<td style=\"width: 10.9854%;\">-7<\/td>\n<\/tr>\n<tr>\n<td style=\"width: 26.3139%;\">All students, combined<\/td>\n<td style=\"width: 32.4452%;\">77%<\/td>\n<td style=\"width: 30.2555%;\">77%<\/td>\n<td style=\"width: 10.9854%;\">0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<table class=\"grid aligncenter\" style=\"border-collapse: collapse; width: 100%; height: 127px;\">\n<caption>Exhibit B:Verify Write Flag Rates by Section (Fictional Data, This Term)<\/caption>\n<thead>\n<tr class=\"shaded\" style=\"height: 55px;\">\n<th style=\"width: 36.9708%; text-align: center; height: 55px;\" scope=\"col\">Instructor<\/th>\n<th style=\"width: 21.7883%; text-align: center; height: 55px;\" scope=\"col\">% of Written Components Flagged<\/th>\n<th style=\"width: 30.2555%; text-align: center; height: 55px;\" scope=\"col\">Instructor&#8217;s Self-Reported Override Rate<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr style=\"height: 18px;\">\n<td style=\"width: 36.9708%; height: 18px;\">Dr. Whitfield (tenure-track)<\/td>\n<td style=\"width: 21.7883%; height: 18px;\">4%<\/td>\n<td style=\"width: 30.2555%; height: 18px;\">Reviews every flag personally; overrides \u223c 60%<\/td>\n<\/tr>\n<tr style=\"height: 18px;\">\n<td style=\"width: 36.9708%; height: 18px;\">Dana Okafor (adjunct)<\/td>\n<td style=\"width: 21.7883%; height: 18px;\">11%<\/td>\n<td style=\"width: 30.2555%; height: 18px;\">Reviews every flag personally; overrides \u223c 70%<\/td>\n<\/tr>\n<tr style=\"height: 18px;\">\n<td style=\"width: 36.9708%; height: 18px;\">Adjunct Instructor C<\/td>\n<td style=\"width: 21.7883%; height: 18px;\">9 %<\/td>\n<td style=\"width: 30.2555%; height: 18px;\">Refers most flags to formal integrity review without independent assessment<\/td>\n<\/tr>\n<tr style=\"height: 18px;\">\n<td style=\"width: 36.9708%; height: 18px;\">Adjunct Instructor D<\/td>\n<td style=\"width: 21.7883%; height: 18px;\">3%<\/td>\n<td style=\"width: 30.2555%; height: 18px;\">Rarely checks the dashboard; flags mostly go unreviewed<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>Considerations<\/h4>\n<h5>Diagnosing the two problems<\/h5>\n<ol>\n<li>The ItemForge gap and the VerifyWrite gap are both examples of AI bias, but they arise from different parts of the AI pipeline. What is the difference between a tool that generates biased content and a tool that evaluates students in a biased way? Does that difference matter for how each problem should be fixed?<\/li>\n<li>Exhibit A shows no gap in the combined average \u2014 only once the data is disaggregated does the pattern appear. What does this suggest about the limits of the kind of monitoring most instructors realistically have time to do?<\/li>\n<\/ol>\n<h5>Weighing the stakeholders<\/h5>\n<ol>\n<li>The adjunct instructors&#8217; concerns are logistical and economic (unpaid labor, time pressure), not indifference to bias. How should Chandrasekaran weigh a genuine resource constraint against a documented equity problem? Is there a version of &#8220;pull ItemForge from finals&#8221; that doesn&#8217;t simply transfer the cost onto the instructors with the least power in the course?<\/li>\n<li>Dana Okafor already overrides most flags using her own judgment. Is an AI-detection tool that experienced instructors mostly learn to override still doing useful work, or has it just become a source of stress and risk for flagged students with no compensating benefit?<\/li>\n<\/ol>\n<h5>Deciding what to do next, with four days left<\/h5>\n<ol>\n<li>What should Chandrasekaran&#8217;s Friday guidance say about ItemForge for the final exam \u2014 keep it, drop it entirely, or something in between (e.g., keep it but require human review of any item involving cultural\/linguistic context)? What are you trading off with your answer?<\/li>\n<li>The Registrar&#8217;s better demographic data would take three to four weeks \u2014 well past finals. Should Chandrasekaran wait for better data before acting, or act now on the imperfect data she already has? What does your answer suggest about how much evidence an institution should require before treating a pattern as a real problem?<\/li>\n<\/ol>\n<h3>Environmental Impact<\/h3>\n<h4>Scenario: The Municipal Zoning and Local Utility Decision<\/h4>\n<p>A city council in a rural, developing region must vote on whether to approve a zoning permit for a tech giant to build a new 500-megawatt hyperscale AI data center campus. The tech giant promises a $5 billion local investment, hundreds of high-paying construction jobs, and a massive boost to the municipal tax base to fund local schools and roads.<\/p>\n<h4>Considerations<\/h4>\n<ul>\n<li>The proposed data center requires massive amounts of power, forcing the regional energy utility to build a new natural gas plant to meet demand.<\/li>\n<li>The facility\u2019s cooling system will draw heavily from the local aquifer, which supplies the town&#8217;s drinking water and agricultural irrigation.<\/li>\n<li>The council must decide if immediate economic growth justifies long-term ecological risks. Approving the project could strain the local power grid\u2014causing residential electricity bills to spike\u2014and permanently deplete the local water supply, threatening the region&#8217;s agricultural economy.<\/li>\n<\/ul>\n<h3>Copyright \/ Intellectual Property<\/h3>\n<h4>Scenario<\/h4>\n<p>Eliza is working on creating flyers for a fundraiser for the local food bank, but she\u2019s really struggling. Eliza knows she has very poor graphic design skills, but she doesn\u2019t know any graphic designers, and she needs to get these flyers up as soon as possible. She wants to support local graphic designers and their work, but she doesn\u2019t feel that she has the funds or time to hire someone. She decides to use generative AI to create the flyer, reasoning that her only alternative is the ugly flyer she has tried to make.<\/p>\n<h4>Considerations<\/h4>\n<ul>\n<li>When is it appropriate for someone to forgo paying a professional for their creative labor and instead use generative AI?<\/li>\n<li>What might be the societal outcomes if creative workers are frequently undervalued in favor of easy-to-access generative AI tools?<\/li>\n<li>What if Eliza had some money to pay for a professional graphic designer to help her, but would prefer to donate the money to the food bank&#8217;s drive? Where does her ethical obligation lie in this situation?<\/li>\n<\/ul>\n<h3>Privacy &amp; Security<\/h3>\n<h4>Scenario:<\/h4>\n<p>Gina is having some health issues and is embarrassed to talk about them with her doctor. She turns to a generative AI chatbot instead and is relieved to get some reasonable-sounding advice. Gina\u2019s mom warns her that generative AI tools train on the conversations that she has and might now have sensitive data about Gina in its training data. Gina is not worried about this and doubts anyone would be able to find the data and trace it back to her.<\/p>\n<h4>Considerations:<\/h4>\n<ul>\n<li>Does Gina\u2019s age matter in this scenario? How would knowing her age change your response to the scenario?<\/li>\n<li>What risks is Gina taking by entering personal data? What is she gaining in exchange? Is this exchange worth it, in your opinion?<\/li>\n<li>How could system-level change remove the incentive for Gina to use a generative AI chatbot for this purpose?<\/li>\n<\/ul>\n<h3>Human Labor<\/h3>\n<h4>Scenario:<\/h4>\n<p>Alani, a computer science student, learns that the generative AI model they use for class projects was trained on data labeled by low\u2011paid workers overseas. These workers spent hours tagging violent or explicit content to make the model \u201csafe.\u201d However, she reasons, now that the technology exists, there is no harm in making use of it \u2013 in fact, if people don\u2019t use the tool, these workers\u2019 suffering would be for nothing.<\/p>\n<h4>Considerations:<\/h4>\n<ul>\n<li>In contexts like this one, to what extent should individual people change their behaviors, as opposed to putting the blame\/responsibility on large corporations? What actions are most likely to make positive change and how does acting in alignment with values play a role?<\/li>\n<li>Much of our technology is only possible because of the exploitation of people in less advantaged parts of the world. Should we stop using technology for this reason?<\/li>\n<\/ul>\n<h4>Scenario:<\/h4>\n<p>A design major named Betsy uses an AI art generator to complete freelance logo commissions. The client is thrilled with the quick turnaround, but Betsy\u2019s classmate Aarathi argues that Betsy is undercutting human artists by relying on automation. Betsy continues to do this freelance work, insisting that she is simply using available tools efficiently, and that generative AI is the way of the future graphic design.<\/p>\n<h4>Considerations:<\/h4>\n<ul>\n<li>Is generative AI use inevitable in every field? What might be AI\u2019s role in professions that currently rely heavily on human creativity?<\/li>\n<li>Do individuals have an obligation to pay and employ creative professionals, even though generative AI tools can sometimes do a good-enough job? What uniquely human value do artists bring to our world?<\/li>\n<\/ul>\n<h3>Power<\/h3>\n<h4>Scenario:<\/h4>\n<p>Sarah has been trying to research potential career options for herself. Sarah\u2019s only electronic device is her phone, which is older, and she is limited in the amount of AI time available to her. Usually, after asking one or two more complicated questions, the AI she is using becomes very slow. She then has to keep going back and forth between other tasks she needs to complete at home and the AI to see if it has responded. Most of the careers the AI is suggesting for her interests require graduate study, but due to her low household income and need to work while studying, getting through an undergraduate degree is going to be quite challenging. AI doesn\u2019t seem to have information about lower barrier-to-entry or part-time jobs that exist in those fields that she could potentially start off with to gain valuable experience. \u00a0When she ask the AI to fix this issue, it thinks for a long time and then just repeats the same information again without listening to her updated request. Some of her classmates have paid for premium AI subscriptions and are not having any of these problems, as AI is quickly suggesting career paths that are within their reach.<\/p>\n<h4>Considerations:<\/h4>\n<ul>\n<li>Who has more power in this situation?<\/li>\n<li>Would having better AI access through libraries help with some of the power issues?<\/li>\n<li>What else could be done to mitigate this?<\/li>\n<\/ul>\n<h2>CONCLUSION<\/h2>\n<p>This chapter scratches the surface of ethical issues related to generative AI, but we hope it provides a helpful overview to spur further discussion and exploration. The content and case studies described here could be a useful way to explore the ethics of AI in the classroom. To ensure AI literacy, it is important to directly address these issues with students.<\/p>\n<p>Generative AI is a rapidly changing technology, and there are likely to be new considerations in the future. We did our best to address ethical issues related to generative AI as they were discussed at the time of writing. As these technologies continue to evolve, some problems may be resolved but also new issues may arise. We recommend re-examining ethical considerations for AI usage on a regular basis.<\/p>\n<h2>HELPFUL RESOURCES<\/h2>\n<h3>Bias<\/h3>\n<ul>\n<li><a href=\"https:\/\/cee.ucdavis.edu\/ai-student-writing\">AI &amp; Student Writing<\/a> \u2013 UC Davis Center for Educational Effectiveness<\/li>\n<li><a href=\"https:\/\/post.parliament.uk\/research-briefings\/post-pn-0712\/\">Use of AI in education delivery and assessment<\/a> \u2013 UK Parliament<\/li>\n<li><a href=\"https:\/\/fas.org\/publication\/modernizing-ai-fairness-analysis-in-education-contexts\/\">Modernizing AI Fairness Analysis in Education Contexts<\/a> \u2013 Federation of American Scientists<\/li>\n<li><a href=\"https:\/\/digital-strategy.ec.europa.eu\/en\/policies\/regulatory-framework-ai\">The The AI ActAct<\/a> \u2013\u2013 European Comm Commission<\/li>\n<li><a href=\"https:\/\/er.educause.edu\/articles\/2025\/3\/ai-procurement-in-higher-education-benefits-and-risks-of-emerging-tools\">AI Procurement in Higher Education<\/a> \u2013 EDUCAUSE Review<\/li>\n<li><a href=\"https:\/\/files.eric.ed.gov\/fulltext\/ED661949.pdf\">Designing for Education with Artificial Intelligence: An Essential Guide for Developers<\/a> \u2013 U.S. Department of Education<\/li>\n<li><a href=\"https:\/\/www.nist.gov\/itl\/ai-risk-management-framework\">AI Risk Management Framework<\/a> \u2013 National Institute of Standards and Technology<\/li>\n<li><a href=\"https:\/\/standards.ieee.org\/ieee\/7003\/11357\/\">Standard for Algorithmic Bias Considerations<\/a> \u2013 IEEE<\/li>\n<li>Standard for Algorithmic Bias Considerations \u2013<\/li>\n<li><a href=\"https:\/\/libraryfreedom.org\/wp-content\/uploads\/2026\/07\/LFP-Critical-AI-in-Higher-Education-Toolkit.pdf\">Critical AI in Higher Education Toolkit<\/a> \u2013 Library Freedom Project<\/li>\n<\/ul>\n<h3>Environmental Impact<\/h3>\n<ul>\n<li><a href=\"https:\/\/www.youtube.com\/@ClimateEmergencyForum\">Climate Emergency Forum<\/a><\/li>\n<li><a href=\"https:\/\/youtu.be\/eOkz6KqqakA?si=eAFN-8HkekkKk8X1\">How Green Is Your Prompt?<\/a><\/li>\n<li><a href=\"https:\/\/www.datacenterwatch.org\/\">Data Center Watch<\/a><\/li>\n<li><a href=\"https:\/\/spectrumnews1.com\/oh\/columbus\/news\/2026\/02\/22\/data-centers--impact-on-the-environment-\">Data centers in Ohio: Economic boost or environmental burden?<\/a><\/li>\n<li><a href=\"https:\/\/www.ceres.org\/resources\/reports\/drained-by-data-the-cumulative-impact-of-data-centers-on-regional-water-stress?\">Drained by Data: The Cumulative Impact of Data Centers on Regional Water Stress<\/a><\/li>\n<li><a href=\"https:\/\/theoec.org\/\">Ohio Environmental Council.<\/a><\/li>\n<li>Oh<a href=\"https:\/\/ohiohouse.gov\/news\/republican\/ohio-house-passes-bill-establishing-the-ohio-data-center-study-commission-142643\">io Data Center Study Commission<\/a>,<\/li>\n<\/ul>\n<h3>Copyright \/ Intellectual Property<\/h3>\n<ul>\n<li><a href=\"https:\/\/copyright.gov\/ai\/\">Report on Copyright and Artificial Intelligence<\/a> \u2013 U.S. Copyright Office<\/li>\n<li><a href=\"https:\/\/www.wired.com\/story\/ai-copyright-case-tracker\/\">AI Copyright Case Tracker<\/a> \u2013 Wired Magazine<\/li>\n<li><a href=\"https:\/\/blogs.gwu.edu\/law-eti\/ai-litigation-database\/\">AI Litigation Database<\/a> \u2013 George Washington University<\/li>\n<li><a href=\"https:\/\/blogs.microsoft.com\/on-the-issues\/2023\/09\/07\/copilot-copyright-commitment-ai-legal-concerns\/\">Microsoft Copilot Copyright Commitment<\/a><\/li>\n<li><a href=\"https:\/\/www.ropesgray.com\/en\/insights\/alerts\/2025\/07\/a-tale-of-three-cases-how-fair-use-is-playing-out-in-ai-copyright-lawsuits\">A Tale of Three Cases: How Fair Use Is Playing Out in AI Copyright Lawsuits<\/a> \u2013 Ropes &amp; Gray Legal Firm<\/li>\n<li><a href=\"https:\/\/nightshade.cs.uchicago.edu\/whatis.html\">Nightshade App<\/a><\/li>\n<li><a href=\"https:\/\/guides.library.cornell.edu\/copyright\/publicdomain\">Copyright Term and Public Domain Chart<\/a> \u2013 Cornell University<\/li>\n<\/ul>\n<h3>Privacy &amp; Security<\/h3>\n<ul>\n<li><a href=\"https:\/\/hai.stanford.edu\/news\/privacy-ai-era-how-do-we-protect-our-personal-information\">Privacy in an AI Era: How Do We Protect Our Personal Information?<\/a> &#8211; Stanford University Human-Centered Artificial Intelligence<\/li>\n<li><a href=\"https:\/\/www.sciencedirect.com\/org\/science\/article\/pii\/S1062737525000605\">Artificial Intelligence (AI) and the Future of Information Privacy<\/a> \u2013 Journal of Global Information Management<\/li>\n<li><a href=\"https:\/\/www.forbes.com\/sites\/federicoguerrini\/2024\/11\/17\/ai-driven-dark-patterns-how-artificial-intelligence-is-supercharging-digital-manipulation\/\">AI-Driven Dark Patterns: How Artificial Intelligence Is Supercharging Digital Manipulation<\/a> &#8211; Forbes<\/li>\n<li><a href=\"https:\/\/www.ed.gov\/sites\/ed\/files\/documents\/ai-report\/ai-report.pdf\">Artificial Intelligence and the Future of Teaching and Learning Insights and Recommendations<\/a> \u2013 Office of Educational Technology (see pg. 32)<\/li>\n<li><a href=\"https:\/\/chat.mistral.ai\/chat\">Mistral AI<\/a> and <a href=\"https:\/\/duck.ai\/\">Duck.ai<\/a><\/li>\n<\/ul>\n<h3>Human Labor<\/h3>\n<ul>\n<li>Humans in the loop; Training AI takes heavy toll on Kenyans working for $2 an hour<\/li>\n<li><a href=\"https:\/\/www.cbsnews.com\/video\/60minutes-2025-06-29\/?intcid=CNM-00-10abd1h\">https:\/\/www.cbsnews.com\/video\/60minutes-2025-06-29\/?intcid=CNM-00-10abd1h<\/a><\/li>\n<li>How big AI companies exploit data workers in Kenya-<a href=\"https:\/\/www.youtube.com\/watch?v=ehkECk2KJjY&amp;t=149s\">https:\/\/www.youtube.com\/watch?v=ehkECk2KJjY&amp;t=149s<\/a><\/li>\n<li>Artificial Intelligence recommendations &#8211; <a href=\"https:\/\/www.unesco.org\/en\/artificial-intelligence\/recommendation-ethics\">https:\/\/www.unesco.org\/en\/artificial-intelligence\/recommendation-ethics<\/a><\/li>\n<li>Generative AI and the Future of Work in America | McKinsey,\u201d accessed January 5, 2025,<\/li>\n<li><a href=\"https:\/\/www.mckinsey.com\/mgi\/our-research\/generative-ai-and-the-future-of-work-in-america\">https:\/\/www.mckinsey.com\/mgi\/our-research\/generative-ai-and-the-future-of-work-in-america<\/a><\/li>\n<\/ul>\n<h3>Power<\/h3>\n<ul>\n<li><a href=\"https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/\">Teaching AI Ethics: Power<\/a>. Leon Furze, (2023). <a href=\"https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/\">https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/<\/a><\/li>\n<li><a href=\"https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf\">Teaching AI Ethics: A Guide for Educators<\/a>. Leon Furze, 2026. <a href=\"https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf\">https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf<\/a><\/li>\n<li><a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3442188.3445922\">On the dangers of stochastic parrots: Can language models be too big?<\/a>\ud83e\udd9c. Bender, E. M., Gebru, T., McMillan-Major, A., &amp; Shmitchell, S. (2021). <em>Proceedings of the 2021 ACM conference on fairness, accountability, and transparency<\/em> (pp. 610-623). <a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3442188.3445922\">https:\/\/dl.acm.org\/doi\/10.1145\/3442188.3445922<\/a><\/li>\n<\/ul>\n<h2>REFERENCES<\/h2>\n<p>AI-Assisted Exam Variant Generation: A Human-in-the-Loop Framework for Automatic Item Creation. (2025). <em>Education Sciences, 15<\/em>(8).<\/p>\n<p>The AP-NORC Center for Public Affairs Research. (October 2025). <a href=\"https:\/\/apnorc.org\/?post_type=project&amp;p=11263\">https:\/\/apnorc.org\/projects\/epic-climate-change-2025\/<\/a><\/p>\n<p>Ashwin, J., Chhabra, A., &amp; Rao, V. (2023). Using large language models for qualitative analysis can introduce serious bias. <em>arXiv preprint<\/em>.<\/p>\n<p>Bias and representation in AI generated text-to-image in education: A systematic review. (2026). [Journal not fully specified in source; verify prior to submission].<\/p>\n<p>Bender, E. M., Gebru, T., McMillan-Major, A., &amp; Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In *<em>Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency<\/em>* (pp. 610-623). Association for Computing Machinery. <a href=\"https:\/\/doi.org\/10.1145\/3442188.3445922\" target=\"_blank\" rel=\"noopener\">https:\/\/doi.org\/10.1145\/3442188.3445922<\/a><\/p>\n<p>Bird, K. A., Castleman, B. L., Mabel, Z., &amp; Song, Y. (2021). Bringing transparency to predictive analytics: A systematic comparison of predictive modeling methods in higher education. <em>AERA Open<\/em>, 7.<\/p>\n<p>Bird, K. A., Castleman, B. L., &amp; Song, Y. (2023). Are algorithms biased in education? Exploring racial bias in predicting community college student success (EdWorkingPaper No. 23-717). <em>Annenberg Institute at Brown University<\/em>.<\/p>\n<p>Bird, K. A., Castleman, B. L., &amp; Song, Y. (2024). Are algorithms biased in education? Exploring racial bias in predicting community college student success. <em>Journal of Policy Analysis and Management<\/em>.<\/p>\n<p>Bisbee, J., Imai, K., Rosenman, E., &amp; Zhou, X. (2024). Synthetic replacements for human survey data? The perils of large language models. <em>Political Analysis<\/em>.<\/p>\n<p>bioRxiv. (2024, December 3). Generative AI divide: How college students&#8217; backgrounds affect their Gen AI literacy.<\/p>\n<p>Bremmer, I., &amp; Suleyman, M. (2023). The AI power paradox: Can states learn to govern artificial intelligence before it is too late? <em>Foreign Affairs, 102<\/em>(5), 26-43. <a href=\"https:\/\/www.foreignaffairs.com\/world\/artificial-intelligence-power-paradox\" target=\"_blank\" rel=\"noopener\">https:\/\/www.foreignaffairs.com\/world\/artificial-intelligence-power-paradox<\/a><\/p>\n<p>Buolamwini, J., &amp; Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. <em>Proceedings of Machine Learning Research, 81,<\/em> 1\u201315.<\/p>\n<p>CalMatters\/The Markup. (2025, July 15). Turnitin helps colleges detect AI writing \u2014 but its cost is more than just a price tag. CalMatters.<\/p>\n<p>Changing Higher Ed. (2026, June 9). AI procurement in higher education and the risk of bias at scale [Podcast episode featuring K. Panetta].<\/p>\n<p>Chauhan, K. (2026). <em>Dead science walking: Publication bias and the AI scientist pipeline<\/em>. arXiv preprint arXiv:2606.04220<em>.<\/em><\/p>\n<p>DeStefano, C., Hackney, J., &amp; Moskal, P. (2026). Understanding the generative AI divide: Faculty and student perspectives in higher education.<em> Online Learning, 30<\/em>(2), 31\u201348.<\/p>\n<p>EDUCAUSE Review. (2025, March 11). AI procurement in higher education: Benefits and risks of emerging tools.<\/p>\n<p>Furze, L. (2023, June 19). <em>Teaching AI ethics: Power<\/em>.\u00a0 <a href=\"https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/\">https:\/\/leonfurze.com\/2023\/06\/19\/teaching-ai-ethics-power\/<\/a><\/p>\n<p>Furze, L. (2026). Teaching AI ethics: A guide for educators. <a href=\"https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf\">https:\/\/leonfurze.com\/wp-content\/uploads\/2026\/02\/Teaching_AI_Ethics_PDF_Version_A4_compressed.pdf<\/a><\/p>\n<p>Garc\u00eda Mathewson, T. (2023, August 14). AI detection tools falsely accuse international students of cheating. <em>The Markup<\/em>.<\/p>\n<p>G\u00e1ndara, D., Anahideh, H., Ison, M. P., &amp; Picchiarini, L. (2024). Inside the black box: Detecting and mitigating algorithmic bias across racialized groups in college student-success prediction. <em>AERA Open<\/em>.<\/p>\n<p>Generalization bias in large language model summarization of scientific research. (2025). [Journal not fully specified in source, PMC-indexed; verify prior to submission].<\/p>\n<p>Huang, R., Wu, H., Yuan, Y., Xu, Y., Qian, H., Zhang, C., Wei, X., Lu, S., Zhang, X., Kan, J., Wan, C., &amp; Liu, Y. (2025). Evaluation and bias analysis of large language models in generating synthetic electronic health records: Comparative study. <em>Journal of Medical Internet Research, 27<\/em>, e65317.<\/p>\n<p>Inside Higher Ed. (2025, April 21). The digital divide: Student generative AI access.<\/p>\n<p>Johnson, D. (2026, March 4). LLMs are getting better at unmasking people online. <em>CyberScoop<\/em>. <a href=\"https:\/\/cyberscoop.com\/ai-deanonymization-risks-online-anonymity-study\/\">https:\/\/cyberscoop.com\/ai-deanonymization-risks-online-anonymity-study\/<\/a><\/p>\n<p>Karell, D. (2026). Not-quite-human tastes: The stylized omnivorousness of LLM survey surrogates. <em>arXiv preprint<\/em>.<\/p>\n<p>Kelly, J, S. Pham, and S. Reilly \u201cHow much water does AI really use?\u201d CBS News Data, June 9, 2026 (<a href=\"https:\/\/www.cbsnews.com\/projects\/2026\/how-much-water-ai-uses\/\" target=\"_blank\" rel=\"noopener\">https:\/\/www.cbsnews.com\/projects\/2026\/how-much-water-ai-uses\/<\/a>)<\/p>\n<p>Lermen, S., Paleka, D., Swanson, J., Aerni, M., Carlini, N., &amp; Tram\u00e8r, F. (2026). Large-scale online deanonymization with LLMs. <em>arXiv preprint<\/em> arXiv:2602.16800.<\/p>\n<p>Li, M., et al. (2025). <em>Understanding and mitigating the bias inheritance in LLM-based data augmentation on downstream tasks<\/em>. arXiv preprint arXiv:2502.04419.<\/p>\n<p>Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., &amp; Zou, J. (2023). GPT detectors are biased against non-native English writers. <em>Patterns, 4<\/em>(7).<\/p>\n<p>Low, A., &amp; Kalender, Z. Y. (2023). Data dialogue with ChatGPT: Using Code Interpreter to simulate and analyse experimental data. <em>arXiv preprint <\/em>arXiv:2311.12415.<\/p>\n<p>McCallum, S. (2023, March 31). ChatGPT banned in Italy over privacy concerns. <em>BBC News.<\/em> <a href=\"https:\/\/www.bbc.com\/news\/technology-65139406\">https:\/\/www.bbc.com\/news\/technology-65139406<\/a><\/p>\n<p>Mishcon de Reya. (n.d.). Generative AI \u2013 IP cases and policy tracker. <em>Mishcon de Reya LLP<\/em>. Retrieved July 24, 2026, from <a href=\"https:\/\/www.mishcon.com\/generative-ai-intellectual-property-cases-and-policy-tracker\">https:\/\/www.mishcon.com\/generative-ai-intellectual-property-cases-and-policy-tracker<\/a><\/p>\n<p>O\u2019Donnell, J and C Crownhart \u201cWe did the math on AI\u2019s energy footprint. Here\u2019s the story you haven\u2019t heard,\u201d MIT Technology Review, May 20, 2025 (<a href=\"https:\/\/www.technologyreview.com\/2025\/05\/20\/1116327\/ai-energy-usage-climate-footprint-big-tech\/\" target=\"_blank\" rel=\"noopener\">https:\/\/www.technologyreview.com\/2025\/05\/20\/1116327\/ai-energy-usage-climate-footprint-big-tech\/<\/a>).<\/p>\n<p>Ouyang, S., Wang, C., Li, Y., &amp; Zhang, T. (2024). An empirical study of the non-determinism of ChatGPT in code generation. ACM Transactions on Software Engineering and Methodology.<\/p>\n<p>Wong, W. (2025, September 19). The new digital divide in higher education: AI as public capability or private advantage? <em>PA Times Online.<\/em> <a href=\"https:\/\/patimes.org\/the-new-digital-divide-in-higher-education-ai-as-public-capability-or-private-advantage\/\" target=\"_blank\" rel=\"noopener\">https:\/\/patimes.org\/the-new-digital-divide-in-higher-education-ai-as-public-capability-or-private-advantage\/<\/a><\/p>\n<p>Parker, Will \u201cThe Americans Striking It Rich in the Data-Center Buildout.\u201d The Wall Street Journal, July 13, 2026 (<a href=\"https:\/\/www.wsj.com\/real-estate\/commercial\/the-americans-striking-it-rich-in-the-data-center-buildout-71c6a8a8\" target=\"_blank\" rel=\"noopener\">https:\/\/www.wsj.com\/real-estate\/commercial\/the-americans-striking-it-rich-in-the-data-center-buildout-71c6a8a8<\/a>).<\/p>\n<p>Soomro, K.A., Kale, U., Curtis, R. et al. (2020). Digital divide among higher education faculty. <em>International Journal of Educational Technology in Higher Education,<\/em> <em>17<\/em>(21). <a href=\"https:\/\/doi.org\/10.1186\/s41239-020-00191-5\" target=\"_blank\" rel=\"noopener\">https:\/\/doi.org\/10.1186\/s41239-020-00191-5<\/a><\/p>\n<p>Rho, J., &amp; Karumbaiah, S. (2026). How students interpret representational cues in AI-generated educational illustrations. In E. G. Blanchard, G. Chen, M. Chi, &amp; S. Isotani (Eds.), <em>Proceedings of the 27th International Conference on Artificial Intelligence in Education (AIED)<\/em>. Springer.<\/p>\n<p>Ribes-Lafoz, M., Navarro-Colorado, B., &amp; Rovira-Collado, J. (2026). Exploring strategies to detect and mitigate bias in AI in education: Students&#8217; perceptions and didactic approaches. Trends in Higher Education, 5(2), 33.<\/p>\n<p>Ritchie, H. (2025, Aug 21). <em>What&#8217;s the carbon footprint of using ChatGPT or Gemini?<\/em> <a href=\"https:\/\/hannahritchie.substack.com\/p\/ai-footprint-august-2025\" target=\"_blank\" rel=\"noopener\">https:\/\/hannahritchie.substack.com\/p\/ai-footprint-august-2025<\/a><\/p>\n<p>Ruta, M. R., Gaidici, T., Irwin, C., &amp; Lifshitz, J. (2025). ChatGPT for univariate statistics: Validation of AI-assisted data analysis in healthcare research. Journal of Medical Internet Research, 27, e63550.<\/p>\n<p>Schmidt, S. (2025, Sep 5). Bartz v. Anthropic: Settlement reached after landmark summary judgment and class certification. <em>Inside Tech Law. <\/em><a href=\"https:\/\/www.insidetechlaw.com\/blog\/2025\/09\/bartz-v-anthropic-settlement-reached-after-landmark-summary-judgment-and-class-certification\">https:\/\/www.insidetechlaw.com\/blog\/2025\/09\/bartz-v-anthropic-settlement-reached-after-landmark-summary-judgment-and-class-certification<\/a><\/p>\n<p>Si, C., Yang, D., &amp; Hashimoto, T. (2024). <em>Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers<\/em>. ArXiv Reprint arXiv:2409.04109<\/p>\n<p>Springer. (2026a). Academic performance and inequality: Can generative AI widen the digital divide in higher education?<\/p>\n<p>Springer. (2026b). From access to ability: Educator&#8217;s perspectives on AI and new digital divides in higher education. Digital Society.<\/p>\n<p>Student-Centered LLM Q&amp;A Study. (2025). Evaluating LLM-generated Q&amp;A test: A student-centered study. In Artificial Intelligence in Education (Springer).<\/p>\n<p>Surveillance and disability in online proctored exams: Student perspectives and design implications. (2025). arXiv preprint.<\/p>\n<p>The homogenizing effect of large language models on human expression and thought. (2026). Trends in Cognitive Sciences.<\/p>\n<p>University of Kansas Center for Teaching Excellence. (n.d.). Addressing bias in AI.<\/p>\n<p>University of Pittsburgh Teaching Center. (n.d.). Encouraging academic integrity.<\/p>\n<p>U.S. Copyright Office (n.d.). Copyright Office Releases Part 2 of Artificial Intelligence Report. <em>Newsnet issue 1060<\/em>. Retrieved July 24, 2026, from <a href=\"https:\/\/www.copyright.gov\/newsnet\/2025\/1060.html\">https:\/\/www.copyright.gov\/newsnet\/2025\/1060.html<\/a><\/p>\n<p>U.S. Copyright Office (n.d.). <em>What is Copyright?<\/em> Retrieved July 24, 2026, from <a href=\"https:\/\/www.copyright.gov\/what-is-copyright\/\">https:\/\/www.copyright.gov\/what-is-copyright\/<\/a><\/p>\n<p>Vanderbilt University Brightspace\/Instructional Technology. (2023, August 16). Guidance on AI detection and why we&#8217;re disabling Turnitin&#8217;s AI detector.<\/p>\n<p>Wang, Y., et al. (2026). <em>BRIDGE the gap: Mitigating bias amplification in automated scoring of English language learners via inter-group data augmentation<\/em>. arXiv preprint arXiv:2602.23580.<\/p>\n<p>Yan, Y., Y. X. Pang, X. Luo, Q. Lin, C. H. Pang, H, Zhang, X. Gao, &amp; T. Wu (2024). Carbon dioxide-focused greenhouse gas emissions from petrochemical plants and associated industries: Critical overview, recent advances and future prospects of mitigation strategies. <em>Process Safety and Environmental Protection, 188, <\/em>406\u2013421.<\/p>\n<p>Yoder-Himes, D. R., et al. (2022). Racial, skin tone, and sex disparities in automated proctoring software. <em>Frontiers in Education, 7<\/em>.<\/p>\n<h2>AI STATEMENT<\/h2>\n<table class=\"grid aligncenter\">\n<thead>\n<tr>\n<th style=\"text-align: center;\" scope=\"col\"><strong>Category<\/strong><\/th>\n<th style=\"text-align: center;\" scope=\"col\"><strong>Label<\/strong><\/th>\n<th style=\"text-align: center;\" scope=\"col\"><strong>Description of AI Use<\/strong><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Content Research<\/td>\n<td><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-content\/uploads\/sites\/216\/2026\/08\/Picture1.png\" alt=\"4 squares within a square\" width=\"55\" height=\"55\" class=\"aligncenter size-full wp-image-160\" \/><\/p>\n<p>Cyborg<\/td>\n<td>Some sections used generative AI to find sources to cite.<\/td>\n<\/tr>\n<tr>\n<td>Writing\u2014Content Generation<\/td>\n<td><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-content\/uploads\/sites\/216\/2026\/08\/Picture2.png\" alt=\"4 squares within a square\" width=\"55\" height=\"55\" class=\"aligncenter size-full wp-image-161\" \/><\/p>\n<p>Handyperson<\/td>\n<td>Some section authors used generative AI to write content, with significant human review. A few case studies were created by generative AI and then edited.<\/td>\n<\/tr>\n<tr>\n<td>Writing\u2014Review &amp; Editing<\/td>\n<td><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-content\/uploads\/sites\/216\/2026\/08\/Picture2.png\" alt=\"4 squares within a square\" width=\"55\" height=\"55\" class=\"aligncenter size-full wp-image-161\" \/><\/p>\n<p>Handyperson<\/td>\n<td>Some section authors used generative AI to review and edit human-written content.<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>&nbsp;<\/p>\n","protected":false},"parent":0,"menu_order":7,"template":"","meta":{"pb_part_invisible":false},"contributor":[],"license":[],"class_list":["post-114","part","type-part","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-json\/pressbooks\/v2\/parts\/114","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-json\/pressbooks\/v2\/parts"}],"about":[{"href":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-json\/wp\/v2\/types\/part"}],"version-history":[{"count":9,"href":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-json\/pressbooks\/v2\/parts\/114\/revisions"}],"predecessor-version":[{"id":299,"href":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-json\/pressbooks\/v2\/parts\/114\/revisions\/299"}],"wp:attachment":[{"href":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-json\/wp\/v2\/media?parent=114"}],"wp:term":[{"taxonomy":"contributor","embeddable":true,"href":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-json\/wp\/v2\/contributor?post=114"},{"taxonomy":"license","embeddable":true,"href":"https:\/\/pressbooks.ulib.csuohio.edu\/usingaiinacademics\/wp-json\/wp\/v2\/license?post=114"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}