Ethics, Bias, and Responsible Use

Module 10 · The Questions That Matter Most

Content current as of April 2026. Specific statutes, court cases, and enforcement regimes evolve; the underlying ethical questions — who’s affected, who decides, who’s accountable — do not.

Learning Objectives

Builds on: Every previous module. Ethics isn’t a chapter you complete and close.

After completing this module, you will be able to:

  1. Identify where bias enters AI systems — and which kinds resist technical fixes
  2. Explain the fairness impossibility result and why “just make it fair” is not a coherent instruction
  3. Evaluate AI tools for privacy, transparency, and accountability
  4. Navigate academic integrity in a way that reflects what detection tools can and cannot do
  5. Understand the current (April 2026) state of AI copyright — training data and output ownership
  6. Apply accessibility principles to AI-generated content
  7. Draft an AI use policy for your classroom, team, or organization that will survive contact with reality
  8. Compare the major responsible-AI frameworks (UNESCO, OECD, EDUCAUSE, EU AI Act)

The Algorithm That Sentenced People to Prison

In 2016, ProPublica published an investigation that reshaped how the world thinks about AI fairness. They examined COMPAS — Correctional Offender Management Profiling for Alternative Sanctions — an algorithm used across the United States to predict whether a defendant would reoffend. Judges used COMPAS scores to inform bail, sentencing, and parole decisions.

ProPublica’s analysis found that the algorithm had a false-positive rate roughly twice as high for Black defendants as for White defendants: it incorrectly labeled Black defendants as “high risk” when they did not go on to reoffend about twice as often as it did White defendants. At the same time, White defendants who did go on to reoffend were more likely to be incorrectly labeled low risk. The algorithm didn’t use race as an input — but it used zip code, employment history, education level, and prior arrests, all of which correlate with race in the United States because of a century of policy choices about housing, policing, and schools.

Northpointe (COMPAS’s maker) pushed back hard. They argued that by a different fairness metric — predictive parity, meaning that a given risk score translated to roughly the same reoffense probability regardless of race — COMPAS was in fact fair. And they were right, in their narrow sense.

Both sides were technically correct. What followed was an uncomfortable result that has shaped algorithmic fairness ever since.

In 2017, Carnegie Mellon statistician Alexandra Chouldechova proved something cleaner and harder to argue with. When groups have different base rates — in this case, different reoffense rates, which themselves reflect decades of inequitable policing and sentencing — you cannot simultaneously achieve:

  1. Predictive parity (equal positive predictive value across groups — a given score means the same thing)
  2. Equal false-positive rates across groups
  3. Equal false-negative rates across groups

…unless your classifier is perfect. For any realistic classifier, satisfying any one of these means violating at least one of the others. It’s not a bug in COMPAS. It’s a mathematical fact about classification when groups differ.

This is the impossibility theorem of fairness (Chouldechova, Big Data, 2017; related results by Kleinberg, Mullainathan, and Raghavan the same year). You can argue about which definition of fairness matters most for criminal justice. You cannot satisfy all of them at once.

People went to prison. People stayed in prison longer. Based in part on a score from an algorithm whose fairness depended on which definition of “fair” you used — and whose designers likely did not realize they were making that choice.

This module isn’t about whether AI is “good” or “bad.” It’s about the questions you must ask every time AI is used to make decisions that affect people’s lives — including your own.

Common Misconceptions About AI Ethics
  • “AI is objective.” AI finds patterns in data created by humans. Human data reflects human systems, which reflect human biases. AI automates and scales whatever pattern was in the data — with mathematical authority that can look like neutrality.
  • “Bias can be fully eliminated.” The Chouldechova result rules out having-it-all when base rates differ. Bias can be reduced, audited, and made transparent — but not eliminated without picking which fairness definition matters for the use case.
  • “AI detection tools reliably catch AI-generated text.” A 2023 Stanford study (Liang et al., Patterns) found that seven widely used GPT detectors flagged 61.3% of human-written TOEFL essays as AI-generated — with near-zero false positives on native-English eighth-grader essays. These tools penalize less-varied vocabulary, which is exactly what non-native English writing tends to look like.
  • “If the data is big enough, bias disappears.” More biased data is more confidently biased. A billion records with the same blind spot is still blind.
  • “Ethics is a one-time review.” Models update. Data drifts. Populations change. Laws change. A policy written in April 2026 needs revisiting by the end of the semester.
  • “Only AI developers need to worry about ethics.” Every user of AI makes ethical choices: what data to enter, what outputs to trust, whose life is affected by the decisions AI helps them make.

Bias in AI Systems — Where It Comes From

Module 2 introduced the data pipeline and bias entry points. Here we go deeper into the consequences, with real cases.

Historical bias

AI trained on historical data inherits historical inequities. If past hiring was biased, the hiring model is biased. If past lending was discriminatory, the lending model is discriminatory. The AI doesn’t create the bias — it automates and scales whatever was already there.

The Amazon hiring tool (2018). Amazon trained a resume-screening model on ten years of applications to technical roles. Because that decade’s applicant pool skewed heavily male, the model learned that “female-coded” words (attendance at a women’s college, membership in a women’s chess club) were negative signals. Amazon scrapped the system once the bias was discovered. The lesson is not that this particular tool was defective; it’s that any tool trained on that data would have produced that result.

Representation bias

If training data overrepresents some demographics, the model performs better on them. The landmark example is Gender Shades (Buolamwini & Gebru, 2018). Evaluating three commercial gender-classification systems (IBM, Microsoft, Face++), the researchers found:

  • Error rate on lighter-skinned men: 0.8%
  • Error rate on darker-skinned women: up to 34.7%

A forty-fold gap. For the darkest-skinned women in the dataset, two of the three systems had error rates near 47% — effectively guessing. Those same systems were being sold and deployed in real decision-making contexts.

Measurement bias

What you measure shapes what the model optimizes for. The most consequential example in healthcare is the Optum risk-stratification algorithm studied by Obermeyer, Powers, Vogeli, and Mullainathan (Science, 2019). The algorithm flagged patients for extra care based on predicted future healthcare costs — a common industry proxy for “who needs the most help.”

But in the U.S., Black patients at the same level of clinical need generate less healthcare spending than White patients, because of lifelong differences in access, trust, and insurance. So using cost as a proxy for need systematically deprioritized Black patients. At a given risk score, Black patients were demonstrably sicker — they had, for example, 26% more chronic conditions than White patients at the 97th-percentile auto-enrollment threshold.

When Obermeyer and colleagues reformulated the algorithm to predict health instead of cost, the share of Black patients in the automatic-enrollment tier jumped from 17.7% to 46.5%. The manufacturer worked with the researchers and reduced the racial bias by roughly 84%. The algorithm was not malicious. It was a perfectly ordinary engineering choice — use the number that’s already in the database — that encoded a century of unequal access.

Aggregation bias

Treating diverse populations as one. A diabetes model trained on all ethnicities may work well on average while failing for specific communities whose risk factors differ. Average performance can disguise catastrophic subgroup performance — which is exactly what Gender Shades exposed for facial analysis.

The fairness paradox, made concrete

Return to the Chouldechova result. The uncomfortable truth it forces into the open: “fair” is not one thing. Deciding which fairness property you care most about is a choice — ideally an explicit, transparent one made before deployment. Hiding that choice inside “our model is 85% accurate overall” lets the real decision happen by accident.

Guided Practice 10.1 — The Bias Audit

A hospital is building an AI to predict which patients are at highest risk for heart disease, so nurses can prioritize follow-up calls. Training data: five years of patient records at the hospital.

Identify at least 4 ways bias could enter this system:

  1. At the data-collection stage
  2. At the labeling / measurement stage
  3. At the training stage
  4. At the deployment stage

Then answer: which definition of fairness would you prioritize — equal accuracy across demographics, or equal false-negative rates (never missing a high-risk patient, regardless of demographic)? Why?

Click for analysis
  1. Collection. Patients with less access — insurance, transport, language — appear less often in the data. The model literally doesn’t know about them.
  2. Measurement. If “number of past visits” is a feature, patients who couldn’t afford to visit appear “healthier” than those who saw doctors regularly. The Obermeyer lesson in microcosm.
  3. Training. If the historical data reflects decades of underdiagnosis in women (whose heart-attack symptoms are famously dismissed), the model learns these patients are lower risk — perpetuating underdiagnosis.
  4. Deployment. If deployed in a community with different demographics than the training hospital, subgroup performance can drift unpredictably.
Fairness choice. In healthcare, most ethicists argue for minimizing false negatives (missing a sick patient) even at the cost of more false positives (over-calling). An unnecessary follow-up call is cheap. A missed one can be fatal. But this tradeoff should be made explicitly, documented, and communicated to the nurses who will use the tool — not buried as an implementation detail.
Guided Practice 10.2 — The Fairness Dilemma

A university uses AI to flag students at risk of dropping out, so advisors can intervene early. The system is 85% accurate overall but has different false-negative rates by demographic:

  • White students: 10% FNR (misses 10% of at-risk students)
  • Hispanic/Latino students: 22% FNR (misses 22% of at-risk students)
  • Black students: 19% FNR (misses 19% of at-risk students)

The model is accurate on average but systematically misses more at-risk students from underrepresented groups.

  1. Is this system “fair”? By which definition?
  2. Would you deploy as-is, fix and redeploy, or scrap entirely?
  3. What information would you need to make this decision responsibly?
Discussion points This scenario reflects documented patterns in student-success prediction. The system is accurate by overall metrics but inequitable in who it fails — and the students it fails most are the ones who need intervention most. Deploying as-is means underrepresented students fall through the cracks at higher rates. Scrapping entirely means no students get flagged and everyone loses. The responsible path is usually: acknowledge the disparity openly, investigate causes (almost always training-data underrepresentation), adjust thresholds per group or retrain on rebalanced data, publish the limitations, and train advisors to treat flags as one input, not a verdict. The worst option is deploying without knowing the disparity exists.

Privacy and Data Security

What happens to your data

When you type a prompt into ChatGPT, Claude, or Gemini, your text travels to the provider’s servers. What happens next depends on the tier and the contract:

Provider Free tier: used for training? Enterprise tier Current pricing (Apr 2026)
OpenAI (ChatGPT / GPT-5) Consumer free/Plus: may be used unless you opt out. API: not by default. ChatGPT Enterprise and API with zero data retention GPT-5.4 $2.50/$15 per M tokens
Anthropic (Claude) Not used for training by default on any tier Team, Enterprise, and API with BAAs available Opus 4.7 $5/$25, Sonnet 4.6 $3/$15, Haiku 4.5 $1/$5 per M
Google (Gemini) Consumer: may be used. Workspace/enterprise: not. Google Workspace AI, Vertex AI Gemini 3.1 Pro $2/$12, Flash $0.50/$3 per M
xAI (Grok) Consumer: terms vary. API with enterprise options Grok 4.1 ~$0.20/$0.50 per M

Prices move every few months; the pattern — free consumer tier uses data, enterprise does not — is stable.

Key privacy principles

  • Never paste sensitive data into free-tier AI tools. Student records (FERPA), patient information (HIPAA), financial data, passwords, proprietary business information — none of these belong in a consumer chatbot.
  • Read the data policy for your tier before using any AI for sensitive work. The policy for ChatGPT Plus is different from ChatGPT Enterprise. They are billed together and they are not the same product.
  • For regulated data, go local or get a BAA. Module 9’s SLM-RAG pattern is the gold standard when patient or student data genuinely can’t leave the building.
  • Assume anything you type could be read by a human. Every major provider retains some conversations for safety review. Don’t type anything you wouldn’t be comfortable having a human at the provider’s trust-and-safety team see.
  • A live legal data point. In January 2026, a federal judge in In re OpenAI Copyright Infringement Litigation ordered OpenAI to produce 20 million anonymized ChatGPT conversation logs as part of discovery in the consolidated copyright suits. De-identification is the court’s chosen privacy protection — but the ruling establishes that user prompts are discoverable evidence in litigation against AI companies. Assume your chat history could, in some future case, be subpoenaed.
✅ Test Yourself: Ethics So Far
  1. COMPAS didn’t use race as an input. How was it biased? (It used features — zip code, employment, prior arrests — that correlate with race because of systemic inequities in housing, policing, and employment.)
  2. Why can’t you satisfy all definitions of fairness simultaneously? (Chouldechova’s impossibility: when base rates differ, equal predictive parity, equal false-positive rates, and equal false-negative rates cannot all hold at once.)
  3. Before pasting text into a free AI tool, what do you check? (Whether the text contains regulated or sensitive data — PII, PHI, student records, financial detail, client work — and whether the tier you’re using has training opt-out.)

AI Ethics Across Fields

The ethical questions change depending on where AI lands. Here’s what “doing it right” looks like in different contexts:

The dilemma. An AI early-warning system in the ER can predict patient deterioration 45 minutes before vital signs shift — but it was trained on data from a suburban academic medical center. Will it work in a rural clinic with different patient demographics?

The stakes. Lives. If the model’s predictions are weaker for certain populations, those patients get slower care — the Obermeyer pattern.

The principle. Validate on your population before deploying. Audit subgroup performance explicitly. In healthcare AI, “works on average” is not good enough — the unlucky 10% of patients where it fails are the ones who most need it to work.

The dilemma. A university wants to use AI to flag at-risk students for early intervention using grades, attendance, LMS engagement, and financial-aid status.

The stakes. Students flagged incorrectly feel surveilled. Students missed by the model get no support. First-generation and low-income students are likely to generate false positives because the model conflates differences in engagement style with risk.

The principle. Prediction is not destiny. Use AI to identify students who might benefit from a conversation, not to produce a risk score that gets treated as a verdict. Share the flag criteria with students. Offer a way to contest.

The dilemma. Your company’s AI resume screener consistently ranks candidates from certain universities higher. You pulled the model because it was biased against women; now the “fixed” model is biased toward one social class.

The stakes. You’re automating the biases of whoever made the past hiring decisions. Your “diverse talent pipeline” becomes a story you tell yourself.

The principle. Audit the outputs, not the inputs. If the model’s shortlist systematically doesn’t reflect the applicant pool’s diversity, something is wrong — regardless of which protected attribute is involved. The fix is usually at the data layer, not the algorithm.

The dilemma. A precision-agriculture AI recommends pesticide applications based on satellite imagery and weather data. It reduces chemical use by 30% on large industrial operations but was never trained on small-acreage organic farms.

The stakes. Small farmers who can’t afford the technology fall further behind. The model’s recommendations may be actively wrong for non-standard practices. Kern County’s small growers live this reality.

The principle. Technology that only works for large operators deepens the equity gap. Validate across farm sizes and practices before claiming broad applicability.

The dilemma. A county uses AI to prioritize building inspections based on complaint history, property age, and code violation records. Neighborhoods with more active complainants get more inspections.

The stakes. Under-resourced neighborhoods that complain less get fewer inspections — and more unsafe buildings. This is the COMPAS pattern applied to code enforcement.

The principle. AI that optimizes on historical patterns entrenches historical neglect. Use AI to find gaps, not to reinforce who already gets attention. In any high-stakes government deployment, mandatory bias audits and human override are the bare minimum.

Academic Integrity in the Age of AI

This is the most contested topic in education. The landscape:

The spectrum of institutional responses

TOTAL BAN ←————————————————————————→ FULL INTEGRATION
"No AI use            "AI permitted            "AI required
 permitted"            with disclosure"          as a learning tool"

Most institutions are converging on the middle — AI permitted with disclosure and guardrails. A CSU system survey of 80,000+ students found that 80.2% of students feel uncomfortable submitting AI-generated work as their own, which suggests students themselves want clear boundaries, not a free-for-all.

The ELEVATE transparency approach

Rather than banning AI or ignoring it, the approach tested across multiple undergraduate courses requires:

  1. Shared transcripts — students submit the full AI conversation link alongside their work
  2. Their own work alongside — students submit their final answer/analysis separately from the AI transcript
  3. Reflection — 3–4 sentences about how AI helped (or didn’t) and what they learned
  4. SMARTS evidence — did the student follow State → Map → Ask → Reason → Think → Summarize?

This approach makes AI use visible and accountable rather than hidden and undetectable. The AI is a tool; the transcript is proof of how it was used. It also reframes the question from “did you cheat?” to “show your process” — which is what instructors actually care about.

AI detection tools: a cautionary note (with numbers)

Tools like Turnitin’s AI indicator, GPTZero, and Originality.AI claim to identify AI-generated text. The evidence says be very careful.

The landmark study is Liang, Yuksekgonul, Mao, Wu & Zou, Patterns (2023). They evaluated seven widely used GPT detectors on 91 TOEFL essays (non-native English writers) and 88 U.S. eighth-grade essays (native English writers):

  • Average false-positive rate on native-English eighth-grade essays: near zero
  • Average false-positive rate on TOEFL essays by non-native English writers: 61.3%
  • At least one detector flagged 97.8% of the TOEFL essays as AI-generated
  • 19.8% of the TOEFL essays were unanimously flagged by all seven detectors

The mechanism is straightforward and uncomfortable: detectors look for low text perplexity — predictable word choice and simpler grammar — as a signal of AI generation. Non-native English writing has the same signal for entirely different reasons. The detectors don’t discriminate against an accent; they discriminate against the statistical fingerprint of a smaller vocabulary and more cautious syntax.

Separately, the bypass result in the same paper: asking ChatGPT to “elevate the provided text by employing literary language” caused detection rates on AI-generated essays to drop dramatically — as low as near-zero on some detectors. Detectors that are easy to fool for native English writers but aggressive against non-native ones are the exact worst combination for use as a disciplinary instrument.

Do not use AI detectors as the sole evidence in an academic-integrity case. A score can be one input, but it cannot be a verdict — especially not in a class with international students. Several universities have already been sued or subjected to regulatory complaints over detector-driven sanctions. The practical rule is the same rule that applies to AI use itself: humans decide, with evidence in hand.

Guided Practice 10.3 — The Academic Integrity Scenario

You’re a professor. A student submits an essay that reads unusually well — better than their previous work. An AI detector flags it as “87% likely AI-generated.” The student — a non-native English speaker — says they used ChatGPT as a brainstorming tool and writing coach but wrote every word themselves.

  1. What do you do? Consider: the evidence, the student’s explanation, the detector’s documented false-positive profile, and your course policy.
  2. What policy would have prevented this ambiguity?
  3. Draft 3 sentences for a syllabus AI policy that addresses this scenario directly.
One approach

The detector score alone is not sufficient. Liang et al.’s 61.3% false-positive rate on TOEFL essays means, for a non-native writer, a flagged essay is roughly coin-flip evidence of actual AI authorship.

The right response depends on your policy. If you have a transparency requirement (shared transcripts), ask to see the ChatGPT conversation. If the transcript shows brainstorming and coaching (not copy-paste), the student followed a legitimate process — and under most permissive policies this is allowed. If the transcript is an extensive back-and-forth resulting in a finished essay, that’s a different conversation.

The policy that prevents ambiguity: “AI use is permitted for brainstorming and feedback. When AI is used, submit the conversation transcript alongside your work. Your final text must be substantially your own writing.” Notice what this doesn’t rely on: a detector.
Guided Practice 10.4 — Write Your Elevator AI Ethics Speech

You have 60 seconds to explain to a colleague why AI ethics matters for your field. Write four sentences:

  • Sentence 1: one specific way AI is being used in your field
  • Sentence 2: one specific risk or ethical concern that use creates
  • Sentence 3: what happens if that risk is ignored
  • Sentence 4: one practical step your organization should take this quarter
Example (healthcare) “Hospitals are using AI risk-stratification tools to decide which patients get follow-up calls. The Obermeyer 2019 study showed that one widely deployed algorithm was cutting the Black patient share of auto-enrollment by more than half because it used cost as a proxy for need. If we deploy these models without subgroup audits, we systematically underserve the patients who need the most help. Every hospital using predictive AI should audit subgroup performance at least annually — and publish the results internally so clinicians know the tool’s limits.”

AI and Accessibility

AI has enormous potential for accessibility — and the potential to create new barriers. Both directions matter.

How AI helps

  • Screen readers and voice interfaces for visually impaired users
  • Real-time captioning for deaf and hard-of-hearing users
  • Text simplification for cognitive disabilities
  • Alt-text generation for images (critical, because it works at scale)
  • Translation for language barriers
  • Personalized pacing for individual learning needs

How AI can harm

  • Bias against disabled users in training data (underrepresentation — the Gender Shades problem extended to disability)
  • Inaccessible AI interfaces that don’t work with assistive technology
  • Voice-only interfaces that exclude non-speaking users
  • Automated benefits decisions without meaningful human review — a well-documented pattern in disability-benefits determinations
  • New surveillance risks disguised as accessibility features
Designing AI Use for Accessibility

When you create AI-powered tools, documents, or content, apply universal design: ensure outputs work with screen readers, provide text alternatives for images, use plain language, and support multiple interaction modes. Module 7’s image generation is only useful to visually impaired users if you also generate descriptive alt text.

Writing Your AI Use Policy

Every classroom, team, and organization needs an AI use policy. The template:

The 6 sections every AI policy needs

  1. Scope. What tools and tasks does this cover?
  2. Permitted uses. What AI use is encouraged or allowed?
  3. Restricted uses. What requires approval or disclosure?
  4. Prohibited uses. What is never acceptable?
  5. Data handling. What information may — and may not — be entered into AI tools?
  6. Review schedule. When and how will this policy be updated?

Example: Classroom AI Use Policy Template

AI Use Policy — [Course Name] — [Semester]

Philosophy. AI tools are permitted and encouraged as learning aids. The goal is to learn to use AI effectively and ethically, not to avoid it.

Permitted. Using AI for brainstorming, outlining, checking grammar, generating practice problems, and explaining concepts. All AI use must be disclosed.

Required disclosure. When submitting work that involved AI assistance, include: (a) the AI tool used, (b) a shared conversation link or screenshot, and (c) 2–3 sentences describing how AI contributed and what you contributed.

Prohibited. Submitting AI-generated text as your own writing without disclosure. Using AI during closed-book exams. Entering other students’ work, or any FERPA-protected information, into any AI tool.

Data handling. Do not enter personally identifiable information, other students’ work, or exam questions into any AI tool.

Integrity. AI detection tool scores alone will not be used as evidence of a violation. Any integrity concern will be investigated with the student, using the submitted transcripts as primary evidence.

Updates. This policy will be reviewed at mid-semester and between semesters as AI tools evolve.

Responsible AI Frameworks

Several major frameworks now guide responsible AI adoption. Knowing they exist helps you evaluate any AI initiative — and makes grant and proposal language easier to write.

The NIST AI Risk Management Framework — four functions

Released January 2023 and widely adopted across U.S. enterprise and government, the NIST AI RMF organizes responsible AI practice into four interlocking functions. It’s voluntary, sector-agnostic, and increasingly referenced in contracts, audits, and regulatory filings. Worth knowing by name:

Function Purpose Representative questions
Govern Cultivate an AI-risk-aware organizational culture. Cross-cutting — infuses the other three. Who owns AI risk decisions? What policies apply? Do personnel have training to play their role?
Map Contextualize the AI system in its operational environment. Identify impacts across technical, organizational, legal, and social dimensions. What is this AI for? Who is affected? What could go wrong? Is the use case a go / no-go?
Measure Assess risk with quantitative metrics and qualitative review. Test for accuracy, bias, robustness, privacy, and security. How well does it work on subgroups? What’s the failure rate? What metrics matter for this use case?
Manage Allocate resources to the risks found in Map and Measure. Mitigate, monitor, and respond to incidents. What’s the bias mitigation plan? The incident-response plan? The deprecation criteria?

The framework doesn’t prescribe tools — it prescribes outcomes. It’s explicitly iterative: as the AI evolves, you cycle back through Map → Measure → Manage. A companion Generative AI Profile (NIST-AI-600-1, July 2024) extends the framework specifically to GenAI risks like hallucination, data leakage, prompt injection, and synthetic-media harms. For any formal AI project, the RMF is the cleanest vocabulary to talk about risk that regulators, auditors, and procurement teams all recognize.

The framework landscape

Framework Source Focus
EU AI Act (2024, phased through 2027) European Union Binding risk-based regulation; enforcement powers live Aug 2026
NIST AI RMF 1.0 (2023) + GenAI Profile (2024) U.S. NIST Voluntary, widely adopted. Govern-Map-Measure-Manage
UNESCO AI Competency Framework for Teachers (2024) United Nations 15 competencies × 5 dimensions × 3 progression levels
OECD + European Commission AI Literacy Framework (2025) OECD / EC 4 domains: Engage, Create, Manage, Design
EDUCAUSE ALTL (2024) EDUCAUSE 4 areas: Technical Understanding, Evaluative Skills, Practical Application, Ethical Considerations
Long & Magerko (2020) Academic 17 AI-literacy competencies across 5 themes
ED-AI Lit (Allen & Kendeou, 2024) Academic Knowledge, Evaluation, Collaboration, Contextualization, Autonomy, Ethics

U.S. policy: a rapidly moving target

Between early 2025 and April 2026, U.S. federal AI policy flipped direction — twice. The short version:

  • Oct 2023 — Biden’s Executive Order 14110 (“Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence”). Mandated red-teaming for high-risk models, safety evaluations, and agency oversight. OMB memos M-24-10 and M-24-18 set standards for federal use and acquisition of AI.
  • Jan 20, 2025 — EO 14110 rescinded on day one of the second Trump administration.
  • Jan 23, 2025 — Executive Order 14179 (“Removing Barriers to American Leadership in Artificial Intelligence”). Reorientation toward “AI dominance” and deregulation. Ordered an AI Action Plan within 180 days.
  • July 23, 2025 — AI Action Plan released. Executive orders on accelerated data-center permitting and preventing “woke AI.”
  • Nov 24, 2025 — Genesis Mission executive order, directing federal research agencies to coordinate AI-accelerated science and share federal research datasets with approved partners.
  • Dec 11, 2025 — Executive order preempting state-level AI regulations (pointed at Colorado’s AI Act, California SB 1047-era bills, and others).

The implication for practitioners is less about the specific policies and more about the landscape: there is currently no binding federal AI regulation comparable to the EU AI Act. NIST RMF remains voluntary, OMB guidance is being rewritten, and state laws (especially Colorado AI Act, effective 2026) are where the enforcement risk lives for now — with the federal government openly trying to preempt them. For any U.S. organization making AI policy, this is a live-wire environment: write policies to the stricter of EU AI Act and applicable state law, and don’t count on federal rules being stable for more than a few quarters.

The environmental cost: AI’s real footprint

Every module until now has treated AI as if it lived in the abstract. It doesn’t. AI runs on data centers that consume electricity and water at scales worth naming.

Per-query cost. The best current benchmarks (Nigam et al., arXiv 2505.09598, 2025) put a short GPT-4o query at ~0.42 Wh. Long prompts against reasoning models like o3 or DeepSeek-R1 can exceed 33 Wh per response — more than 70× a GPT-4.1 nano query. A Washington Post / University of Colorado Riverside analysis estimated a 100-word ChatGPT-4 response at ~0.14 kWh of electricity and ~519 mL of water (for cooling — depends heavily on data-center location; Microsoft’s 2024 sustainability report cited 1.8–12 L per kWh across its global fleet, with Ireland efficient and Washington state less so).

Training cost. GPT-3 training consumed an estimated 1,287 MWh and ~502 metric tons of CO₂-equivalent. GPT-4-scale training likely used 1.9–4.9 million liters of water across cooling cycles. Frontier models released since 2024 (GPT-5, Gemini 3, Grok 4) haven’t published numbers, but compute requirements scale with model size.

System-level projections. Global data-center electricity demand is projected to grow from roughly 415 TWh (2024) to ~945 TWh by 2030 — a jump from 1.5% to ~3% of global electricity demand, driven disproportionately by AI. U.S. AI data-center water consumption is projected to rise 200–300 billion gallons annually by 2030.

The ethical implications are real but nuanced. The resource cost of a single AI query is negligible compared to a Google search plus the work it replaces. But aggregated across billions of queries a day, the numbers matter — especially in water-scarce regions where AI data centers have been sited. Three defensible positions for practitioners:

  • Don’t run a flagship reasoning model when a small model would do. A 7B local model at ~0.05 Wh/query beats o3 at 33 Wh/query by three orders of magnitude on environmental cost. Module 9’s SLM-RAG argument holds here too.
  • Batch and cache. The same levers that cut your bill also cut your footprint — caching prevents redundant compute, batching improves GPU utilization.
  • Location-aware siting matters. Hyperscalers (Microsoft, Google, AWS, Meta) publish location-level water/energy data. Where your inference runs isn’t a neutral fact.

There’s a reasonable argument that AI’s net environmental impact could be positive — if it accelerates climate modeling, grid optimization, and materials discovery. There’s a reasonable counterargument that the cost is large and uncertain enough to warrant restraint. Both views are live. Acknowledging the footprint at all is the minimum bar; ignoring it because the numbers are uncomfortable is not.

📋 Ready to Write Your Own AI Policy?

The Policy Templates appendix includes four ready-to-customize templates: Classroom AI Use Policy, Team/Department Guidelines, Organization-Wide Policy, and Student AI Disclosure. Start from a template — don’t start from scratch.

Module 10 Summary

Key Takeaways

  1. Bias is systemic, not accidental. AI automates and scales whatever was in the training data — from hiring to healthcare to criminal justice. The Amazon, COMPAS, Gender Shades, and Obermeyer cases are not outliers. They’re the shape of the problem.

  2. The fairness paradox. Chouldechova’s impossibility result: when base rates differ between groups, you cannot simultaneously achieve predictive parity, equal false-positive rates, and equal false-negative rates. “Fair” is a choice — make it explicitly.

  3. Privacy depends on the tier. Free consumer AI tools may use your data for training. Enterprise tiers with BAAs offer more protection. For regulated data, local (Module 9’s SLM-RAG) is often the only genuinely compliant option — and prompts are now discoverable evidence in litigation.

  4. Academic integrity is best handled through transparency (shared transcripts + disclosure + reflection), not through AI detection tools. Liang et al.’s 61.3% false-positive rate on TOEFL essays makes detector-driven discipline a civil-rights problem as well as an accuracy problem.

  5. Copyright for AI content, April 2026. Human authorship is required (Thaler cert denied). Purely prompt-driven output is not copyrightable; AI-assisted work with meaningful human control is. Training-data fair use is unresolved — three district courts, three different emphases, with Bartz v. Anthropic settled at $1.5B.

  6. Accessibility is both opportunity and risk. AI can break down barriers or build new ones. Design for universal access.

  7. Every organization needs an AI policy covering scope, permitted uses, prohibited uses, data handling, and a review cadence. The EU AI Act’s Aug 2, 2026 enforcement date is a useful forcing function.

  8. The NIST AI RMF (Govern-Map-Measure-Manage) is the cleanest vocabulary for talking about AI risk across procurement, audit, and regulatory contexts. It’s voluntary but widely adopted; for any formal AI project, speaking this language makes everything else easier.

  9. U.S. federal AI policy is in flux. Biden’s EO 14110 was rescinded January 20, 2025; Trump’s EO 14179 reoriented toward “AI dominance”; a December 2025 order preempts state-level rules. Write policies to the strictest applicable regime (EU AI Act + relevant state law); don’t count on federal rules being stable for more than a few quarters.

  10. AI has a real environmental footprint. A long reasoning-model query can use 33 Wh — >70× a small-model query. Data-center electricity demand is projected to double by 2030. The equity lever and the sustainability lever are the same lever: use the smallest model that solves your problem.

  11. Ethics isn’t a module you complete and forget. It’s a lens you apply to every tool, every prompt, every decision — for as long as AI is in the workflow.

💭 Think About It

The COMPAS algorithm didn’t set out to be racist. Its creators didn’t code in racial discrimination. They built a system that predicted recidivism using factors correlated with outcomes — factors that were themselves shaped by decades of systemic inequity in housing, policing, and employment. The algorithm reflected the world as it was, not the world as it should be. This is the fundamental ethical challenge of AI: it learns from the past, and we deploy it to shape the future. If we train on a biased world and deploy without intervention, we automate that bias at scale — and dress it up in the authority of a number. Every time you use AI, you are making a choice about which patterns from the past you want to carry forward — and which ones you want to break.

🔬 Lab 10: Write Your Own AI Policy

Lab 10: Write Your Own AI Policy

Time: 45 minutes | Tools needed: The policy template above + knowledge of your specific context

Part A — Context Analysis (10 min)

  1. Choose your context: a specific course you teach/take, your team at work, your department, or your organization.
  2. List: What AI tools are people already using? What tasks? What data is flowing through them?

Part B — Draft the Policy (20 min)

  1. Using the 6-section template, draft a real AI use policy for your context. Be specific — name the actual tools, the actual tasks, and the actual data.

Part C — Stress Test (10 min)

  1. Ask an AI chatbot to evaluate your policy: “Review this AI use policy. Identify 3 gaps or ambiguities that could cause problems. Suggest specific improvements.”
  2. Revise based on the feedback that holds up — ignore the feedback that’s generic.

Part D — Reflection (5 min)

  1. What was hardest? What will you need to revisit in three months as tools evolve?

Submit. Your drafted policy, the AI’s critique, your revised version, and your reflection.

🎮 Module 10 Quiz

Previous: Module 09 — AI Infrastructure | Next: Module 11 — AI Across Disciplines