One of the most important philosophical distinctions in the AI debate is the difference between what AI can do and what it ought to do. David Hume identified this gap centuries ago: you cannot derive a moral "ought" from a factual "is."

In education, this means that even if AI can grade papers faster, teach concepts more efficiently, and personalize learning better than any human teacher — it doesn't follow that schools should replace teachers with AI. The question requires engaging with values, not just capabilities.

The integration of AI into education is precisely such a matter. When someone argues "AI is better at teaching math, so we should use AI to teach math," they are committing Hume's error — leaping from a descriptive claim to a normative one without justifying the values that bridge the gap.

Sparrow and Flenady's 2025 paper "Bullshit Universities" (AI & Society 40, 5285–5296) articulates this clearly. Drawing on Harry Frankfurt's concept of "bullshit" — content produced without concern for truth — they argue:

1. It would be misguided to replace teachers with AI that generates text without understanding.
2. The history of automation in other industries shows it is naïve to think AI can assist teachers without eventually replacing them, because "money talks" — economic pressures will push institutions to cut costs.
3. The dream of AI teaching neglects three essential features: the importance of "learning how" for "learning that," teachers as role models, and the social nature of education.

The paper explicitly invokes the is/ought distinction, acknowledging it is possible for computers to replace teachers but arguing they ought not. My related blog post →

Gert Biesta's framework (2010, 2014, 2020, 2022, 2024) identifies three purposes of education:

Biesta argues that if education serves only qualification and socialization, it risks becoming training or indoctrination. Subjectification — involving freedom, ethical agency, and the "beautiful risk" of education — requires interruption, existential confrontation, and irreducibly human encounter.

Coelho et al. (2025, British Educational Research Journal) applied this directly to generative AI, arguing AI creates:

• A "placebo effect" — the illusion of subjectification without authentic effort
• A "nocebo effect" — students' retreat from becoming autonomous subjects due to offloaded responsibility

Subjectification is the "most misunderstood yet foundational dimension" — and it's exactly what's missing when AI bypasses the struggle of learning.

UNESCO released a suite of foundational documents at Digital Learning Week 2025 (September 2–5, Paris):

NYC Public Schools (March 24, 2026) released the most detailed U.S. district AI policy after 25 rounds of feedback with 1,000+ stakeholders. The framework uses traffic-light categories:

All AI tools must pass a 10-step data privacy review (ERMA). The explicit principle throughout: AI cannot do a teacher's job — it can support it.

The broader state landscape:

• Ohio (HB 96, August 2025) — Mandated formal AI policies for every public district by July 1, 2026
• Tennessee (March 2024) — First state to legally require all districts to publicly post AI use policies
• LAUSD (January 2026) — Introduced AI literacy curriculum in grades 6–12; adopted cautious stance after AllHere vendor collapse
• Chicago Public Schools — Designated 2024–25 as a "learning year"; full integration planned 2025–26
• White House (April 2025) — Executive Order "Advancing Artificial Intelligence Education for American Youth" established White House Task Force; "Pledge to America's Youth" secured 200+ tech company commitments

The EU AI Act (Regulation (EU) 2024/1689, in force August 1, 2024) classifies education as a high-risk domain under Annex III. Four specific use cases are regulated:

1. AI systems determining access/admission to educational institutions
2. Evaluating learning outcomes
3. Assessing appropriate education level
4. Monitoring/detecting prohibited behavior during tests

Key provisions:

• Requirements include risk management, data governance, human oversight, and accuracy standards
• Fines up to €35 million or 7% of global turnover
• Emotion recognition systems banned in educational institutions (effective February 2, 2025) — classified as "unacceptable risk"
• Article 4 mandates AI literacy for all staff deploying AI systems

For U.S. educators, the EU Act foreshadows the direction global regulation is heading. Even where it doesn't directly apply, it provides the most mature regulatory framework worth studying.

The evidence is simultaneously encouraging and cautionary:

Multiple rigorous studies confirm AI detection tools have structural limitations that should make anyone uneasy about punitive use:

Institutional response:

• At least 12 major universities (Yale, Johns Hopkins, Vanderbilt, Waterloo) have disabled Turnitin's AI detection
• Education Week reported Black students are disproportionately falsely accused
• New York City reversed its ChatGPT ban after recognizing prohibition was neither effective nor productive
• OpenAI itself stated detection tools are "not reliable enough given that educators could be making judgments about students with potentially lasting consequences"

The productive approach: reframe the question. Not "how do we catch cheaters?" but "what do we actually want from and for our students?" That's a question about values — which means it belongs to philosophy, not software.

Any policy built around prohibition is fighting a losing battle. The numbers are overwhelming:

The implication is clear: policies must help students use AI well, not pretend they won't use it at all.

Effective AI use policies must start with explicit value judgments grounded in normative ethical theory — not vague appeals to "responsibility." Here is the process I recommend:

1. Define your values using normative ethics. For example: "We ought to foster educational growth in our students." This is a value judgment — it requires philosophical justification. Which framework supports it? Utilitarianism (maximizing good outcomes)? Virtue ethics (cultivating character)? Deontological duty?

2. Test those values with thought experiments specific to your context. If you value transparency, does that mean students must disclose all AI use? What about spell-check? Grammar suggestions? Where is the line, and why? (See my thought experiments page for worked examples →)

3. Accept imperfection. Like the English language, your policy will never be perfect. Let go of the illusion that achieving perfection is possible and begin working with your team on common goals. As I tell my students: "We made it all up and we never have been and likely never will be perfect."

4. Involve stakeholders — educators, students, parents, and technology experts. NYC's 2026 AI policy went through 25 rounds of feedback with 1,000+ stakeholders. Multiple perspectives strengthen the framework.

5. Review continuously. Technology evolves faster than policy. Your policy must keep pace.

Read my full analysis on The Examined Classroom →

Surveys and professional development sessions surface the same concerns repeatedly. These aren't abstract anxieties — they represent genuine structural barriers to thoughtful AI adoption. Any ethical framework that ignores them is incomplete.

Two additional concerns surface persistently: over-dependence (students losing the capacity for independent thinking) and deepfakes (the erosion of trust when any image, audio, or video can be fabricated). Both are fundamentally about the same thing — the relationship between technology and human judgment.

The pattern across these concerns is instructive. None of them can be resolved by technical solutions alone. Each one requires a normative judgment about what education is for — which brings us back to ethical frameworks.

Rather than adopting a single ethical theory and applying it universally, a pluralistic approach maps each principle to the philosophical tradition that grounds it most naturally. This framework emerged from professional development work with K–12 educators and draws on classical and contemporary moral philosophy.

The value of this mapping is that it makes the philosophical commitments explicit. When a school board says "we value transparency," they are implicitly invoking Kant. When they say "we value equity," they are implicitly invoking Rawls. Making these connections visible helps educators reason more clearly about trade-offs — because different traditions sometimes pull in different directions.

Abstract principles become actionable when tested against concrete stakeholder impacts. The ethical matrix — adapted from Ben Mepham's work in bioethics — provides a structured way to do this. It cross-references stakeholder groups (students, teachers, parents/community) against ethical dimensions (wellbeing, autonomy, fairness) to surface tensions that broad principles alone can't capture.

For example, an AI tutoring system might score well on student wellbeing (personalized pacing) but poorly on teacher autonomy (if the system prescribes what gets taught next) and ambiguously on community fairness (if only some families can access it at home).

The matrix doesn't produce a single "right answer." Its purpose is to make hidden trade-offs visible so that the people making decisions — educators, administrators, families — can do so with their eyes open.

The theoretical frameworks above need grounding in real products that educators encounter. Each of these tools illustrates different ethical trade-offs — and each sits at a different point along the autonomy-versus-automation spectrum.

The common thread across these tools: the ones that work best ethically are the ones designed to keep the human — student or teacher — in the loop. The ones that raise the most concern are those that optimize for efficiency at the expense of agency.

One of the most important findings in recent AI-in-education research comes from Bastani et al., who studied high school math students given access to AI tutoring tools. The results were paradoxical:

This finding challenges the intuition that "more practice = more learning." It matters what kind of practice. If AI removes the productive struggle — the point where genuine learning happens — then the practice is hollow. Students may feel more competent while actually becoming less so.

The implication for educators: AI tools should be designed (or configured) to scaffold thinking rather than replace it. A tutoring system that asks "what do you think the next step is?" before revealing the answer preserves struggle. One that simply shows the solution does not.

Khan's aspiration is the right one. The question is implementation: does the AI actually tutor (ask questions, probe understanding, hold back answers), or does it merely dispense information? The difference matters enormously.

The trajectory of AI in education is not a single line — it is several concurrent developments, each with distinct ethical implications. Five areas deserve particular attention from educators thinking about the next three to five years.

Ethical frameworks become real when they encounter hard cases. Three thought experiments from the "Navigating the AI Frontier" presentation are designed to surface disagreements within educator teams — not to resolve them, but to make the underlying values visible.

These thought experiments are available in full on the Thought Experiments page, where each includes structured discussion prompts and connections to specific ethical frameworks.

Compliance asks "Is this allowed?" Ethics asks "Is this right?" The two questions overlap, but they are not the same — and educational leaders who treat them as identical end up reactive instead of principled. Ethics goes beyond legality: something can be permitted and still corrode the educational mission.

Ethics provides a toolkit for evaluating AI implementation decisions. Three traditions show up most often in education debates:

The Jisc Ethical Framework offers guidance tailored to AI applications in education — applying these traditions to concrete deployment questions like data governance, classroom transparency, and stakeholder consultation. It is one of the more usable starting points for leaders who want a structured ethics toolkit rather than a list of slogans.

The practical implication: "applied tech ethics" — fairness, transparency, accountability — is not a checklist but a system of trade-offs. A tool may score well on transparency and poorly on equity. Leaders need a framework that lets them name those trade-offs honestly, not just resolve them with a rubber stamp.

One of the most common arguments for calm in the AI conversation is that prior technological transitions also looked alarming and ultimately reshaped — rather than eliminated — work. The Industrial Revolution, agricultural mechanization, and the digital transformation are all cited as evidence that adaptation is possible.

The complacent reading is that AI is "just another transition." But four features distinguish it:

Historical parallels are useful as cautionary tales — they remind us that transitions can be managed and that doom-saying is rarely accurate. But the speed, scope, and target of the AI transition are genuinely new. Treating it as ordinary risks under-preparing for the actual shift.

Translating principles into practice requires structured conversation. The following four prompts — designed for use in leadership PD sessions — have surfaced productive disagreement across multiple cohorts of school administrators. Used well, they reveal the values a team holds without realizing it.

A leadership group that agrees on policy language but disagrees on Prompt 3 ("where should the locus of effort actually sit") doesn't actually agree on policy. The prompts are designed to expose those disagreements before they become operational problems.

The following case is designed for leadership discussion. It deliberately leaves the moral evaluation underspecified — different ethical frameworks will yield different verdicts, and that is the point. Three students each submitted excellent work, but their approaches to learning varied:

Three structured questions for discussion:

1. Assessing true understanding. What assessment strategies — beyond the submitted artifact — would surface the difference between Alex, Jordan, and Taylor? Oral defenses? In-class follow-ups? Process journals?
2. Adapting policies. A single policy must be fair to all three. Is "disclosed AI use is acceptable" enough, or does the policy also need to address depth of engagement?
3. Building trust. What role does transparency about AI use play in building trust between students and staff — and how does the policy signal that trust, not enforcement, is the goal?

The case has no single right answer. It is designed to make the team's actual operating values visible — including the ones the team did not realize it held.

This is the question every other AI-in-education question eventually reduces to. Detection, policy, assessment design, professional development — all of these are means. The end is being able to say, with justification, that learning has happened.

The honest answer is uncomfortable: in many traditional assessments, we never could say it confidently. AI didn't break the assessment system; it exposed assessments that were always proxying for learning rather than measuring it. Multiple-choice tests, formulaic essays, and book reports were vulnerable to substitution long before generative AI existed.

What changes the answer:

• Process visible alongside product. If the only artifact is the final draft, you cannot tell who learned what.
• Performance under variation. A student who has truly learned can apply a concept in a new context, not just reproduce it in the original one.
• Conversation, not just paperwork. A two-minute exchange about the work reveals more than a dozen pages of submitted text.
• Honesty about the question itself. "This student learned the material" is a value-laden judgment. AI hasn't made it harder to make — it has made it harder to avoid making.

The action plan that follows from this reflection is unglamorous: revise policies, organize teacher training to adapt assessment practices, evaluate tools that address AI challenges. None of these are individually novel. What is novel is doing them with a coherent ethical framework underneath — so that the policies, assessments, and training all answer to the same value commitments.