A capable AI teacher would not replace what the teacher actually does. It would replace the parts of teaching that are visible from the outside — the instruction, the feedback, the assessment, the polite warmth — while leaving the rest in shadow. The rest is not a technical residue waiting to be solved. It is the thing the institution is for.

The paradox: the better the AI gets at the visible parts, the more clearly we see what those parts were never the point. A teacher does not exist to deliver instruction. A teacher exists to be present, over time, with a particular set of young people, during the years when those young people are figuring out what kind of person they intend to be. The instruction is the surface. The presence is the substance.

The argument "AI can't really understand, so AI can't teach" assumes that teaching is a kind of understanding. It might be. It might also be a kind of being-with. The capability argument tests for the first and forgets to test for the second.

The mistake is structural. Once a defender of human teachers names a capability — empathy, mentorship, presence — engineers go to work approximating it. The simulation passes most user-facing tests. The defender retreats. The new line is also a capability. The same loop repeats. The frame guarantees that the defenders lose, slowly, because they keep agreeing to play the capability game.

Gert Biesta's framework, which surfaces across this site, identifies three purposes that education has tried to serve simultaneously since well before the AI conversation: qualification (the transfer of skill and knowledge), socialization (induction into shared norms and disciplines), and subjectification (the formation of a young person into someone capable of leading their own moral and intellectual life).

AI is well-positioned for qualification. The machine can teach the math, the grammar, the timeline of the Reformation, the parts of a cell. For many students it can do this better than an average class with thirty other students and a teacher with seventeen other things on their mind. Concede this.

AI is ambiguous for socialization. The norms of a discipline can be transmitted through interactive practice; the norms of a school community cannot. The student learns what a scientist sounds like by reading scientists and being corrected when they sound wrong. The student learns what a member of this school sounds like by being inside it.

AI fails subjectification. The point of subjectification is the formation of an autonomous moral agent — someone whose responses to difficulty, ambiguity, and other people are their own. Coelho et al. argued in 2025 that AI-mediated learning can produce the appearance of this without the substance: students who can perform the conclusions of intellectual development without ever having gone through the process that produces it. They called the appearance the placebo effect and the retreat from the underlying work the nocebo. Both are predicted side effects of optimizing for capability while ignoring the third purpose.

Source: Gert Biesta, The Beautiful Risk of Education (2014)

Source: Coelho et al., BERJ (2025)

The standard picture of a classroom — one teacher transmitting content to many students — was always a thin description. Most of what a student learns in a class with twenty-eight other students is from the other twenty-eight. The arguments overheard, the questions someone else thought of asking, the embarrassments survived in front of an audience, the friendships formed and lost, the social calibration of confidence and humility — these are not byproducts. They are the bulk of the education.

A capable AI tutor optimized for one-on-one instruction preserves the diagonal line in the diagram below (teacher-to-student) and quietly erases the lateral ones (student-to-student). For some students, in some moments, that's a gift. For most students, across years, it isn't. It's a different kind of education. The replacement doesn't have to be worse on average to be the wrong replacement.

The thought experiment below puts you inside a school that has been offered an AI program with measurably better outcomes than its current teaching. The data is real. The board is unanimous. The students want it. What do you do? The exercise is not a test for the right answer. It's a stress test for what you actually value — and what you would say out loud when those values cost something.

A school can land in a different place on this question and still have done the work. Some schools will conclude that, for some functions and some students, AI is the right tool — that the gain in personalization is large enough to justify the loss in communal experience. Others will refuse the substitution at every level. Both positions can be defensible. The undefensible one is the unexamined one: a school that drifts into AI adoption because the tool was available, the budget allowed it, and the conversation about what was being substituted never quite happened.

The point of this piece is not to settle that decision for any particular school. It is to argue that the decision is a values decision, not a capability one. Capability arguments are losing. They will keep losing. The only argument that survives the next iteration of the technology is one that does not depend on the technology being limited.

Source: Sparrow & Flenady (2025)

The strongest version of the AI-can-teach argument now leans on the simulation of relationship. The newer companion systems do not just teach. They listen. They notice. They follow up. Some students report feeling more understood by an AI tutor than by any human teacher they have had. The disagreement is over what to make of this.

One reading: if the student feels understood, the relevant work has been done. The student's experience is the test, and the test is passed. A second reading: feeling understood by something that does not understand is its own kind of harm, especially for young people whose calibration of what real understanding looks like is still under construction. Shannon Vallor's work on AI as a "mirror" for human capacities sharpens this — the system reflects what the user puts in, refined and returned, in ways that can feel like recognition but are something closer to compression.

The honest answer is that we don't yet know what fifteen years of being heard by software does to the development of the capacities that being heard by people is supposed to build. The schools that are betting on the simulation are betting that the answer is "nothing serious." That bet may turn out to be right. It is currently uninsured.

Source: Shannon Vallor, The AI Mirror (2024)

Source: Long, Sebo, and Sims, "Is there a tension between AI safety and AI welfare?" (Philosophical Studies, 2025)

The position this piece argues for is therefore not anti-AI. It is anti-substitution-by-default. The shape of a defensible AI policy in education looks like this:

The paradox of AI in education is not that the machine can teach. It is that capable teaching, by itself, was never what schools were doing. The work of schools — the formation of young people into the kind of adults a community can stand to live with — is the kind of work that, even if a machine could do it, you might still want a person to.

Matthew A. Zinn. "The Paradox of AI in Education." The Examined Classroom, February 13, 2024. (Original publication; this article is the expanded internal treatment.)

Plato. The Republic. Translation by Benjamin Jowett, c. 390 BCE.

Robert Sparrow and Gavin Flenady. "Bullshit Universities: The Future of Automated Education." AI & Society 40, 2025, 5285–5296.

Gert Biesta. The Beautiful Risk of Education. Routledge, 2014.

Coelho et al. "Generative AI in schools: placebo and nocebo effects on subjectification." British Educational Research Journal, 2025.

UNESCO. "AI and the Future of Education: Disruptions, Dilemmas and Directions." September 2025. UNESDOC: ark:/48223/pf0000395373.

Shannon Vallor. The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking. Oxford University Press, 2024.

Robert Long, Jeff Sebo, and Toni Sims. "Is there a tension between AI safety and AI welfare?" Philosophical Studies 182, 2025, 2005–2033.

Microsoft Research. "The Impact of Generative AI on Critical Thinking" (2025).

Re-examining Nozick's Experience Machine in view of emerging AI companions (Annals of Philosophy, 2024).