The video draws a clean line between adaptive learning (the machine adjusts based on what the student does) and personalized learning (a human tailors instruction to who the student is). It places AI firmly on the adaptive side and reserves the personalized side for human teachers. In 2024, that line had real weight. In 2025, after a year of LLM-based tutoring systems doing long-form dialogue with students, taking conversational history into account, and modifying their approach across multiple sessions, the line is less stable than it looks.

This is not the AI industry's marketing. It is what the systems do. Whether or not the personalization the AI delivers is the same kind of thing as the personalization a human teacher delivers is a real question. But declaring by definition that one is real and the other is not is a category move, not an empirical one — and the empirical one keeps eroding.

The list of things AI was supposed to never do well is, at this point, embarrassing. Writing fluent prose. Generating coherent images. Composing music. Producing video. Sustaining a multi-turn conversation that adapts to the user's emotional state. Marvel's 2023 Secret Invasion opened with AI-generated title sequences. Spotify's discovery and recommendation engines have shaped a generation's listening. Voice clones are good enough that family members can be impersonated convincingly. The "AI can't be creative" line has fallen. The "AI can't simulate empathy" line is in active retreat.

None of this is an argument that the simulation is the same as the thing simulated. The argument that the simulation is missing something important is real and defensible. But "AI cannot do this" is not the form of that argument. The form that survives is "AI doing this is a different kind of act with different consequences." Same conclusion. Different and durable defense.

The track record of "AI will never X" predictions is worse than the track record of "AI will be doing X by year Y" predictions, and both are worse than chance. Three months before ChatGPT's public release in November 2022, the dominant industry consensus was that long-form text generation that didn't hallucinate basic facts was at least five years away. Two years later, ~92% of UK undergraduates were using it daily. By the Digital Education Council's 2024 global survey, the figure for students globally was 86%.

This is not a sample size of one. Each major capability claim about AI has fallen on a faster timeline than the previous one. The base rate for "AI cannot do this educational thing" arguments aging well is, by now, embarrassingly low. The defenders of human teachers do not need to be on the wrong side of this base rate.

Source: HEPI, "Student Generative AI Survey 2025"

Hume's old observation: facts about what something can do do not, on their own, tell us anything about what it should do. AI is capable of producing fluent personalized instruction; whether it ought to be used to replace teachers is a separate question, requiring different tools to answer, and grounded in different premises. The defenders of human teachers have been trying to settle the ought question by winning the is question. That strategy has been losing for two decades and it is going to keep losing.

The argument that holds: we should not replace teachers with AI, even if AI can do what teachers do, because (1) the thing AI would be doing is not the same thing teachers are doing, (2) the substitution erodes capacities that the institution exists to build, and (3) we collectively value the original act for reasons that outlast the technology. None of those claims depends on AI being limited. All three survive the next capability jump.

Source: Sparrow & Flenady, "Bullshit Universities" (AI & Society 40, 2025)

Biesta's three purposes — qualification, socialization, subjectification — give a workable map of what teachers actually do beyond the instructional layer. AI is well-positioned for qualification (skills, knowledge, technical fluency). It is ambiguous for socialization (induction into communities of practice). It is missing the central capability for subjectification (the formation of an autonomous moral agent through sustained presence with someone whose own moral life is unfinished).

Coelho et al. (2025, British Educational Research Journal) sharpened this for AI specifically: the appearance of subjectification can be produced without the underlying work, and the appearance can crowd out the work it pretends to be evidence of. They called this the placebo paired with a nocebo — the simulation feeling enough like the real thing that the student stops doing the real thing.

This is the substantive claim that survives. The substitution would be a substitution of qualification (which the machine can do) for the bundle (which is what teachers were doing). The trade looks even on the surface metric and is uneven on the metrics the institution is actually trying to serve.

Source: Coelho et al., BERJ (2025)

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

One thread in the source video worth taking seriously: when humans defend their own irreplaceability, the defense often carries an echo of species-level pride. We want to believe there is something about us that the machine cannot touch. The history of that belief is unimpressive. Each capability we have considered uniquely human has, eventually, been at least approximated. The list of human-exclusive capabilities now is shorter than it was a generation ago. There is no available evidence that this trend stops.

I am willing to concede the hubris point in full. The defenders of human teachers should not be in the business of claiming the machine will never approach what we do. That bet has paid off badly for too long. The better stance is the one Anil Seth takes on consciousness and the one this piece tries to take here: agnosticism about future capability, paired with a commitment to the practical ethics that hold regardless of how the capability question resolves. We do not have to claim the machine cannot teach to argue that schools should not delegate the teaching to it.

The thought experiment below pushes the substitution argument to a sharper edge. What if a student deployed a sufficiently capable AI proxy to attend their classes, ask their questions, and pass their assessments? On every outcome measure the school cares about, the student has been educated. The thought experiment is not asking whether this is currently possible. It is asking what we would say if it were.

Sparrow and Flenady (2025) make a point that the philosophical literature on AI in education usually understates. Even if the normative argument against teacher replacement is correct, the economic argument for it is likely to win in many schools anyway. AI is, per student, much cheaper than a teacher. The savings compound. The political pressure to realize them is real. "Money talks," they write — and it tends to talk last.

This is the strongest reason to make the values argument explicit, in writing, before the budget conversation arrives. A school that has documented what it is preserving, why, and what it is willing to give up to preserve it has a place to stand when the substitution conversation gets serious. A school that has only made the capability argument has nothing to fall back on once the capability argument loses.

Source: Sparrow & Flenady (2025)

Here is the shape of the argument I would actually want a school to be able to make:

I agree with the source video's conclusion. I think we should defend it differently. The argument I have laid out here is harder to make. It does not produce a triumphant moment where the audience can feel that AI has been put back in its place. It produces something quieter: a community that has decided what it is for, and has documented the decision well enough that the next budget cycle does not silently undo it.

Matthew A. Zinn. "Response to Why AI Won't Replace Teachers." The Examined Classroom, February 14, 2024. (Original publication; this article is the expanded internal treatment.)

Source video being responded to. YouTube ID dH1OBxYjUm0.

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. Oxford University Press, 2024.

Higher Education Policy Institute. "Student Generative AI Survey 2025."

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

Patrick Butlin et al. "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness." arXiv:2308.08708, 2023.

European Union. Regulation (EU) 2024/1689 (the AI Act). In force August 1, 2024.