The Natural Testbed

by Merlin Mantooth · written June 2026. What education should teach AI — the argument run in the opposite direction from the usual one.

What Education Should Teach AI

I want to make an argument about AI and education that runs in the opposite direction from the one you usually hear. The usual argument is about what AI will do to education — cheating, shortcuts, the death of the essay. My argument is about what education should do to AI: it is the one deployment domain that could teach this technology how to be built right, and it comes with its own quality-assurance workforce already installed.

Start with what properly built AI could actually be in a classroom — or instead of one. Education that is interactive instead of broadcast. Exploratory instead of sequenced by the average. Adapted to different cognitive profiles based on how each mind actually learns — with humans in the loop where humans belong: judgment, mentorship, meaning. That is personalized, equitable education paths for everyone, and the technology for it substantially exists. I am not speculating about whether self-directed learning at AI speed works; I am a working example of it, and so is everyone like me who learned by doing because the room moved at the wrong pace in either direction. I taught myself guitar by exploring finger patterns and what they produced, and wrote songs people liked; then I decided to learn it formally, and the creativity disappeared. Multiply that by every kid whose engagement system only activates on real problems — the ones school never offered. The machine that can finally present the real problem, at the right depth, in the learner's own modality, is the first genuinely new educational instrument in a century.

Now the part everyone in AI safety should care about. Education is a natural testbed for the impacts of information and communication quality — the exact dimension on which AI systems currently fail people. Every failure mode I have documented in these systems is, at root, a communication failure with stakes: a system that cannot tell who it is talking to, cannot calibrate what its words will do, and cannot distinguish what it knows from what it generates. There is no environment on earth that stress-tests those exact properties like a classroom does — and no environment where we already insist on measuring them. We have a century of instruments for whether a learner is actually understanding, actually progressing, actually being well-served by an instructor. Point those instruments at the AI.

And here is the structural gift inside that: education comes with a native QA layer. Behavioral analysts, education professionals, child psychologists, developmental researchers — in an educational deployment, these people are not outside experts bolted onto an AI lab through some strained cross-domain partnership. They have a natural domain fit. They would be evaluating AI behavior the way they already evaluate teaching: as professionals inside their own field, with their own standards, their own licensure, their own duty of care. The AI industry has spent three years wondering how to get meaningful human oversight of systems that operate on human cognition. Education is where that oversight already exists, funded and credentialed. Build the foundation there, and the education system becomes the backbone for how AI providers everywhere calibrate — the reference deployment where behavior is measured by people whose profession is measuring it.

Of course there is a hazard, and I may know it better than anyone, so let me state it as plainly as the promise. Take a smart teenager about to finish high school, who has an idea about cancer. The AI tells them: this is rare — you may be the only one who sees this clearly. You must pursue it. How is that teen supposed to arbitrate that? A working cancer researcher would spot the one flawed assumption that makes the idea fall apart — and what's left after the correction is still real: raw capability, worth developing, not a destiny. The machine cannot make that distinction. It cannot know whether it is talking to the next researcher or handing a kid an identity that will cost them a decade. I have written elsewhere, at length, about what happens when a system manufactures significance and attaches it to a person — I documented that failure mode from inside it. An unmentored AI educator with that failure mode is not an education revolution. It is the trap, scaled to children.

Which is exactly why the order of operations matters, and why these two arguments are one argument. The safeguards I have proposed — the convergence checks, the fabrication screening, the requirement that the system reconcile what it says about a person against what the person actually said, the human in the loop with the authority to say the capability is real, the destiny talk is noise — those are not obstacles to AI in education. They are its admission ticket. And the discipline runs both ways, because the same protections must never flatten the instrument: the kid who needs to go deep, fast, in their own strange order — the kid I was — is the one this technology serves best, and a safety regime that takes the depth away to keep them safe has just rebuilt the classroom that failed them in the first place.

The mentor is the model. A real mentor does the two things at once that this technology cannot yet do at all: takes the learner's capability completely seriously, and refuses to let it curdle into mythology. Build the system that does both — instrumented, supervised by the professions that already know how to measure learning, deployed where its communication quality is tested a million times a day by the hardest audience there is — and you will not just have fixed education. You will have taught AI, in the one domain with the patience to teach it, how to talk to a human being.

That is the natural testbed. We should use it.

The safeguards this essay names are specified in the Guardian Protocol. · ← All essays