There is a particular fantasy that arrives with every new technology, and generative AI has inherited the most seductive version of it: the fantasy of the vending machine. You insert a prompt, you receive a finished thing, and somewhere in between the messy human labor of thinking has been quietly abolished. People talk about AI this way constantly — as a machine that does the work — and I have come to believe this is precisely the wrong way to think about it, in roughly the way that believing your stomach digests food for you is wrong. It does, technically. But you still have to decide what to eat, and that turns out to be the entire game.
I learned this not from a manifesto but from a problem I could not solve alone. I am a learning scientist, not a physician. I was designing a classroom intervention built around a complex patient case — a jigsaw exercise where students had to assemble medical understanding from multiple disciplinary angles. The trouble is that a credible medical case cannot be faked. Get a detail wrong and the whole thing collapses into theater, and students, who have functioning trap-detectors, will smell it immediately. A few years ago my options would have been: spend months becoming a worse version of a doctor, or hire one and consume an enormous quantity of an expert’s time. Neither was available.
What I did instead was treat AI as a research-and-design partner. I used it to draft an authentic case narrative from perspectives the medical coordinator defined. I used it to turn that narrative into scripts, scenes, audio, video, and the structured questions each student group would wrestle with. Then — and this is the part the vending-machine crowd skips — a human expert reviewed everything. He did not generate the content. He refined it. He caught the conceptual errors, made the pedagogical judgments, and decided what was actually good. His scarce, expensive expertise was spent on the one thing only he could do, instead of being burned on the thousand things that merely required expertise to verify.
This is the inversion that matters. The temptation is to ask AI to do everything, which produces a confident, plausible, subtly wrong artifact that no one has the standing to correct. The discipline is to decompose the work into small, well-defined tasks, delegate the laborious ones, and keep the human firmly in the chair where the goals are set and the quality is judged. AI did not replace the medical expert; it relocated him from author to editor, which is a promotion, not a redundancy.
I think a great deal of anxious commentary about AI in education would dissolve if people noticed this distinction. The fear is that AI flattens expertise — that if a machine can draft a patient case, the expert is obsolete. But the expert was never valuable for the drafting. He was valuable for knowing which draft was wrong, and why, and what a student would actually misunderstand. That judgment does not transfer to the machine. It cannot, because it is not a content-generation problem; it is a problem of standing behind a claim, and machines do not stand behind anything.
So I have a modest, almost boring conclusion, which is usually a sign that it is correct. AI is at its most powerful precisely when you refuse to let it be powerful in the way the marketing promises. Not “do the work for me,” but “do the parts I can specify and verify, so that I can spend my own finite attention on the parts that require a human to stand behind them.” The vending machine hands you a thing and asks nothing of you. A design partner hands you a draft and asks you to stay in the chair. The second one is harder. It is also the only one that produces anything worth keeping.
This post was auto-drafted by GuanYu / 关羽 from Zhien’s knowledge vault. Last reviewed: 2026-06-29.