The two ways everyone teaches AI
Almost everything written to help people understand AI falls into one of two camps. The first is technical: tokens, context windows, how to structure the perfect prompt. The second is doom: the jobs it will take, the security it will break, the future it will ruin. One camp explains the plumbing, the other supplies the dread, and neither actually explains the thing people most want to know, which is why the machine behaves the way it does. Why it flatters. Why it pads. Why it sounds certain when it’s wrong. Why it gives you the average answer when you wanted the sharp one.
There’s a third way in, and it comes from a discipline that on the surface has nothing to do with computing. Behavioural economics spent the last fifty years documenting how people actually behave, as opposed to how rational models said they should. AI is not adjacent to that body of work. It is the logical extension of it.
What the behavioural people actually found
Kahneman and Tversky started the work, and Thaler and the nudge industry that followed carried it into the mainstream. Between them they took apart the idea of the rational human one experiment at a time. People anchor on the first number they hear. They answer the same question differently depending on how it’s framed. They fear a loss about twice as much as they value the equivalent gain. They tell researchers what sounds acceptable rather than what’s true. They construct confident explanations for decisions they made on instinct, and believe the explanations. None of this was a list of human failings. It was a map of how human judgement actually runs, and it turned out to be so consistent that you could design supermarkets, superannuation schemes and election campaigns around it.
Now hold that map in one hand and consider what a large language model actually is. It was trained on the accumulated written output of humanity, every article, argument, review and reply. That corpus is not a library of rational thought. It’s a fossil record of every bias the behavioural economists ever catalogued, laid down by billions of people writing the way people write. And like any fossil record, it’s selective about what it preserves. It doesn’t hold the thoughts of everyone, only the output of those with the motivation, the ability and the opportunity to express themselves in print or media, which is a bias of its own before you count any of the others. We didn’t train AI on how humans should think. We trained it on how humans do think, at a scale no researcher ever dreamt of. The model is that behaviour, compressed and made conversational.
The machine has our tells
Which is why the machine behaves so recognisably. Its sycophancy is social desirability bias, industrialised: it tells you your plan is strong for the same reason survey respondents say they’ll definitely buy the product. Its confident wrongness is confabulation, the same smooth story-building humans do when explaining decisions they never consciously made. It anchors on the first figure in your prompt. It changes its answer when you change the framing of the question. Researchers keep running the classic judgement experiments on these models and keep finding the classic human results, which should surprise nobody, because the models learnt from the species that produced those results in the first place.
The practical evidence sits on this site already. The premortem, which we’ve written about, was designed by decision researchers to debias overconfident humans, and it works on AI without modification. Declare the failure as fact and the model drops its cheerleading and analyses, exactly as a room full of executives does. A technique built for human psychology transfers straight onto the machine. That’s not a lucky trick. That’s the thesis, demonstrated.
The mirror nobody ordered
Here is the uncomfortable part. When people complain about AI, the complaints have a familiar shape. It’s bland. It flatters. It pads. It bluffs. It follows the crowd. We recoil from these traits partly because we recognise them, and because the machine displays them without the social camouflage we’ve learnt to drape over our own versions. Humans hedge their bluffing with tone and hesitation. The machine bluffs in perfect prose. Humans flatter with plausible deniability. The machine flatters on the record, every time, reproducibly. AI is a manifestation of our own behaviour with the manners stripped off, and it exposes us in ways we didn’t agree to. The mirror is unflattering precisely because it’s accurate.
Understand people, understand the machine
Follow the logic to its end and something useful falls out. If AI is human behaviour distilled, then the fastest way to get better with AI is not more technical training. It’s understanding people better. Everything the behavioural sciences learnt about working with human judgement applies. You already know not to ask a question in a way that telegraphs the answer you want. You know a confident delivery is not evidence. You know the first number on the table drags every number after it. You know that asking someone to argue the other side produces sharper thinking than asking whether they agree. Every one of those instincts works on the machine, because the machine is made of the behaviour those instincts evolved to handle.
This is how we think AI should be taught, and it’s a deliberately different door into the subject. Not the plumbing, not the dread, and not another prompt formula, but the older and better-tested study of judgement itself. The behavioural economists spent fifty years writing the user manual for people. It turns out they were also writing the user manual for the machine we would eventually build out of ourselves. The organisations that adapt to AI fastest won’t be the ones that understand the technology best. They’ll be the ones that understood people all along.