Too Smart to Use It

I lean on an AI for a good part of my working day, and I have already written about what that help is worth and what it quietly costs. This is about a different thing, one that took me longer to notice: the number of people I know who are plainly sharper than I am, in their own field and out of it, who have looked at the same tool and decided it is not for them.

One of them is a lawyer who can find the load-bearing sentence in a forty-page contract faster than I can find my keys. She tried ChatGPT once, asked it something well inside her expertise, watched it cite a case that does not exist, and closed the tab the way you end a conversation with someone who has just lied to your face. That was two years ago. She has not been back. When I told her it had become the most useful instrument on my desk, she gave me a look I have come to know, the one that revises its estimate of your judgment gently downward.

I assumed people like her were the holdouts, a stubborn minority, and that the tide of the competent was flowing the other way. Then I went to look at who actually uses these tools, expecting the figures to sort the world neatly into the curious and the incurious. They did something stranger. Nearly everyone has now heard of ChatGPT; as of early 2026 just under half of American adults have used an AI chatbot at all, only about a quarter open one on a given day, and just over half never touch one (Pew Research Center). That much you might have guessed. The shape of the divide is the surprise. Use does not thin out as you climb the ladder of schooling and skill; it climbs with it. A person with a postgraduate degree is nearly three times as likely to have used ChatGPT as someone who left school at eighteen (Pew). The researchers who build and study these systems are wildly more hopeful about them than the public — fifty-six percent of the experts expect a good outcome against seventeen percent of everyone else, one of the widest gaps Pew has ever recorded on a technology.

So I had the question backwards. The puzzle is not why the unequipped stay away. It is why, among people with every advantage, so many still have not really begun — nearly half of those postgraduates have never opened it, and in the professions where it should pay off most, law and software and journalism and marketing, close to six in ten practitioners had not used it for their work (PNAS). And it is why, of the ones who do try, so many leave after a single encounter.

What I have come to think is that intelligence gets you to the doorway and no further. It wins you the free trial. What happens after is governed by five things that have little to do with how smart you are — and two of them, it turns out, get worse the smarter you are. The tool that was supposed to reward the clever has a couple of trapdoors built for exactly them.

The blank rectangle

Every tool you have ever used told you what it was for. A camera has a shutter; a calculator has a keypad; the purpose is moulded into the object, and you could no more make a phone call with a hammer than mistake one for the other. What arrives when you open ChatGPT is an empty rectangle and a blinking cursor. It is the first mass-market tool in history that ships with no built-in purpose at all.

That sounds like freedom. To most people it is closer to a small paralysis. Half a century of research into why anyone adopts a technology keeps returning to the same variable, and it is not power and not even ease — it is whether the person can picture the thing improving something they already do, what the literature calls perceived usefulness, the strongest single predictor of whether a tool gets picked up at all (Davis). A camera gives you that for nothing; its job is written on its face. The blank rectangle gives you nothing. You have to bring the job yourself, and if you cannot think what to bring it in the first minute, the tool has already failed its audition.

Run AI through the old checklist that predicts how fast an innovation spreads and it fails almost every line (Rogers). Is it plainly better than what it replaces? You cannot say, because you have not decided what it replaces. Does it fit the way you already work? No — it asks you to invent new habits. Is it simple? No. Can other people see the point of it when you use it? No; whatever happens, happens invisibly, inside a private window nobody else reads. It passes one test only: you can try it instantly, for nothing, at no risk. Which is why the trial numbers are astronomical and the actual habit is rare. Hundreds of millions of people have walked up to the rectangle, typed one thing, felt underwhelmed, and walked off. And because nobody had told them what it was for, most of them reached for the nearest familiar thing and decided it was a search engine.

It is not a search engine

Type a query into a search box and you get back a ranked list of things other people wrote. Type the same words into a language model and you get back something that has never existed before, assembled a word at a time out of probabilities — a plausible paragraph that might be right, might be confidently wrong, and comes with no links to check. The machine is not looking anything up. It is predicting what a good answer would look like. If you believe you are Googling, all of that is a defect. Once you know you are not, some of it becomes the whole point.

When researchers sat non-experts down with these tools and watched, the thing that most reliably wrecked the results was not a shortage of intelligence, it was the instinct to treat the machine like a person (CHI 2023). People were polite to it, believing courtesy earned better answers. They took one lucky prompt as a standing rule. They assumed that because it produced fluent language it must, in some way, understand them. The researchers' verdict was flat: our reflexes for talking to another human do not merely fail here, they "fundamentally mislead." And the wrong picture does not fade with practice — it sets. In one survey two-thirds of Americans allowed that ChatGPT might be conscious, and the more heavily they used it, the more mind they granted it, not less (Colombatto & Fleming). Familiarity, which dissolves most misconceptions, deepens this one.

The one-strike rule

Come back to my lawyer and the case that never existed. Her response — one lie, tab closed, finished — feels like nothing more than good sense. It is also, measured with care, a specific and slightly irrational reflex, and it has a name.

In a run of experiments, people watched a statistical model and a human forecaster both work, then bet their own money on whose later predictions to trust (Dietvorst, Simmons & Massey). The model was clearly the better bet; in some versions it made half the errors the humans did. It changed nothing. The instant people saw the model slip once, they dropped it and backed the human — often the very human they had just watched do worse. We keep a running credit line open for people and hand machines a one-strike rule. You take a dozen wrong turns a year and never once think to distrust your own driving; the satnav reroutes you badly a single time and you doubt it for a month.

It burns worse because the failures land where we least expect them. We can forgive a machine a hard problem and cannot forgive it an easy one, and a language model fails precisely on the easy ones — the fake citation, the wrong date, the function that was never in the library — while gliding through things that ought to be harder. Worse, it fails in the same confident, level voice it uses when it is right, so nothing in the tone flags the drop. The very authority that makes a correct answer persuasive makes a wrong one feel like a betrayal.

Here is the first trapdoor. You would expect expertise — knowing enough to catch the machine out — to make someone a better user of it. It does the reverse. In the same research, the people who discounted the model most sharply, and lost the most accuracy for it, were the experts: seasoned professional forecasters trusted the algorithm less than novices did, and were measurably worse off (Logg, Minson & Moore). The more your own judgment has earned its keep, the more a machine that contradicts you reads as an affront rather than a second opinion. My lawyer is not failing to understand the tool. She is doing exactly what expertise trains you to do — trust the instrument you have spent twenty years calibrating, which is yourself — and that reflex, which serves her everywhere else, is the one that shut the tab.

What repairs this in the lab, tellingly, is not a better model. It is a steering wheel. Let people adjust an algorithm's output even trivially, even in ways that make it slightly worse, and their willingness to use it more than doubles (Dietvorst 2018). It was never really about accuracy. It was about keeping a hand on the wheel. Which is why the shapes of AI that stick — the ones that offer you a line to accept or reject rather than promising to do the whole job — are the ones winning. They leave you holding the wheel.

It feels like cheating

There is a barrier here that has nothing to do with whether the tool works and everything to do with how using it makes you look. In a set of experiments with several thousand people, workers who used AI were judged by others as lazier, less competent, and less diligent than workers who produced the same thing another way — and, sensing this, they hid it (PNAS 2025). The tool that makes you faster also makes you look like you did not earn it.

The penalty attaches to the confession, not the work. Shown poems with no labels, readers liked the AI-written ones best; told afterward which was which, they marked the very same poems down (Porter & Machery). Not a word had changed. Only the story of where the words came from. Underneath sits an old instinct: we read effort as a proxy for worth. People rate a poem, a painting, a suit of armour as better the more toil they believe went into it (Kruger et al.). AI's entire offer is to delete the visible toil and keep the result, which strips out exactly the cue we use to decide a thing is any good. Something that arrives without evident labour feels, below the level of argument, like something that was not really made.

The consolation buried in that research is that the penalty is tribal, not moral. Whether owning up to AI helped or hurt depended entirely on who was judging: managers who did not use it themselves marked down the candidates who did, while managers who used it rewarded them. It is a border, and borders move; as more people cross, the shame drains out. But that is thin comfort while you are the one deciding whether to admit, in a meeting, that the sharp line came from a machine.

The frame gets installed early. A generation was taught in classrooms that reaching for AI is a species of cheating, and students now talk about using it in the vocabulary of self-reproach, lazy and dishonest, and conceal it even where no rule forbids it. And this is the second trapdoor for the accomplished. The more your standing rests on being the capable one, the person others come to, the more a tool that quietly does the thing you are admired for reads as a threat rather than a gift. Needing it is a confession. For a certain kind of expert the sentence "the machine did it better" is not a note about productivity; it is a small wound. Far easier not to open the tab.

Knowing what to ask

Suppose you clear all of that — you bring your own purpose, you stop treating it as Google, you make peace with its lies and with how it looks. One thing is still left, the least discussed and maybe the deepest: using AI well is a skill, unevenly held, and nearly invisible even to the people who have it.

The cleanest demonstration is a single experiment that different people quote as proof of opposite things (Dell'Acqua, Mollick et al.). Hundreds of consultants were given a batch of tasks. On the ones that sat inside the model's range, those with AI did far more, far faster, at markedly higher quality — and the weakest performers gained the most, the tool hauling them up toward the best. That is the "AI levels the field" study everyone cites. But the same experiment slipped in a task built to sit just outside the model's competence, and there the picture inverted: people using AI reached the right answer less often than people working alone, because the machine was confidently wrong and they went along with it. The ones who came out ahead were not those who leaned on it hardest. They were the ones who knew when to stop.

What divides those two groups is not typing speed or secret phrases. It is judgment about the tool — knowing which of your problems it can actually take, and feeling the wrongness in a fluent, confident, mistaken reply. When a study in Kenya handed the same AI assistant to small-business owners, the strong performers pulled ahead and the struggling ones went backwards, and the difference was not how they worded their questions but which questions they brought: the strong asked things that could be answered, the weak asked the tool to fix a drought (Otis et al.). On a clean, well-defined task — draft this, condense that — AI genuinely does flatten the gap between strong and weak (Noy & Zhang). The moment the task becomes choosing what to ask in the first place, the old advantages come flooding back, because that choosing is itself the rarest skill.

This is the quiet engine under all the rest. The people who get the most from AI tend to be the ones already good at saying exactly what they want and noticing when an answer is off — which is to say the tool most rewards a particular articulate, sceptical, self-auditing cast of mind, and is close to useless, or worse, for someone who cannot yet tell a good answer from a plausible one. It does not hand everyone the same gift. It amplifies a talent some people have spent their lives building and others have not, and it does so smoothly enough that the talent stays out of sight. Which is also why the conviction that AI is useless seals itself shut: the people who use it least rate it lowest (PwC), take that as confirmation, and never build the skill that would change the verdict.

The honest part

I do not want to turn this into a sermon, because the sceptics are holding a few real cards. For all the adoption, AI still touches only a sliver of actual work — by one careful estimate, somewhere between one and five percent of hours (Bick, Blandin & Deming) — and most organisations that have "adopted" it have no result to show for it, with only a small fraction capturing real value (McKinsey). The jagged frontier cuts both ways: there are whole territories where the honest answer is that the tool will make you worse, and the person who senses that and keeps clear is not timid, they are calibrated. "I tried it and it lied to me" is a true report. My claim is narrower than "everyone should use it for everything," which is false. It is that these five reflexes make capable people write the whole thing off for the exact tasks where it would, in fact, beat them — and that writing-off costs more than it feels like it costs.

Not the barrier we assumed

I think about my lawyer and the case that was never real. She was not wrong about what she saw; the tool did lie to her, in the confident voice it keeps for the truth. She was wrong only in the size of the conclusion she took from it — that a thing which errs is a thing without use — and she was wrong in the most understandable way there is, the way her whole training pushes her to be wrong.

The people I know who get the most out of these tools are not smarter than she is. Most of them are not smarter than average. What they have is an odd, slightly unnatural composure. They can pick up a tool that has no purpose until they give it one, that will lie to their face without blinking, that they have to stop mistaking for a mind, that makes them feel faintly as though they cheated, and that only pays out if they already half-know the answer — and use it anyway, lightly, for the things it is good at, with both hands kept on their own judgment. That composure is not intelligence. In a couple of places intelligence works directly against it. Which means the barrier was never the one we assumed. It is not that some people are too dim to understand AI. It is that being clever was never the thing that was going to save them.