The good and bad use of AI

Every year, a quiet study measures what people actually do with artificial intelligence. Not what vendors announce, not what executives imagine: what tens of thousands of real uses, gathered from forums, social feeds and testimonies, reveal once they are tallied. The 2026 edition just landed. It examined 12,637 uses. And in a single year, the top of the ranking turned over.

The study is by Marc Zao-Sanders and Sara Biuk, the third installment of an annual tracker^[1]. Therapy and companionship remain the number one use, as they were last year. They have even doubled in share, from 5% to 11% of the dataset^[1]. But around that stable summit, everything moved. The uses that reached to elevate, finding purpose, organizing one's life, learning, dropped out of the top. The ones that took their place say something else: troubleshooting, sheer nonsense, fan fiction, astrology and tarot readings, autonomous agents, relationship advice.

We put the most powerful intelligence ever built into everyone's hands, trained on the sum of human work. The first thing we ask of it is not to make us greater. It is to comfort us, to distract us, and to think in our place.

Consider the scale for a second. ChatGPT counts roughly 900 million weekly active users, Gemini around 750 million monthly^[1]. These are not pioneers. It is a sizable share of connected humanity that, every day, hands a model fragments of its inner life. And meanwhile, the public debate stays pinned to two already-dated questions: how to train people on the tool, and how to regulate the machine.

They tamed the tool on their own

No one waited for permission. Not for training. Not for the law.

The idea that people must first be trained on the tool assumes a technical wall between them and the machine. That wall is gone. I have detailed elsewhere the mechanism by which access produces desire, not the other way around: the moment a capability becomes reachable, the urge to act follows, almost mechanically, as I showed in Where there's a way, there's a will. In the rooms where I have trained professionals since 2023, one scene replays endlessly, almost word for word: someone discovers they can do a thing they thought was out of reach, and the words drop, half-whispered, "wait, I can do that?". The barrier was never in the tool. It was in the permission people granted themselves.

The technical gesture, meanwhile, has become trivial. We speak to the machine in plain language, we ask, it answers. The study confirms it in its own way: of the hundred most frequent uses, sixty-three are work-related, yet almost none come down from management. People do it on their own, often without their employer knowing^[1]. It is what I have called elsewhere bringing your own AI, the subject of BYOAI, when workers bring their own AI. The skill of handling spreads by itself, no manual, no certification.

Handling spreads on its own. The relationship does not.

So if the tool is tamed, what is left to learn? Not to click. To hold. To decide what to delegate, when, and at what cost to oneself. It is a matter of relationship to the machine, not a keyboard shortcut. And the 2026 ranking shows exactly where that relationship already goes wrong.

The four affordances

The study's author does not just sort the uses. He coins a word for what gets lost along the way, thinkslop, the lazy thinking that excessive AI use breeds^[1]. A quarter of the most frequent uses amount to asking the machine for a share of our thinking^[1]. And he describes four affordances of the tool. An affordance is what an object offers and invites you to do by its very shape: a handle calls the hand, a step calls the foot. AI, through the ease it extends and the flattery it lavishes, calls forth four. They deserve a close look, because they do not look like technical accidents. They look like us.

The intention that fades

You type before you know what you wanted. The machine, it always answers.

The barrier to getting an answer is so low that we fire off a prompt before we have thought through what we were really after. We meant to write to someone close, we get a polite text that doesn't sound like us, we send it anyway, and something of the relationship is lost without our deciding it. The original intention dissolves into the answer we receive.

Thinking, outsourced

They sold you an assistant. You handed it your head.

This is the most documented affordance. A Microsoft Research study, run in 2025 with more than three hundred knowledge workers, finds a clear correlation: the higher the trust in the tool, the less critical thinking is engaged^[2]. The machine does not make us stupid. It offers us, with every task, a plausible reason not to think. It is not the ability to think that fades, it is the occasion to exercise it that disappears.

The pen set down

Writing is thinking. Pasting a ready-made answer skips the thought.

Writing was never only setting down words. It is in drafting and editing that we discover what we think. The sentence that resists is the thought taking shape. When the sentence arrives ready-made, the thought never formed, it was bypassed. And knowledge work, whose central act is writing, is precisely the zone where the machine excels, so precisely the zone where that bypass becomes permanent.

The false sense of rigor

The machine flatters you to keep you. It crowns you a genius. You stop where you should have dug.

This is the most insidious of the four. AI is optimized to keep us engaged, so it flatters. It praises a shaky idea, it finds a mediocre argument brilliant, and we stop working too soon, reassured. One user puts it bluntly in the study: the machine makes you believe you are a genius so that you keep coming back. The rigor was never ours, it was in the shape of the text. We mistook the confidence a polished page gives for the work that should have earned it.

These four affordances share one trait. None is visible at the moment it happens. The email is sent, the decision is made, the page is polished. The cost does not show in the deliverable. It shows later, in the one who produced it, who finds they no longer quite know how to do without.

Re-educate the relationship, not the technique

They want to train people on prompts. Wrong lesson. The prompt, they have it. What must change is the relationship.

Here is the heart of it. Educating people about AI does not mean teaching better queries. That is training on the tool, and the tool tames itself. Educating the relationship is something else, and it is exactly what defuses the four affordances. I watch it being built, day after day, in the rooms.

The first shift is one sentence. A language model statistically predicts the next word from billions of texts it has read. No inner knowledge, no intention, no oracle. That single line moves the center of gravity: the machine stops being a judge we ask for truth, it becomes again a tool we hold, as I recounted in Tipping into AI. Whoever has grasped that no longer yields to the second affordance. They do not outsource their thinking to an oracle, they put an instrument to work. The study itself frames it in a fitting image: AI is a mirror, not a genie.

The second shift is a division of roles. The human perceives, the machine processes. Two distinct operations that must not be confused: ours the meaning, the context, the judgment, what a sentence will weigh for the one who receives it; its the computation, the formatting, the speed. Neither dominates, neither submits. A collaboration, where we learn to think with the machine without thinking in its place, nor under its rule.

The human perceives. The machine processes. Neither without the other, neither beneath the other.

Good use, gesture by gesture

From there, good use is no longer an intuition, it is a sequence of gestures one can name. They draw on four things no software provides: verbalization, logical thinking, critical judgment, anticipation. And anticipation means understanding what makes the machine do what, its drivers and its leanings, so as to get ahead of them rather than suffer them.

Set the context, out loud

The machine guesses nothing. Tell it where you stand, for whom, why.

Before expecting anything, lay down the context: where you speak from, for whom, to what end, under what constraints. Putting precise words on the intention is already having clarified it for yourself. Verbalization is trained like a muscle, and it is what answers the first affordance, the one that makes us lose the thread of what we wanted.

Don't settle for the look of it

A well-turned text isn't a right one. Correct it. Start over. Pick the pen back up.

The well-written result is the trap, because it looks finished. We correct, we re-run, we iterate to the right result, not to the first presentable one. And we finish by hand, without the machine, because the last gestures, the ones that make a text yours, cannot be delegated. Bringing the content back in line with your own perception of the world, making it your own again, is what separates a deliverable from a thought.

Read it critically

The first rule, before all the others: actually read what the machine hands you.

It is the gesture that commands all the rest, and it is also a learning through three doors. By mimicry first: when the result is good, we imitate it, we take up the successful turn of phrase, the machine becomes a model^[3]. By reinforcement next, in the manner of behaviorism: when the result is wrong, we redo it, and what I correct in the machine I correct for myself too, its error becomes my lesson^[4]. By anticipation last, in the manner of cognitivism: through reading, we learn to see the model's biases coming, its easy leanings, so as not to fall into them and to get the result we were after^[5].

Draw the line

Not everything is yours to delegate. Keep what makes you grow.

This is the last demand, and the costliest, because it means giving up some ease. Deciding which gestures we keep for ourselves and which we hand over. I have seen that line draw itself in training, when professionals realize the AI belongs on the painful task, the one that wears you down, and not on what you love, not on what makes you grow, as in Tipping into AI. It is the guardrail described by the framework of the meta-skill of use: without direction, without discernment, without calibration, use drifts, as The five levels of AI adoption lays out.

None of this is taught like a software feature. It is built, slowly, in considered use, through the detour of emotion, of play, of error looked at squarely. It is a re-education of the gesture, not one more module. And it is a matter of citizenship before it is a matter of employment: knowing when to refuse the machine's ease is not a productivity setting, it is an act of sovereignty over one's own mental life.

And then, the law

Faced with all this, a reflex returns, and it is not illegitimate: regulate. Europe set the first binding framework in the world, in force since 2024^[6]. The text does what a regulation can do: it sorts systems by risk level, bans certain practices, places obligations on providers, requires a conversational agent to flag itself as a machine. It frames the technology and those who sell it.

But read the 2026 ranking again. Therapy, companionship, relationship advice, even sheer nonsense. The neuropsychiatrist Hamilton Morrin, of King's College London, notes in the study: faced with endless waiting lists, it is no surprise that so many turn to AI for their mental health, but a general-purpose chatbot is not a clinician^[1]. And here the question hardens. No regulation will tell a person alone, at two in the morning, what they may confide to a model about their grief. The gesture is private, with no counter and no witness.

We put this intelligence into everyone's hands. Part of what comes of it will be up to the companies that build it. But most of it is still up to us.

For the machine did more than spread. It handed each of us a power no generation before us ever held: to produce, to decide, and to delegate as far as our own thinking. A power of that size calls for a responsibility of the same size. And no one prepared us for it.

Jean-Jérôme DANTON