You open a chat window, you type a request, you wait for an answer. It is the most natural gesture in the world, and it is also the one that explains most of the disappointments. Because we do not address AI as a tool. We address it as an oracle, and an oracle that is wrong with confidence is more dangerous than a tool we know how to handle.
We want an oracle, we get a mirror
The first time you seriously use a generative AI, you expect one precise thing: the answer. The finished text, the finished image, the solution that drops into your lap. And sometimes it does, which keeps the illusion alive. But as soon as you move beyond trivial requests, the result starts to resemble what you gave it: vague if the request was vague, generic if the intention was soft, brilliant only if you already knew what you were looking for.
That is the first honest observation to lay on the table. AI is not an oracle that knows things you do not. It is a mirror that reflects back to you, amplified, the clarity or the fog of your request. When the answer is disappointing, the first question to ask is almost never “why is the tool bad?”, but “what did I actually ask it?”. That shift changes everything, because it puts the responsibility back on the side where it has an effect.
Saying this is not belittling the tool. A mirror that thinks fast, never tires and offers twelve variations in a minute is a considerable instrument. But an instrument. It has no intention, no taste, no stake. It wants nothing. The whole value of the exchange therefore depends on what you want, and on your ability to put it into words.
Psychiatrist or therapist
There is an image that helps to understand what is really going on. When a person is unwell, they can consult two kinds of professionals, and they do not work the same way. The psychiatrist makes a diagnosis, names the disorder, prescribes. You go to them for an answer from outside, an authority that rules. The therapist almost never gives the answer. They ask questions, they rephrase, they hand the ball back, until the person themselves puts words on what they already confusedly knew.
Most people approach AI as a psychiatrist: give me the verdict, give me the finished solution. And that is precisely where the relationship seizes up, because the tool then returns an answer of authority that is not one, plausible and hollow. Productive use looks far more like the therapist’s work. You use it to clarify, to bring things to the surface, to formulate an intention that was there but that you could not say.
This requires accepting a slightly vexing idea: on your own project, the answers are already in you, or in the client, or in the market. Somewhere, you know what this brand has to say, what this campaign has to provoke, why the first direction does not work. That knowledge is there, but buried, shapeless. The role of a good tool, like that of a good therapist, is not to bring you an answer from outside. It is to force you to bring out the one you were already carrying.
- “Give me the answer.”
- You expect a verdict from outside.
- You accept the first result delivered.
- When it fails, it is the tool’s fault.
- “Help me formulate what I am looking for.”
- You use the exchange to clarify.
- You follow up, refine, set aside.
- When it fails, you sharpen the question.
The real bottleneck: we are bad at asking questions
If the tool is a mirror, then the skill that counts is not knowing how to operate it technically. It is knowing how to ask a question. And here, we have to be frank: most people, including seasoned professionals, are fairly bad at this exercise. Not for lack of intelligence, but because nobody ever taught it. All of our training is about how to find answers. Almost none of it is about how to formulate questions.
We see it every day. People ask “make me a nice logo”, “write me a catchy text”, “propose a campaign”: orders that contain no context, no constraint, no intention. Then they are surprised by the interchangeable result. But no human creative director would do better with a request like that. They would start by asking twenty questions. The difference is that the human colleague asks those questions for you, whereas the tool takes your vagueness literally and hands it back as is.
It is good news disguised as bad. The quality of your exchanges with an AI is a fairly brutal indicator of the clarity of your own thinking. A confused request betrays a confused intention. In that sense, the tool does not replace thinking: it exposes its state. Those who progress fastest are not those who learn phrasing tricks, they are those who agree to clarify what they want before opening the chat window.
The pros, without indulgence
Since we are talking about a tool, we might as well name honestly what it does well. The first advantage is speed of exploration. Seeing ten directions in an hour, where it used to take ten, changes the nature of the work. You no longer cling to your first idea for fear of the cost of the next one. You can afford to set aside, to compare, to throw away.
The second advantage is the externalization of thought. Formulating a request forces you to get out of your head an idea that was going round in circles there. That effort alone clarifies, even before the answer. Many people discover what they want by trying to explain it to the machine, exactly as we often understand a problem by telling it out loud to a colleague.
The third is endurance. The tool does not tire, does not take offence, does not count its hours. You can ask it for the twentieth variation without guilt, test an absurd hypothesis just to see, go back. That freedom to iterate without friction has real creative value, provided you keep your hand on the judgment, which remains entirely human.
The cons, without panic
The other side must be stated with the same frankness. The most dangerous flaw is not error, it is confident error. The tool is wrong in the same perfectly assured tone it uses when it is right. There is no signal, no hesitation in the voice. It is up to the user to know, and therefore to check, which presupposes already knowing the subject. AI is a formidable accelerator for those who know, a trap for those who do not.
The second pitfall is levelling. A tool trained on everything that exists tends, by nature, toward the average of what exists. Without strong direction, it produces the plausible, the already-seen, the consensual. Very good for roughing out, dangerous for anything that has to stand out. A brand that delegates its visual writing to the average of the web ends up looking like the average of the web.
The third is the false impression of finish. A clean, well-presented result looks finished when it often is only on the surface. You are tempted to stop there, because it “does the job”. That is the precise moment when the craft reasserts itself: recognizing that a correct thing is not a right thing, and that the gap between the two is exactly what you pay a professional to close. The tool bears no responsibility for the result. You do.
Asking a better question
Everything therefore comes down to a single skill, and it can be learned. A good question, addressed to a tool as to a human, almost always contains four things: a context, a constraint, an intention and a refusal. The context situates: for whom, in what setting, with what history. The constraint bounds: what format, what length, what tone, what limits. The intention says what you are trying to provoke, not only to produce. And the refusal, often the most useful, says what you above all do not want.
It is not a magic recipe, it is a discipline of clarity. The same one that makes the difference between a good and a bad brief addressed to a human designer, a subject we covered in detail in our article on briefs. The good news is that this skill has nothing specific to AI. By learning to question a machine better, you mostly learn to question better, full stop: your clients, your colleagues, your own ideas.
“Suggest a slogan for a microbrewery.”
No context, no constraint, no intention. The result will be correct and interchangeable.
“Neighbourhood microbrewery in Montreal, a clientele of regulars more than tourists. I want a short slogan that says belonging to the neighbourhood, with no pun on beer, and without the word craft.”
Context, intention and refusal. The result will have an angle, even if a decision still has to be made.
In the end, an article about AI always ends up talking about something other than AI. The tool replaces neither taste, nor intention, nor responsibility for the result. It makes visible, faster and more crudely than before, the clarity or the fog of the person using it. Used as an oracle, it disappoints. Used as a therapist, it helps to formulate what you already knew. We go into this relationship with the tool in our articles on the prompt and creative direction and on AI that does not create but reveals (article in French). And if you want to see how this stance concretely changes a project, tell us about yours: the first thing we will do is ask you the right questions.