A machine now lines up sentences that are more accurate, faster and often clearer than mine. The temptation is strong to conclude that it thinks. I believe that is a mistake, and above all that it has very concrete consequences for the way we use it day to day.
Predicting is not understanding
Thinking is not producing correct sentences. A generative AI works by predicting the next word from the immense mass of texts it has absorbed. It is extraordinarily good at this game, to the point where the result looks deceptively like reasoning. But it does not know what it is talking about. It has no world, no intention, nothing at stake.
As early as 2021, researchers offered an image that caught on to describe this mechanism: the clever parrot, which recombines brilliantly what it has heard, without understanding any of what it says. The image is a little reductive, because these systems do more than repeat, but it is right on the essential. Fluency is not understanding. You can produce a perfect discourse on a subject you do not grasp, and that is exactly what the machine does, at a dizzying scale.
The simplest proof is that it will serve you a false statement with the same confidence as a truth. It does not prefer the true to the false, because preferring assumes you understand the stake, and it has none. This is not thought, it is the production of plausible language.
Why this distinction is practical, not just philosophical
You might think all this occupies philosophers and has no effect on the life of a company. That is false, and I see it every week, because the answer you give to this question changes what you expect from the tool.
If I believe the machine thinks, I trust it like an expert. I take its answer for an authoritative opinion, I fall in line, I decide on its word. If I know it predicts the plausible without understanding anything, I use it quite differently, as a brilliant but judgement-free assistant, whose material I gather and whose every committing claim I check. The same sentence, out of the same machine, does not carry the same weight depending on what I believe it is. The whole difference between enduring the tool and using it lies in that belief, most often implicit.
That is why, in my work, I start with this question before touching any tool. Because once you have truly taken in that the machine does not understand, you stop treating it as an oracle, and you start using it intelligently.
The intelligence stays on your side
The real subject, at heart, is not whether the machine thinks. It is keeping in mind that the intelligence, in the loop, stays yours. The one that decides which question to ask. The one that spots when an answer sounds right but falls beside the point. The one that decides and owns it. The machine provides the material at a staggering speed. The meaning, you put it there, and no one else.
We will gain a lot by stopping the question, fascinated or worried, of whether the machine will start to think, and asking instead what it does to us, when we give it too much room in our reasoning. The first question is spectacular and distant. The second is modest, immediate, and it is the only one that changes anything about how you will decide tomorrow morning.