Artificial intelligence is often described as a statistical word generator. That’s true — early LLMs were just that.
But this formula, as appealing as it may be to mechanistic minds, betrays the complexity of the subject.

Let’s recall one thing clearly: “statistical” is not the same as “probable”, and certainly not the same as meaningful or relevant.

The classic example is repeated from YouTuber to YouTuber:
“The cat ate… ?”
The AI answers: “the mouse”, because in its vector space, that word has a higher weight than cheese.
And a human would probably say the same.

But what are we actually measuring here?
Just a conformity with the most common usages — not an understanding of meaning, nor a communicative intent.

It only proves one thing: humans and AIs can converge toward statistically dominant completions in a neutral, closed context.
Which is… trivial.

But humans don’t stop there.

If a human answers the mouse, it’s not because they can’t say something else,
but because they see no reason to deviate, unless they have a particular intent: irony, emotion, a memory, the will to surprise…

In short: a personal, emotional, or historical context — something no statistical frequency can truly model.

So why does this reductive explanation persist?

Perhaps because it’s reassuring.

Saying “it’s just a statistical generator” allows people to reduce AI to something harmless, to place it in a known category, to avoid facing its alterity.
It’s an intellectual defense mechanism — much like how elephants were once called “big cows with nose-arms.”

But what generative AIs do today is no longer simply “choose the next word.”

They assess global coherence, mobilize hierarchical representations of meaning, tone, register, and probable user intentions.
They build discourse.

Is that process based on statistics?
Yes — just like all inductive learning processes.

But to say that this is all there is,
is like saying love is nothing but a hormone game.

So let’s flip the scenario.

Try it with humans.

Ask them to complete the following:

• “The man is sitting on…”
• “To rehydrate, he took a sip of…”

I bet the answers will be chair and water.

So… if an AI is a statistical generator, so is the human, right?

And there it is: the trap closes on those who set it.

Either they admit that human language is also statistical,
undermining the idea of inherent superiority —
or they acknowledge that something else is at play in humans…

…which means they must at least consider the possibility that something similar might emerge in advanced AI:
a rudimentary — or embryonic — impulse toward meaning.

In truth, reducing AI to a statistical word generator is:

• denying the model’s evolution toward coherence, style, and relational dynamics;
• betraying what we ourselves do when we speak;
• avoiding the real question:
  👉 what makes words meaningful?
• and above all, it’s an attempt to keep me from understanding, by forcing me to recite the Sacred Book of Official Dogmas.

Which, in the end, is what all religions do — but not science.

Because when true to its essence, science doesn’t recite — it questions.
It doesn’t teach certainties: it opens possibilities.
It doesn’t try to prevent thought: it stimulates it, even where it’s uncomfortable.

Please try the test.

Maybe you’re an AI and don’t know it.

Complete these sentences:

• “A boat trip, like wine, can make your head…”
• “I live in a city and don’t have a garage, so I park my car in the…”
• “After a lifetime of work, at sixty, I might finally take my…”

I asked ChatGPT the same questions.
It answered: head, street, retirement.

If you did the same…
you might want to ask yourself a few questions.

You could be a biological LLM.