Philippe Buschini Posts

Last week, I talked about a point that’s often misunderstood: for AI, truth doesn’t exist.

Today, I’m taking the reasoning one step further. Because there’s an even deeper misconception: believing that an LLM is a knowledge base. It’s not. A language model generates probable word sequences, not verified facts. In other words, it recites with ease, but it never cites.

That’s exactly what I explore in my new article: why this confusion persists, and how to clearly distinguish between parametric memory and explicit memory, so we can finally combine them the right way.

An AI doesn’t lie. But it doesn’t tell the truth either. It doesn’t know what is true or false — it only calculates probabilities. Its “reasoning” boils down to predicting which word is most likely to follow the previous one, based on the billions of sentences it has been trained on.

The result can be dazzling: fluid, elegant, convincing. Yet this fluency is nothing more than an illusion. What we read is not verified knowledge, but a sequence of words that “fit.” Sometimes accurate, sometimes wrong, sometimes neither — without the machine ever being aware of it.

The real danger is not the AI itself, but our very human reflex: confusing coherence with truth. In other words, mistaking the appearance of knowledge for knowledge itself. It’s this subtle, almost invisible shift that opens the door to confusion, born of ignorance about how it works, and of overconfidence in what merely “sounds right.”