Do LLMs actually understand what they're saying?

Nobody actually knows for sure, and that's the honest answer — AI researchers are genuinely split on this, with surveys of the field showing something close to a 50-50 divide.

The skeptical case

An LLM is trained to predict the next word in a sequence of text, and that's it. It has no eyes, no hands, no memory of touching a hot stove — none of the grounded experience that humans build "understanding" out of. And it shows: the same model can explain a concept perfectly in one sentence and confidently contradict itself in the next, which isn't something a mind that truly grasped what it was saying would typically do.

The complicating case

These models solve word problems they've never seen before, walk you through their own reasoning step by step, and generalize to situations that look nothing like their training data. Pure memorization and pattern-matching struggle to explain that on their own. Some researchers now argue the models are doing something that at least functions like reasoning, even if it doesn't feel the way human reasoning feels from the inside.

The honest verdict: there's no scientific consensus here, and part of the fight is really about what "understanding" even means as a testable claim rather than a philosophical one. Anyone who tells you it's obviously settled — in either direction — is selling you certainty the field itself doesn't have.