Correct—they don’t “know” in the epistemic sense, but they do encode a latent world model that shows up as useful priors.

Put differently: GPT-4 isn’t a knowledge base, it’s a *Bayesian autocomplete* over dense vectors. That’s why it can draft Python faster than many juniors, yet fail a trivial chain-of-thought step if the token path diverges.

The trick in production is to sandwich it: retrieval (facts) LLM (fluency) rule checker (logic). Without that third guardrail, you’re betting on probability mass, not truth.