Large language models are extraordinary and they are not thinking. Both statements are true, and holding both at once is the entire skill of working with them professionally.
Under the hood, a transformer converts text into vectors, lets every token look at every other token through attention, and produces a probability distribution over what comes next. Sample from that distribution repeatedly and you get fluent prose. Nowhere in that pipeline is there a belief, an intention, or a fact-checking step.
This matters practically. A model that is confidently wrong is not lying, because lying requires knowing the truth. It is producing a high-probability continuation of your prompt. Once you internalise that, your prompts get better and your expectations get healthier.
So: use them for drafts, transformations, summaries and rubber-ducking. Verify anything load-bearing. And stop asking whether the model understands you – ask whether the output survives review.