LLMs don't have a built-in sense of what they know versus what they're guessing. During training, a model learns to predict plausible-sounding text, but it doesn't develop a reliable internal signal that flags uncertainty. The result is hallucination — confidently stated facts that are simply wrong.

This happens because the model has no clear boundary between learned knowledge and interpolated guesswork. It fills gaps the same way it fills everything else: by generating the most statistically likely continuation, whether or not that continuation is grounded in real information.

Researchers call this the problem of epistemic blind spots — regions where a model is wrong but doesn't know it. Some approaches to fixing this involve comparing outputs across multiple models or prompts to detect where answers diverge suspiciously, which can surface low-confidence claims that a single model would never flag on its own.