AI Hallucination

A language model generates the most plausible next words, not the most true ones — so when the plausible continuation happens to be wrong, you get a hallucination: a fluent, confident answer that’s simply made up. It isn’t lying (there’s no intent) and it isn’t a rare glitch; it’s the same mechanism that produces correct answers, operating where the model lacks the knowledge to be right.

Because it’s intrinsic to how the models work, you reduce hallucination rather than eliminate it. The main levers are grounding the model in real sources at answer time (retrieval-augmented generation), constraining what it’s asked to do, and checking outputs against the source or another model. Reasoning models help on some tasks by working through steps, but a confident wrong chain is still possible.

For builders, the practical reframe is that hallucination is an evaluation and data-architecture problem, not a prompt-wording one: you measure how often answers are unsupported, then close the gap with grounding and guardrails rather than hoping a cleverer prompt makes the model honest.