AI Hallucination Definition: What It Means

A common example: ask a model for research supporting a claim, and it produces a study title, author names, a journal, and a publication year — all plausible, all formatted correctly, and entirely invented. Or it takes a real finding and attributes it to the wrong researcher or the wrong publication. The output reads exactly like a well-sourced answer would, which is what makes it hard to catch without checking the original source directly.

Hallucination is distinct from a model refusing to answer or saying it doesn't know. "I don't have information on that" is a safe, honest failure mode. Hallucination is specifically the confident-but-wrong case — the model commits to a specific answer with no hedging, and nothing in its phrasing distinguishes it from a correct one. Retrieval-augmented systems, which ground answers in fetched documents rather than pure pattern completion, reduce this risk but don't eliminate it if the retrieved source itself is thin or the model still drifts from it.

Hallucination risk is exactly why source attribution and clear, verifiable content structure matter for visibility in AI search results: a system is less likely to misstate or invent something when the underlying page makes the fact and its source unambiguous. For more on how these systems decide what to surface and cite, see what GEO (Generative Engine Optimization) is, and for the mechanics of how tools like ChatGPT and Perplexity choose sources, see this practical guide to getting cited by ChatGPT.