AI hallucination is when a language model produces fluent, confident text that is factually wrong or entirely invented. The term is worth knowing because a smooth answer is not evidence of a correct one, and it flags which outputs you should verify before trusting them.
A one-sentence definition
AI hallucination is a model stating a fact, citation, or number that sounds right but has no basis in reality. It happens because the model predicts likely words rather than looking anything up, so the sentence can read perfectly while the content is fabricated.
A concrete example
The clearest hallucination is an invented source. Ask ChatGPT to support a claim with a reference, and it can return a well-formatted link that points to a page which never existed. Answers can be wrong, so important claims and citations need review. Tools built around citations are not exempt: Perplexity attaches sources to each answer, yet a cited answer can still rely on a weak or unsuitable source, and its summaries may omit scope, caveats, or context from the original.
How it differs from related problems
Hallucination differs from an ordinary mistake because it has nothing to trace back to. A plain error usually comes from misreading real information, while a hallucination has no origin: the model fills the gap on its own. It also differs from outdated information, where the source was once correct but has since changed.
When to watch for it
Watch for hallucination whenever the output turns into a decision or carries facts you cannot afford to get wrong. Casual chat can absorb a small slip, but these cases deserve a second look:
- Specific numbers, amounts, or dates
- Citations, sources, and clickable links
- Legal, medical, or financial conclusions
- Anything you will forward or put into a formal document
The check itself is simple: put each key claim back against its original source. Even with a citation-first tool like Perplexity, open every source and confirm it actually supports the point, since model, file, and research allowances vary by plan and a listed source is not proof it was read in full. Before uploading anything sensitive, review the tool’s privacy and data controls first.
