Hallucination

How it works

A hallucination is output stated with confidence that is false or unsupported: invented facts, citations, quotes, function names, API parameters or numbers. It happens because models generate plausible text rather than looking facts up.

Rates vary by model, task and domain, and makers measure them with different tests, so figures from different reports are hard to compare. Hallucinations are more likely on obscure topics, events after the knowledge cutoff, exact numbers and long sources the model must summarize. Some models are trained to say they don't know more often.

Give the model the source material and ask it to quote or cite it, use tools for lookups, and verify anything high-stakes. In code, rely on tests and type checks rather than trusting that an API exists.

Worked example

Illustrative: asked to load a config file with a less common library, a model writes load_config(path, strict=True). The function exists but has no 'strict' parameter, so the code fails at runtime; a test or a look at the docs catches it.

Common mistakes

• 1Trusting citations or links without opening them.
• 2Asking for a summary of a document the model was never actually given.
• 3Treating fluent, detailed answers as more likely to be correct.