Hallucination (AI)

In short: A hallucination is a convincing-sounding but false or invented answer from a language model.

In more detail: Language models predict the most likely next token; they do not check facts. When they lack knowledge, they fill the gap with something plausible, such as invented sources, numbers or laws.

In Depth

• Common with: rare facts, exact numbers, quotations, references, legal questions.
• Remedies: RAG with sources, follow-up questions, checking against a second source, low temperature.
• Recognising: answers that are too smooth with no uncertainty, sources that cannot be found.
• Responsibility: Whoever uses AI answers remains responsible for correctness.

See also: Temperature and sampling, Prompt engineering