Hallucination

A hallucination is a confident but false or invented statement produced by an AI model.

  • Using models
  • Updated

Language models generate plausible text instead of looking facts up, so they can invent names, numbers, quotations and citations. Hallucinations are more likely on obscure topics, in long outputs, and when a question has no good answer. Speech recognition models can hallucinate too, writing words that were never spoken; OpenAI says this can happen with Whisper.

You can reduce hallucinations by giving the model source material to work from, asking it to say when it does not know, using a stronger reasoning model and checking important claims yourself. None of these removes the risk.

Related terms

  • Retrieval-augmented generation (RAG) Retrieval-augmented generation lets a model answer from your own documents by retrieving the relevant passages and adding them to the prompt.
  • Benchmark A benchmark is a standardized test used to compare models, such as a set of maths, coding or knowledge questions; scores are useful but easy to over-read.

Keep reading

All terms