Hallucinations

Reviewed by Proof · Updated

Definition

A hallucination is when a model states something false or made up as if it were true.

Key takeaways

  • Models predict likely text. Likely-sounding text isn't always true.
  • Hallucinations often look confident: fake quotes, wrong dates, made-up links, or functions that don't exist.
  • Giving the model the source text and asking it to answer only from that cuts them down a lot.
  • Letting the model say "I don't know" helps.
  • For anything that matters, check facts in code or with a person.

Why they happen

A language model learns patterns from text, not a list of checked facts. When it doesn't know something, it can still produce a fluent answer that fits the pattern. It has no built-in sense of which of its statements are true.

Some things make hallucinations more likely:

  • Questions about rare, recent or very specific facts.
  • Requests for exact numbers, citations or URLs.
  • Pushing the model to answer when it should decline.
  • High Temperature settings.

How to reduce them

  • Ground the answer. Paste the relevant source and say "Answer only from the text below." See Retrieval-Augmented Generation.
  • Allow uncertainty. Add "If the answer isn't in the text, say you don't know."
  • Ask for quotes. Have the model quote the lines it relied on, then check that they're really in the source.
  • Use tools. Let the model look things up or run code instead of guessing. See Tool Calling.
  • Check outputs. Validate code by running it, and check links and numbers before you publish them.

Example

curl https://api.luv13.ai/v1/chat/completions \
  -H "Authorization: Bearer $LUV13_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "luv13/glm-5.3-flash",
    "messages": [
      {"role": "system", "content": "Answer only from the provided text. If the answer is not there, reply: I do not know."},
      {"role": "user", "content": "Text: The library opens at 9 a.m. on weekdays.\n\nQuestion: What time does it open on Sunday?"}
    ]
  }'

A good reply here is "I do not know", because the text doesn't say.