Hallucination
In shortA hallucination is a fluent, confident AI answer that is false, invented or not backed by the sources it was given.
In plain words
A language model writes the text that fits best, not the text it has checked. When it lacks a fact, it can still produce a smooth answer, and that answer may be wrong. Researchers describe such output as nonsensical or unfaithful to the source (see the survey by Ji et al.). Anthropic's guide says the same: even advanced models can generate text that is factually incorrect or inconsistent with the context you gave.
Typical cases: an invented quote, a made-up link or court case, a wrong number, or a summary that adds a point the document never made.
Why it matters
- Risk: Wrong facts in legal, medical, financial or technical work cause real harm. The tone gives no warning, because wrong answers sound as sure as right ones.
- Cost of checking: You save time on drafting and spend some of it on verifying. Plan for that.
- No full cure: Anthropic states that its techniques reduce hallucinations but do not remove them, so validate what matters.
What helps, according to Anthropic's guide:
- Allow the model to say "I don't know".
- For long documents, ask for word-for-word quotes first, then base the answer on them.
- Ask for a source for each claim, and drop claims without one.
- Tell the model to use only the documents you provide.
- Run the prompt several times and compare; differences can flag a weak spot.
RAG applies the same idea at scale: it hands the model the right passages so it answers from them.
Example
You ask for three sources on a niche topic. The model lists three neat titles with authors and years. Two exist. One does not. You find out only by searching for it. A safer prompt supplies the papers and asks: "Answer only from these texts and quote the sentence you rely on."
Common misconceptions
- "A newer model never hallucinates." Rates may fall, but no provider claims zero.
- "If it sounds sure, it is right." Confidence in tone tells you nothing about accuracy.
- "Web search ends the problem." Search adds sources, but the model can still misread or misquote them.
Questions and answers
Why does AI make things up?
A language model predicts likely text. When it lacks a fact it can still produce a fluent answer, and nothing in the process checks it against reality.
How can I reduce hallucinations?
Give the source text, allow "I don't know", ask for quotes or citations, and check key facts yourself. These steps reduce errors but do not remove them.
Can hallucinations be eliminated?
Not today. Anthropic's own guide says its techniques reduce them but do not eliminate them.
Read next
- RAG (retrieval-augmented generation)
RAG (retrieval-augmented generation) first looks up relevant passages in your documents, then gives them to the model so it answers from them.
- Knowledge cutoff
The knowledge cutoff is the date after which a model's training data stops, so it knows nothing about later events unless you give it fresh information.
- Temperature
Temperature is a setting that controls how random a model's word choices are: low gives steady answers, high gives more varied ones.
- Prompt
A prompt is the input you give an AI model: your question or task, plus any instructions, examples and text it should use.
- Large language model (LLM)
A large language model (LLM) is an AI system trained on huge amounts of text to predict the next token, which lets it write, summarise and answer.