An AI hallucination is a generated claim that is false, fabricated, or unsupported by the relevant source material. It can sound fluent and confident. Confident wording is not evidence that a model knows a fact.
What does a hallucination look like?
- A plausible article citation whose title, authors, or identifier do not exist.
- A summary that adds a decision absent from the meeting notes.
- A product answer that invents a feature or states an old policy as current.
- A response that presents an unknown date as a confirmed fact.
Inventing a character in a requested fictional story is different from inventing a source in a factual report. Creative generation is useful; fabricated evidence should not be presented as one of hallucination's benefits.
Why does it happen?
Language models generate responses using learned patterns and the context available to them. A plausible continuation need not be true. OpenAI's research discussion describes how training and evaluation that reward guessing can contribute to false answers. That is one explanation of the problem, not a claim that every error has the same cause.
A worked citation-checking example
Suppose an answer cites a paper called “Perfect Recall in Everyday Chatbots.” This is a deliberately invented title for the example.
- Open the supplied source link; do not rely on the citation's appearance.
- Check the title, author, date, and identifier against the publisher or repository.
- Find the passage supporting the specific claim.
- If the source is missing or does not support the claim, remove the claim or mark it unverified.
A real source can also be cited incorrectly. Finding a matching title is only the first check.
A prompt that makes checking easier
Answer using only the source passages below.
For each factual claim, give the passage label and a short supporting quote.
If the passages do not answer the question, state what is missing.
Do not invent citations, dates, or quotations.
Question: [question]
Passages: [permitted source material]This prompt can make evidence gaps more visible, but it does not guarantee factuality or faithful quotes. Compare the response with the original material yourself.
Apply the answer-checking routine, or use the learning exercise to check an explanation against course material.
What does not solve the problem by itself?
- Lowering temperature.
- Asking the model whether it is sure.
- Requesting sources without opening and reading them.
- Using a larger model or a longer prompt without verifying its claims.