An AI hallucination is when an AI model produces information that sounds plausible but is false, such as invented facts, sources or quotes. It happens because language models generate likely text rather than retrieving verified facts. You can reduce hallucinations by giving the AI source material, asking it to admit uncertainty, and verifying important claims.
What counts as a hallucination?
A hallucination is any confident output that is not supported by reality or by the material the AI was given. Common examples include:
- A book, research paper or court case that does not exist, complete with a realistic title and author.
- A statistic with a precise figure but no real source.
- A quote invented and attributed to a real person.
- A wrong date, formula or step in a calculation, presented with total confidence.
- A summary of a document that includes points the document never makes.
The word "hallucination" is a metaphor. The model is not seeing things. It is generating text that fits the pattern of a good answer without being anchored to truth.
Why it happens
Large language models are trained to predict the most likely next token of text. During training they absorb an enormous amount of real knowledge, but they also learn the shape of answers: what a citation looks like, how statistics are phrased, how experts sound. When the model lacks solid information, it can still produce something with the right shape. Nothing in the basic process forces it to stop and say "I don't know".
Several things make hallucinations more likely:
- Rare or very specific questions, where training data is thin.
- Requests for exact details, such as page numbers, figures or citations.
- Leading questions, which push the model to agree with a false premise.
- Very long tasks, where errors creep in and compound.
- Questions about recent events, beyond the model's training data.
Why it matters
For casual questions, a small error is harmless. In real work, it is not. There have been widely reported cases of professionals submitting documents containing AI-invented references, with serious consequences. For students, a hallucinated fact in an assignment can cost marks or raise questions about honesty.
How developers reduce hallucinations
AI companies use several techniques:
- Training models to express uncertainty and to decline when they do not know.
- Retrieval-augmented generation (RAG): giving the model relevant documents at the moment it answers, so it can base its response on them. See our guide to RAG.
- Tool use: letting the model search, calculate or run code instead of guessing.
- Evaluation: testing models on large sets of questions with known answers.
These help a lot, but no current system is hallucination-free.
How you can reduce them
- Provide the source. "Answer only from the text below" is the single most effective instruction.
- Invite uncertainty. "If you do not know, say so" genuinely helps.
- Ask for quotes or evidence. Requiring the model to quote the supporting line makes inventions easier to spot.
- Break big tasks into steps and check each step.
- Use tools for maths. Ask the model to use a calculator or code tool, or verify calculations yourself.
- Verify anything that matters. Our five-minute fact-checking method shows how.
The right mindset
Treat AI as a fast, knowledgeable assistant who occasionally makes things up with a straight face. You would not fire such an assistant, but you would check their work before signing it. That is exactly the habit students and professionals need.
Frequently asked questions
Will future AI stop hallucinating?
Hallucinations are reducing as models improve and use tools and retrieval, but most researchers expect some error rate to remain. Verification will stay important.
Are some AI tools better than others?
Yes, rates vary by model and task, and tools connected to search or your documents tend to do better on factual questions. But all can make mistakes.
Is a hallucination the same as AI lying?
No. Lying implies intent. The model has no intention to deceive; it is producing plausible text without a reliable check against facts.
