What Is an AI Hallucination?
Why AI chatbots sometimes make things up in a confident voice, and how to check and reduce it.
TL;DR
- An AI hallucination is an answer that sounds plausible but is false, misleading or invented, such as a made-up citation [1][2].
- It happens because language models predict likely text rather than check facts [2].
- Hallucinations can be reduced but not fully eliminated [2].
- The best defence is to ground the model in trusted sources and verify anything important yourself.
What a hallucination is #
A hallucination is a response that contains false or misleading information presented as fact. Typical examples are fake studies, web links that do not exist and wrong details about real people or organisations [1][2]. Some researchers prefer the word "confabulation", and others argue that "hallucination" wrongly makes software sound human, so you will see both terms [1].
Real cases show the stakes. A lawyer filed court papers that included case citations supplied by a chatbot that turned out not to exist [1][2].
Why hallucinations happen #
- It predicts, it does not look up. Generative models pick likely next tokens from patterns in training data, and they do not inherently know what is true [2].
- Training can reward guessing. Training on next-word prediction rewards producing something even when the model lacks the information, and some researchers argue that evaluations also favour guessing over saying "I don't know" [1].
- Flawed or thin data. Errors, gaps and biases in training text can become learned behaviour [2].
- Ambiguous or leading prompts. A question that assumes something false can invite a made-up answer [1].
- Adversarial input. Crafted prompts can steer a model towards misleading output [2].
How to reduce hallucinations #
| Approach | Idea |
|---|---|
| Retrieval-augmented generation (RAG) | Let the model look up trusted documents before answering [1][2] |
| Clear prompts | State the scope, give the source text and allow "I don't know" [2] |
| Human review | Keep a person in the loop for high-stakes content such as legal or medical text [2] |
| Better data and tuning | Train on curated, high-quality material, including examples where the right answer is "I don't know" [2] |
| Monitoring | Test outputs against known correct answers [2] |
Governments are also paying attention. NIST has published a risk-management profile aimed at helping organisations identify risks specific to generative AI [3].
How to spot a hallucination #
- Be careful with precise names, numbers, quotes and citations, since these are often invented.
- Ask for the source, then open the source yourself.
- Ask the same question another way and compare answers.
- Cross-check with a reliable website or expert before acting on health, legal or money questions.
A confident tone tells you nothing about accuracy [2].
FAQ #
Why does AI make things up? #
Because it generates statistically likely text rather than retrieving verified facts, so a gap in its knowledge can still be filled with fluent guesswork [2].
Can hallucinations be fixed completely? #
Not at present. Techniques such as retrieval and review reduce them, but sources say they cannot be eliminated entirely [2].
Does a longer context window stop hallucinations? #
Not by itself. Giving the model the right source text helps, but more text is not automatically better [4].
Is a hallucination the same as a lie? #
No. A lie needs intent, and a language model has none. The result is still a false statement, which is why checking matters [1].
Sources #
All sources accessed 2026-10-27.
- Hallucination (artificial intelligence). Wikipedia. https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence)
- What are AI hallucinations? IBM Think. https://www.ibm.com/think/topics/ai-hallucinations
- AI Risk Management Framework. National Institute of Standards and Technology. https://www.nist.gov/itl/ai-risk-management-framework
- Context windows. Anthropic documentation. https://platform.claude.com/docs/en/build-with-claude/context-windows