What is an AI hallucination?
When a model states something false with total confidence — and why that is the normal case, not a glitch.
A hallucination is when a language model tells you something false, fluently and without hesitation. Invented citations, plausible-but-wrong dates, functions that don't exist in a library, a confident summary of a document that says something else.
The word is a little misleading. It implies malfunction. What's actually happening is the model working exactly as designed.
Why it happens
A language model generates likely continuations of text. It is not consulting a database, and — unless something has explicitly handed it a document — it is not looking anything up. Asked for a citation, it produces text shaped like a citation: plausible author, plausible journal, plausible year. Whether that paper exists is not a question the process asks.
There is no separate "truth" step. Fluency and accuracy come from the same mechanism, which is precisely why wrong answers read as smoothly as right ones. A model's confidence is not evidence that it is right. Its certainty of tone is produced by the same process whether the content is sound or invented, so a fluent, assured answer tells you nothing you can act on. That is the single most important thing to internalise.
They are not all the same thing
"Hallucination" gets used for several distinct failures with different fixes:
| Type | What it looks like | What actually helps |
|---|---|---|
| Fabrication | Invented papers, cases, APIs, quotes | RAG — supply real documents |
| Staleness | Confident answer that was true two years ago | Retrieval, or just check the date |
| Faithfulness failure | Misstates a document it was given | Ask for quotes; check them |
| Reasoning error | Arithmetic or logic goes wrong | Tools — a calculator, a code interpreter |
| Overreach | Answers a question it has no basis to answer | Prompt it that "I don't know" is acceptable |
Grouping these together is why mitigation discussions go in circles. RAG fixes fabrication and staleness. It does nothing for arithmetic.
What reduces it
- Ground it in documents. Retrieval is the highest-leverage fix by far. A model reading a real document is a different proposition from a model recalling one.
- Demand quotes. "Quote the sentence you based that on" is a cheap, effective check. A fabricated quote is far easier to catch than a fabricated claim.
- Give it tools. Don't ask a text predictor to do arithmetic. Give it a calculator.
- Permit uncertainty. Models are trained to be helpful, and an answer is more helpful-looking than a refusal. Explicitly saying "say you don't know if you don't know" measurably helps.
What does not reduce it
Asking "are you sure?" is not verification. The model will often revise a correct answer into an incorrect one simply because you implied doubt. Its self-assessment is generated by the same process that produced the original error.
Verification means checking against a source. There is no shortcut, and this is why everything published here carries its sources — so you can do exactly that.
What we addedA taxonomy of the distinct failure types people lump together as "hallucination", and which mitigations actually address which.
This article was researched and drafted with AI assistance from the sources listed below, then checked and edited by Fiqhro Dedhen before publication. How we work.
Sources
3 cited · 3 primary
- 1PrimaryarXiv (Huang et al.)A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions
Survey establishing the taxonomy this article follows.
arxiv.org · accessed 17 Jul 2026
- 2PrimaryarXiv (Lewis et al.)Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
The grounding technique that addresses fabrication and staleness.
arxiv.org · accessed 17 Jul 2026
- 3PrimaryarXiv (Ouyang et al.)Training language models to follow instructions with human feedback
Explains the helpfulness training that makes models reluctant to say "I don't know".
arxiv.org · accessed 17 Jul 2026