Hallucination
When an AI produces something fluent and confident that is simply false — fluency is not the same as accuracy.
Reviewed July 10, 2026Stable
When not to use it
- As a catch-all for every model error. Calling a formatting failure or a retrieval miss a "hallucination" hides the real bug and stops you fixing it.
- As a reason to distrust the model on everything. Hallucination rates vary enormously by task. Summarising a document you supplied is not the same risk as recalling a citation from memory.
Reach for something else instead
- Grounding with retrieval so answers come from supplied text rather than memory.
- Constrained generation — pick from a list, fill a schema — removes the room to invent.
- Verification passes, where a second call checks the first against the source, catch more than prompt-based pleading.
Read more on the blog
- AI in journalism: 45% of news answers had a flawTwenty-two broadcasters in eighteen countries evaluated 3,000 AI answers about the news. Forty-five per cent carried a significant issue, and the worst performer failed on 76%.
- AI in law: 1,313 filings sanctioned in 106 countriesA researcher has catalogued 1,313 court proceedings involving AI-fabricated content, 496 involving licensed attorneys. Single-matter sanctions went from $5,000 to $55,597 in two years.
- Why AI hallucinates: the confident lie is a feature, not a bugAI models don't hallucinate because they're broken. They hallucinate because we trained and scored them in a way that rewards confident guessing over honest uncertainty, and that has a mathematical floor. The real mechanism, the 2026 research that pinned it down, and what actually reduces it.
- How to reduce AI hallucinations: what actually worksYou cannot instruct a model into being truthful, because the process that invents a fact is the same one that recalls a real one. Every technique that measurably reduces hallucination works by adding something outside the model. Here is what works, in order of impact, and what only appears to.
Further reading
- Ji et al. (2022), Survey of Hallucination in Natural Language Generation — the taxonomy worth having before you use the word.
- Maynez et al. (2020), On Faithfulness and Factuality in Abstractive Summarization — hallucination measured on a task where the source text was right there.
- Bender et al. (2021), On the Dangers of Stochastic Parrots — the argument that fluency without grounding is the design, not the bug.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Asking the model not to hallucinate. It cannot tell when it is; that's what makes it a hallucination.
- Trusting cited sources without checking. Fabricated citations often have real-looking authors, journals, and DOIs.
- Assuming RAG solved it. Retrieval reduces hallucination; models still blend, misread, and over-extend the retrieved text.
At a glance
FieldLanguage & LLMs
Core ideafluent but false output
Root causetrained for plausibility, not truth
Main defensegrounding + human verification
DifficultyBeginner
Where this sits
25 concepts come first. Understanding it opens up 2 more.
12Levelsteps in
25Needs firstconcepts
2Opens up1% of 310
2Areascrosses fields
Computed from the prerequisite graph, not assigned. How this works