AI hallucinations, explained
Short answer
AI hallucinations are confident answers from a chatbot that are false or unsupported: invented facts, altered quotes, fake court cases and citations to sources that do not exist. They happen in every major tool. In the EBU/BBC study, 45% of AI news answers had at least one significant issue.
What an AI hallucination is
An AI hallucination is output from a language model that reads as fact but is wrong, made up, or not supported by the source the model points to. The term covers more than one failure:
- Invented facts. A date, figure, name or event that is simply false.
- Fabricated sources. A citation, link, paper or court case that does not exist, or exists but says something else.
- Altered quotes. Real words changed, merged or attributed to the wrong person.
- Unsupported summaries. A summary of a document that adds claims the document never made.
- Repeated falsehoods. A false claim from the web, passed on as true.
The common thread is confidence. A hallucinated answer looks exactly like a correct one. ChatGPT Search signalled uncertainty only 15 times in 200 source questions and never declined to answer, while getting 134 of them wrong, in the Tow Center citation test (March 6, 2025).
Why it happens
A language model writes by predicting likely next words from patterns in its training data. It has no built-in check that a sentence is true. When the facts it needs are rare, recent, or missing from its training, it can still produce a fluent answer, because fluent text is what it was trained to produce.
Training and testing also reward guessing. OpenAI's own research argues that scoring models on accuracy alone pushes them to guess rather than admit they do not know. In its figures, gpt-5-thinking-mini and o4-mini had near-equal accuracy on SimpleQA (22% vs 24%), but the model that abstained 52% of the time had a 26% error rate, against 75% for the model that abstained 1% of the time (OpenAI, Why language models hallucinate, vendor claim, September 5, 2025).
Web search does not remove the problem. NewsGuard found that the share of non-answers from 10 leading chatbots fell from 31% to 0% in a year, while the rate of repeating false news claims rose from 18% to 35%. Live web search made chatbots answer more, and repeat more falsehoods (NewsGuard, August 2025 audit).
Newer is not always better. OpenAI reported that its o3 reasoning model hallucinated on 33% of PersonQA questions about public figures, against 16% for the older o1 (o3 and o4-mini System Card, vendor claim, April 16, 2025).
The full explanation, cause by cause, is on why does AI hallucinate?
How often it happens
There is no single hallucination rate. Each test measures a different task, so a tool's rate on one test cannot be set against another tool's rate on a different test. Within one test, the comparisons are clear.
| Test | What it measures | Result | Date | Source |
|---|---|---|---|---|
| EBU/BBC News Integrity | News answers from ChatGPT, Copilot, Gemini and Perplexity with at least one significant issue | 45% of 2,709 responses | Oct 21, 2025 | EBU and BBC |
| BBC study, round one | News answers from the same four assistants with significant issues | 51% | Feb 11, 2025 | BBC |
| Tow Center citation test | Wrong answers across 8 AI search tools asked to name a news article's source | More than 60% | Mar 6, 2025 | Columbia Journalism Review |
| NewsGuard False Claims Monitor | Prompts on false news narratives where 10 chatbots repeated the false claim | 35% | Sep 4, 2025 | NewsGuard |
| Vectara Hallucination Leaderboard | Lowest rate of summaries with unsupported content, best model on the board | 1.8% | Oct 22, 2026 | Vectara |
By tool, the spread is wide. In the EBU/BBC study, Perplexity had significant issues in 30% of news answers, ChatGPT 36%, Copilot 37% and Gemini 76%. In NewsGuard's audit, Claude repeated false claims 10% of the time and Gemini 16.67%, while ChatGPT and Meta AI both sat at 40% and Perplexity at 46.67%.
The table on AI hallucination rates by model lists every published rate we hold, with its test, date and source. For a ranking that weighs all of them, see the most accurate AI assistants.
How to spot one
Hallucinations rarely announce themselves, but they cluster in predictable places. Be most careful when an answer:
- Cites a source you cannot open. Grok 3 produced 154 broken or invented links in 200 answers in the Tow Center test. A link that fails is a warning sign for the claim beside it.
- Quotes someone word for word. In the BBC's first round, 13% of quotes attributed to BBC articles were altered or missing from the original (BBC, February 2025).
- Gives a precise number with no source. Exact figures look authoritative and are easy to invent.
- Covers recent or breaking news. Meta said its assistant asserted, in some answers, that the July 2024 assassination attempt on Donald Trump did not happen, and called these hallucinations (Meta, July 30, 2024).
- Names legal cases, papers or reports. Courts have now addressed AI-hallucinated content in 2,241 decisions in Damien Charlotin's database (count checked October 29, 2026).
- Summarizes a source it says it read. Google AI Overviews gave answers not supported by the sources they cited 56% of the time in the New York Times and Oumi analysis reported by Futurism (April 8, 2026).
Invented references are the most damaging kind for students and professionals. Does ChatGPT make up sources? covers how often it happens and how to catch a fake.
How to protect yourself
You cannot switch hallucinations off, but a few habits stop them reaching your work:
- Treat every answer as a draft. Use AI to find leads and structure, not as the final source.
- Open the sources yourself. Check that each link works and that the page says what the answer claims.
- Verify numbers, dates and quotes at the origin. Not in another summary of it.
- Ask for uncertainty. Tell the assistant it may say "I don't know". The OpenAI figures above show that abstaining cuts errors.
- Cross-check what matters. Run high-stakes questions past a second source, not a second chatbot.
- Pick the tool for the task. Tools differ by test: see is AI accurate? for each one.
The five-step routine on how to fact-check an AI answer turns these habits into a checklist you can run in minutes.
No AI assistant is free of hallucinations, and a confident tone is never evidence that an answer is right.
For which assistants hallucinate least across the published tests, see the AI accuracy rankings.
Questions
What are AI hallucinations?
AI hallucinations are answers from an AI tool that sound confident but are false or unsupported. They include invented facts, made-up citations, fake court cases, altered quotes and summaries that add claims the source never made. Every major chatbot produces them; in the EBU/BBC study, 45% of news answers had at least one significant issue.
Why do AI chatbots make things up?
Chatbots predict likely words rather than look up verified facts, so they can produce fluent text about things they do not know. Training and scoring also reward a guess over an honest "I don't know". OpenAI's own figures show a model that abstained 52% of the time made 26% errors on SimpleQA, against 75% for one that almost never abstained (vendor claim).
Which AI hallucinates the least?
It depends on the test. Claude repeated false news claims least often in NewsGuard's August 2025 audit, at 10%. Perplexity had the fewest significant issues in the EBU/BBC news study, at 30%. On the AA-Omniscience leaderboard, the model behind Gemini 4 Argon (High) has the lowest hallucination rate listed, at 4%.
Can AI hallucinations be fixed?
They can be reduced, not removed. Models that abstain when unsure make fewer errors, and grounding answers in documents helps. But even the best model on the Vectara leaderboard adds unsupported content to 1.8% of summaries (October 2026), and web search raised false-claim repetition in NewsGuard's audits.
How can I tell if an AI answer is a hallucination?
Check the parts that are easiest to invent: links, quotes, exact numbers, dates and named cases or papers. Open each source and confirm it exists and says what the answer claims. Be extra careful with breaking news and with any reference you plan to cite. The fact-check routine covers each step.
Sources
- Independent study
- Independent study
- Independent study
- News audit
- Benchmark
- Benchmark
- Vendor claim
- Vendor claim
- Vendor claim
- News audit
