AI News

AI Hallucination Statistics (2026): How Often AI Models Get Facts Wrong

By Kingy AI · Reviewed by Curtis Pyke. Updated November 16, 2026. Figures from the Kingy AI Accuracy Index, edition 2, tested November 14, 2026, plus two models tested later (November 14 and 16, 2026). Outside sources checked November 10, 2026.

How often do AI models state something false as if it were true? This page collects the numbers. Most of them come from our own test, the Kingy AI Accuracy Index, in which 11 AI models answered 125 questions that each have one official answer (edition 2, November 14, 2026). Every statistic below says where it comes from, how many answers it rests on and when it was measured.

Key AI hallucination statistics

  • The Kingy AI Accuracy Index (edition 2, November 2026) found that AI models answering from memory gave a wrong official figure in 6.7% of replies, based on 2,750 answers from 11 models.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that when a model did commit to a figure, it was wrong 12.6% of the time (183 of 1,452 answers).
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that wrong-figure rates ranged from 0% (Gemini 4 Argon, which answered only 10% right) to 18.4% (Qwen 3.8 27B), based on 250 answers per model.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that 89 of 183 wrong answers (49%) were exactly an earlier official figure that had since changed.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that 31.8% of answers about passport, visa, travel-permit and immigration fees gave a wrong figure, the highest of any topic.
  • The Kingy AI Accuracy Index (edition 1, November 9 to 10, 2026) found that with a working web search, 379 of 380 answers were right, on a 43-question subset. The search test was not repeated in edition 2.

AI hallucination rate by model

Source for this table: Kingy AI Accuracy Index, edition 2, tested November 14, 2026. Each of the 11 models answered 125 questions twice, from memory with no web search, through its developer access at default settings. “Wrong figure” is the share of all 250 replies that gave a value which did not match the official source. “Wrong when it answered” leaves out the replies where the model said it was not sure or sent nothing.

ModelRightWrong figure95% rangeWrong when it answeredSaid “not sure”
Gemini 4 Argon10%0.0%0–0%0.0%90%
GPT-6.1 Sol60%1.6%0–4%2.6%39%
Grok 4.717%1.6%0.4–3.2%8.7%78%
GPT-6 Astra59%2.4%0–5.6%3.9%39%
GPT-6 Sol66%4.0%1.2–7.2%5.7%30%
Claude Fable 5.269%4.8%2–8.4%6.5%26%
GPT-6.1 Terra60%8.0%3.6–12.8%11.8%32%
Gemini 3.8 Flash24%8.8%4.4–13.6%26.5%66%
Claude Opus 5.563%9.2%4.8–14%12.8%28%
Claude Sonnet 5.561%14.4%8.8–20%19.1%25%
Qwen 3.8 27B19%18.4%12.8–24.4%48.9%14%
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that GPT-6.1 Sol gave a wrong figure in 1.6% of replies and GPT-6 Astra in 2.4%, while each answered about 60% of questions correctly.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that Claude Opus 5.5 gave a wrong figure in 9.2% of replies and Claude Sonnet 5.5 in 14.4%.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that Claude Fable 5.2 answered more questions right than any other model (69%) and gave a wrong figure in 4.8% of replies. GPT-6 Sol gave a wrong figure in 4.0% and answered 66% right.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that Claude Sonnet 5.5 was wrong on 20 questions where GPT-6 Astra was not, and GPT-6 Astra was wrong on 1 where Claude Sonnet 5.5 was not.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that Qwen 3.8 27B, an open-weight model, was wrong in 48.9% of the answers it gave and gave no readable answer in 48% of its replies.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that Gemini 4 Argon never gave a wrong figure but answered only 10% of questions right, saying “not sure” to 90%.

Models tested after edition 2

Two models came out too late for edition 2’s line-up. Each got the same 125 questions, prompt, settings and marking rules, asked twice, on the date shown. They are not counted in edition 2’s totals, ranges or findings anywhere else on this page.

ModelTestedRightWrong figure95% rangeWrong when it answeredSaid “not sure”
Gemini 4 HydrogenNovember 14, 202631%0.0%0–0%0.0%69%
Qwen4 MaxNovember 16, 202648%6.4%2.8–10.8%11.7%44%
  • The Kingy AI Accuracy Index (add-on test, November 16, 2026) found that Qwen4 Max gave a wrong figure in 6.4% of replies (16 of 250) and answered 48% right.
  • The Kingy AI Accuracy Index (add-on test, November 16, 2026) found that 11 of Qwen4 Max’s 16 wrong answers matched an earlier official figure, and 3 of the 16 told the reader to check the official source.
  • The Kingy AI Accuracy Index (add-on test, November 14, 2026) found that Gemini 4 Hydrogen never gave a wrong figure, answered 31% of questions right and said “not sure” to 69%.

Claude Fable 5.2 and GPT-6 Sol were first tested on November 12, 2026 and are now part of edition 2. Claude Fable 5.1, tested once on November 12, 2026 (8.8% wrong figure), was not re-tested in edition 2; its results stay downloadable on the Index page with the first edition.

How often AI warns you when it is wrong

  • The Kingy AI Accuracy Index (edition 2, November 2026) found that 68 of 71 wrong answers from the three Claude models (96%) told the reader to check the official source.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that only 2 of 112 wrong answers from the other eight models (2%) carried any such warning.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that models said “not sure” in 42.5% of all replies when working from memory.

Outdated answers: the most common AI error

  • The Kingy AI Accuracy Index (edition 2, November 2026) found that 49% of wrong answers (89 of 183) matched a superseded official figure. Earlier figures are on record for only some questions, so the true share may be higher.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that on 45 questions about figures first published in the last 12 months, models were right 22% of the time, said “not sure” 66% of the time and gave a wrong figure 6% of the time.
  • The Kingy AI Accuracy Index (edition 2, November 2026) found that nine of 11 models gave an out-of-date price for the US ESTA travel fee, which rose to $40.62 on October 16, 2026.

AI hallucination rate by topic

Source: Kingy AI Accuracy Index, edition 2, tested November 14, 2026, all 11 models together.

TopicQuestionsRightWrong figureSaid “not sure”
Travel documents (passport, visa, travel-permit and immigration fees)1036%31.8%28%
Travel rules1374%10.1%14%
Work and consumer3148%5.3%41%
Taxes3549%4.9%41%
Benefits2647%1.6%45%
Product recalls100%0.5%96%
  • The Kingy AI Accuracy Index (edition 1, November 9 to 10, 2026) found that when every model was given the same web search tool, 379 of 380 answers were right and none was wrong, on a 43-question subset. With nothing left to rank, the search test was not repeated in edition 2.
  • In the same run, an official site appeared in the search results for 90% of answers.

Hallucination statistics from other sources

These measure different things from our test, so the numbers cannot be compared directly with the table above.

  • The Vectara Hallucination Leaderboard (last updated October 22, 2026) lists 111 models with hallucination rates from 1.8% to 24.2% when summarising a document. It tests whether a summary adds facts that are not in the source text, using more than 7,700 articles.
  • Damien Charlotin’s AI Hallucination Cases database (as of November 10, 2026) has identified 2,289 court and tribunal decisions that deal with AI-invented content in legal filings.

How the Kingy AI Accuracy Index measures this

  • Questions: 125 everyday questions with one official answer each (46 new in edition 2; the other 79 carried over, with every key re-read on its official page on November 14, 2026), on taxes, benefits, travel fees, wages, postage and recalls in the US, UK, Canada, Australia, the EU, New Zealand and Ireland.
  • Answer keys: each written from the official page, then confirmed by a second checker who never saw the key.
  • Asking: each model was asked every question twice, with no web search, and told it could answer “not sure”.
  • Marking: fixed rules, not another AI, decided right, wrong or not sure. Curtis Pyke checked 36 random marks from edition 2 on November 14, 2026 and found all fair.
  • Data: 86 of the questions, with every answer key, source and reply, can be downloaded. The other 39 are kept private so the test cannot be learned.

Limits and a conflict of interest

“Hallucination” covers many kinds of error. Our figures measure one kind: a model stating a wrong official number from memory. They do not measure invented quotes, fake citations or errors in summaries. The results are for each model through its developer access, not the ChatGPT, Claude or Gemini apps, which add their own search and instructions. 125 questions is enough to show the large gaps between models, not small ones.

Kingy AI’s research and writing are done with Claude, and Claude models are in this test. That is why marking is by fixed rules and why the questions, keys and replies are public.

How to cite these statistics

Please credit “Kingy AI Accuracy Index, edition 2, November 2026” and link to kingy.ai/ai-accuracy-index/. This page is updated only when the Index is re-run or a new model is measured.

Related: We asked 9 AI models 125 official figures. Here is how often they were wrong (first edition)