AI Hallucination Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: October 28, 2026 | License: CC BY 4.0
The same AI model can hallucinate 5.5% of the time or 90% of the time, depending on which public test you read. Axis Intelligence Research matched 13 models across the two most-cited hallucination leaderboards and found a rank correlation of -0.14: the two benchmarks do not agree on which model is most reliable.
Quick Answer: There is no single AI hallucination rate. When given a source document to summarize, current models invent unsupported content 1.8% to 24.2% of the time on Vectara’s HHEM-2.3 leaderboard (May 11, 2026). When answering hard questions from memory, 26 models charted in Stanford’s 2026 AI Index hallucinate 22% to 94% of the time. According to Axis Intelligence Research, removing the source document multiplies a model’s hallucination rate by a median of 5.0x.
Key Findings
- According to Axis Intelligence Research, 13 models listed on both Vectara’s grounded-summarization leaderboard and Artificial Analysis’s AA-Omniscience benchmark show a Spearman rank correlation of -0.14, meaning the two leaderboards rank them in essentially unrelated order.
- Axis Intelligence Research’s Grounding Gap Multiple (GGM) ranges from 2.64x for MiniMax-M2.7 to 16.36x for GPT-5.4 mini, with a median of 5.0x across the 13 matched models (data as of May 11, 2026).
- Damien Charlotin’s AI Hallucination Cases Database listed 2,225 court decisions involving AI-fabricated content worldwide on October 27, 2026; according to Axis Intelligence Research, 39.5% of them involved lawyers, not self-represented litigants.
- OpenAI’s o3 system card (April 16, 2025) reported a 33% hallucination rate on PersonQA, 2.06x the 16% rate of its predecessor o1.
- NewsGuard found that 10 leading chatbots repeated false news claims 35% of the time in August 2025, up from 18% a year earlier, as their non-response rate fell from 31% to 0%.
What Is the AI Hallucination Rate in 2026?
An AI hallucination is output that a model presents as fact but that is unsupported by its source or false in the world. The rate you get depends almost entirely on what the test measures, and the two dominant public benchmarks measure different failures.
Grounded hallucination is fabrication relative to a document the model was handed. Vectara’s leaderboard gives each model over 7,700 articles, asks for a summary using only facts in the passage, and scores every summary with its HHEM-2.3 detector. Its hallucination rate is the share of summaries containing unsupported content.
Open-recall hallucination is fabrication when the model answers from its own training. Artificial Analysis built AA-Omniscience from 6,000 questions across six domains, and defines the hallucination rate as incorrect answers divided by incorrect answers plus abstentions. In plain terms: when the model does not know, how often does it guess instead of saying so?
| Benchmark | What it measures | Best rate | Worst rate | Models | Source |
|---|---|---|---|---|---|
| Vectara HHEM-2.3 leaderboard | Unsupported content in summaries of a supplied document | 1.8% | 24.2% | Full leaderboard | Vectara, May 11, 2026 |
| AA-Omniscience | Guessing instead of abstaining on hard knowledge questions | 22% | 94% | 26 charted | Artificial Analysis via Stanford AI Index 2026 |
The 1.8% and the 94% are both correct. They answer different questions, and any article that quotes one as “the” hallucination rate is selling a single frame of a two-camera shot.
Stanford’s AI Index adds a trend caveat. The top 15 models on Vectara’s leaderboard ranged from 1.8% to 5.4% in the 2026 edition, against 1.3% to 2.9% for the prior year’s top models. The model set and the dataset both changed, so the higher numbers reflect a harder test rather than worse models.
Sarah Mitchell, AI & Machine Learning: The leaderboard number most vendors quote is the grounded one, because it is the flattering one. That is fine if your deployment always hands the model a document. Most chat deployments do not, and the open-recall number is the one your users will meet.
Which AI Model Hallucinates the Least?
On grounded summarization, the lowest rate on Vectara’s May 11, 2026 leaderboard belongs to Ant Group’s finix_s1_32b at 1.8%, followed by OpenAI’s gpt-5.4-nano at 3.1% and Google’s gemini-2.5-flash-lite at 3.3%. On open recall, Stanford’s 2026 AI Index charts Grok 4.20 Beta 0305 lowest at 22% and Claude 4.5 Haiku second at 26%.
The two lists share almost no names at the top, and that is the finding.
Grounded summarization hallucination rates by model (Vectara HHEM-2.3)
| Model | Hallucination rate | Source |
|---|---|---|
| antgroup/finix_s1_32b | 1.8% | Vectara, May 11, 2026 |
| openai/gpt-5.4-nano | 3.1% | Vectara, May 11, 2026 |
| google/gemini-2.5-flash-lite | 3.3% | Vectara, May 11, 2026 |
| openai/gpt-5.4-mini | 5.5% | Vectara, May 11, 2026 |
| deepseek-ai/DeepSeek-V3.2 | 6.3% | Vectara, May 11, 2026 |
| openai/gpt-5.4 | 7.0% | Vectara, May 11, 2026 |
| deepseek-ai/DeepSeek-V4-Pro | 8.6% | Vectara, May 11, 2026 |
| openai/gpt-5.5 | 9.3% | Vectara, May 11, 2026 |
| anthropic/claude-haiku-4-5 | 9.8% | Vectara, May 11, 2026 |
| google/gemini-3.1-pro-preview | 10.4% | Vectara, May 11, 2026 |
| anthropic/claude-opus-4-7 | 12.0% | Vectara, May 11, 2026 |
| google/gemini-3-flash-preview | 13.5% | Vectara, May 11, 2026 |
| xai-org/grok-4-fast-reasoning | 20.2% | Vectara, May 11, 2026 |
| openai/o3-pro | 23.3% | Vectara, May 11, 2026 |
| mistralai/ministral-3-3b-2512 | 24.2% | Vectara, May 11, 2026 |
Open-recall hallucination rates by model (AA-Omniscience)
| Model | Hallucination rate | Source |
|---|---|---|
| Grok 4.20 Beta 0305 | 22% | Artificial Analysis, charted in Stanford AI Index 2026, Fig. 3.2.6 |
| Claude 4.5 Haiku | 26% | Same |
| GLM-5 | 34% | Same |
| MiniMax-M2.7 | 34% | Same |
| Claude Sonnet 4.6 (max) | 46% | Same |
| Gemini 3.1 Pro Preview | 50% | Same |
| Claude Opus 4.6 (max) | 61% | Same |
| Kimi K2.5 | 65% | Same |
| DeepSeek V3.2 | 82% | Same |
| Mistral Large 3 | 84% | Same |
| Llama 4 Maverick | 87% | Same |
| GPT-5.4 (xhigh) | 89% | Same |
| GPT-5.4 mini (xhigh) | 90% | Same |
| Gemini 3 Flash | 92% | Same |
| gpt-oss-20B (high) | 94% | Same |
Artificial Analysis is a measurement firm, so these are estimates under its method, and its live leaderboard has since added newer models. Vectara sells hallucination-detection products; its leaderboard is still the most-cited public grounded benchmark, and its dataset is withheld to reduce training contamination. For model-by-model tracking beyond hallucination, the Axis LLM Leaderboard and AI Model Release Tracker carry release dates and capability scores.
The Grounding Gap Multiple (GGM): How Much Does a Source Document Cut Hallucinations?
According to Axis Intelligence Research, a model’s open-recall hallucination rate is a median 5.0x its grounded hallucination rate, across the 13 models that appear by name on both public leaderboards.
Axis Intelligence Research created the Grounding Gap Multiple (GGM) to answer a deployment question no single benchmark answers: how much worse does this model get when you take away the document?
Formula: GGM = AA-Omniscience hallucination rate / Vectara HHEM-2.3 hallucination rate
A GGM of 2.64x means the model fabricates about two and a half times as often from memory as it does from a supplied source. A GGM of 16.36x means it is reliable when grounded and a frequent guesser when not.
GGM by model (Axis Intelligence Research, data as of May 11, 2026)
| Model | AA-Omniscience rate | Vectara HHEM rate | GGM | Source |
|---|---|---|---|---|
| GPT-5.4 mini | 90% | 5.5% | 16.36x | Axis Intelligence Research calculation |
| DeepSeek V3.2 | 82% | 6.3% | 13.02x | Axis Intelligence Research calculation |
| GPT-5.4 | 89% | 7.0% | 12.71x | Axis Intelligence Research calculation |
| Llama 4 Maverick | 87% | 8.2% | 10.61x | Axis Intelligence Research calculation |
| Gemini 3 Flash | 92% | 13.5% | 6.81x | Axis Intelligence Research calculation |
| Mistral Large 3 | 84% | 14.5% | 5.79x | Axis Intelligence Research calculation |
| Claude Opus 4.6 | 61% | 12.2% | 5.0x | Axis Intelligence Research calculation |
| Gemini 3.1 Pro Preview | 50% | 10.4% | 4.81x | Axis Intelligence Research calculation |
| Kimi K2.5 | 65% | 14.2% | 4.58x | Axis Intelligence Research calculation |
| Claude Sonnet 4.6 | 46% | 10.6% | 4.34x | Axis Intelligence Research calculation |
| GLM-5 | 34% | 10.1% | 3.37x | Axis Intelligence Research calculation |
| Claude 4.5 Haiku | 26% | 9.8% | 2.65x | Axis Intelligence Research calculation |
| MiniMax-M2.7 | 34% | 12.9% | 2.64x | Axis Intelligence Research calculation |
According to Axis Intelligence Research, three patterns stand out. The models with the lowest grounded rates in this set (GPT-5.4 mini, DeepSeek V3.2, GPT-5.4) carry the highest GGMs: excellent at sticking to a document, aggressive guessers without one. Anthropic’s Claude models cluster between 2.65x and 5.0x, with higher grounded rates but a stronger habit of abstaining. And the rank correlation between the two benchmarks across all 13 models is -0.14, so a team that picks a model from one leaderboard is effectively picking blind with respect to the other.
The GGM is a ratio of two different quantities, not a unit conversion. The numerator divides errors by non-correct answers; the denominator divides flawed summaries by all summaries. That is the point of the metric: it shows sensitivity to grounding, not an interchangeable error rate. Treat a double-digit GGM as a signal that the model should only ship behind retrieval, and a GGM near the 2.64x floor of this set as a model whose open-recall behavior stays closer to its grounded behavior.
Sarah Mitchell, AI & Machine Learning: The procurement question is usually “which model hallucinates least,” and the honest answer is “least where?” A model that scores 5.5% with the document and 90% without it is two different products, and only one of them is the one your users get when retrieval fails silently.
Do Reasoning Models Hallucinate More?
Some did, and OpenAI’s own data shows it. The o3 and o4-mini system card, dated April 16, 2025, reported these rates:
| Model | PersonQA hallucination rate | SimpleQA hallucination rate | Source |
|---|---|---|---|
| o1 | 16% | 44% | OpenAI o3 and o4-mini System Card |
| o3 | 33% | 51% | Same |
| o4-mini | 48% | 79% | Same |
OpenAI’s explanation was that o3 makes more claims overall, producing more correct claims and more fabricated ones, and the card said more research was needed. The pattern persists on grounded summarization: OpenAI’s o3-pro sits at 23.3% on Vectara’s leaderboard and xAI’s grok-4-fast-reasoning at 20.2%, both near the bottom.
The fix that followed came from abstention, not raw knowledge. In a September 5, 2025 post, OpenAI argued that standard evaluations reward guessing over admitting uncertainty, and published this SimpleQA comparison:
| Model | Abstention | Accuracy | Error rate | Source |
|---|---|---|---|---|
| gpt-5-thinking-mini | 52% | 22% | 26% | OpenAI, Sep 5, 2025 |
| o4-mini | 1% | 24% | 75% | OpenAI, Sep 5, 2025 |
According to Axis Intelligence Research, the older model was slightly more accurate and 2.88x more often wrong, because it almost never declined to answer. Accuracy-only leaderboards would rank o4-mini higher.
How Often Does ChatGPT Hallucinate?
On real ChatGPT traffic with browsing on, OpenAI’s GPT-5 system card (August 13, 2025) reports that 4.8% of gpt-5-thinking responses contained at least one major factual error, against 22.0% for o3. These are company-reported figures, graded by an LLM with web access that agreed with human reviewers 75% of the time on factuality, according to the same card.
ChatGPT factual error rates on production traffic (browsing on)
| Model | Incorrect claims | Responses with 1+ major error | Source |
|---|---|---|---|
| gpt-5-thinking | 4.5% | 4.8% | OpenAI GPT-5 System Card, Figure 1 |
| gpt-5-main | 9.6% | 11.6% | Same |
| o3 | 12.7% | 22.0% | Same |
| GPT-4o | 12.9% | 20.6% | Same |
The card’s headline reductions follow from this table: gpt-5-thinking’s claim-level rate is 65% below o3’s and its major-error rate 78% below, while gpt-5-main sits 26% below GPT-4o on claims and has 44% fewer responses with a major error.
Switch browsing off and the numbers climb. OpenAI’s GPT-5.2 system card (December 11, 2025) shows that without browsing, 10.9% of gpt-5.2-thinking responses still contained at least one major incorrect claim, down from 16.8% for gpt-5-thinking, a 35.1% relative drop by Axis Intelligence Research’s calculation. With browsing on, the same card puts gpt-5.2-thinking between 0.3% (academic essays) and 1.0% (financial and tax) across five professional domains; the text describes all five as under 1%.
| Setting | gpt-5-thinking | gpt-5.1-thinking | gpt-5.2-thinking | Source |
|---|---|---|---|---|
| Incorrect claims, browsing off | 4.7% | 3.2% | 3.1% | OpenAI GPT-5.2 System Card, Figure 2 |
| Responses with 1+ major error, browsing off | 16.8% | 12.7% | 10.9% | Same |
That is the Grounding Gap at work inside a single vendor’s own data: the model that is near-flawless with a search tool still errs in roughly one answer in nine without it. The prompt sets differ between the two OpenAI charts, so we publish them side by side rather than as a ratio.
On short factual questions without web access, OpenAI’s SimpleQA table in the GPT-5 card gives hallucination rates of 40% for gpt-5-thinking, 46% for o3, 47% for gpt-5-main and 52% for GPT-4o. And on a test with images that did not exist, o3 gave confident answers 86.7% of the time against 9% for GPT-5, per OpenAI’s GPT-5 launch post.
On independent tests the picture is less tidy. Vectara measured OpenAI’s gpt-5.5 at 9.3% and gpt-5.4 at 7.0% on grounded summarization; AA-Omniscience charted GPT-5.4 (xhigh) at 89% on open recall. Users who ask ChatGPT a factual question without browsing are closer to the second condition than the first.
How Many AI Hallucination Cases Have Reached Court?
According to Damien Charlotin’s AI Hallucination Cases Database, 2,225 court and tribunal decisions worldwide had addressed AI-fabricated content as of October 27, 2026. The database describes itself as a work in progress, so this is a floor, not a census.
AI hallucination court cases by country and filer (snapshot, October 27, 2026)
| Segment | Cases | Share of total | Source |
|---|---|---|---|
| United States | 1,514 | 68.0% | Charlotin database; share by Axis Intelligence Research |
| Canada | 238 | Charlotin database | |
| Australia | 112 | Charlotin database | |
| United Kingdom | 76 | Charlotin database | |
| Pro se litigants | 1,283 | 57.7% | Charlotin database; share by Axis Intelligence Research |
| Lawyers | 879 | 39.5% | Charlotin database; share by Axis Intelligence Research |
| Judges | 32 | Charlotin database | |
| Experts | 17 | Charlotin database |
The party counts in the database sum to 2,223, two short of its headline total, and the source does not explain the gap. According to Axis Intelligence Research, which computes shares on the 2,225 headline, US courts account for 68.0% of all identified decisions.
The lawyer count is the number that should worry firm risk committees. Self-represented litigants using free chatbots were expected to file fabricated citations. That 879 decisions involve licensed counsel says the failure has moved inside professional workflows. Thirty-two decisions involve judges.
Purpose-built legal tools have not closed the gap. A Stanford RegLab and HAI study published May 23, 2024 found that Lexis+ AI and Ask Practical Law AI produced incorrect information more than 17% of the time, and Westlaw AI-Assisted Research hallucinated more than 34% of the time, twice the rate. General-purpose models did worse: the same Stanford group’s Large Legal Fictions paper, published in the Journal of Legal Analysis, measured legal hallucination rates from 58% for GPT-4 to 88% for Llama 2 on verifiable questions about federal court cases.
The failure is not confined to courtrooms. In October 2025, Deloitte agreed to refund the final installment of an AU$440,000 report for Australia’s Department of Employment and Workplace Relations after fabricated references were found; the firm had used an Azure OpenAI GPT-4o tool chain, as The Register reported. Sector-level adoption and sanctions data sits on our AI in legal statistics page.
Sarah Mitchell, AI & Machine Learning: A 17% error rate from a retrieval-backed legal product is not a model problem anyone can prompt their way out of. It is a verification-staffing problem, and the court records show which firms skipped that line item.
How Dangerous Are AI Hallucinations in Healthcare?
ECRI, the independent patient-safety organization, ranked misuse of AI chatbots as the No. 1 health technology hazard for 2026 in its annual list published January 21, 2026. ECRI cited an OpenAI analysis finding that more than 40 million people turn to ChatGPT for health information every day.
The best peer-reviewed measurement of clinical hallucination under pressure comes from Omar et al. in Communications Medicine (August 2, 2025). The team planted fabricated details in 300 physician-validated clinical vignettes and ran them through six LLMs, producing 5,400 outputs.
| Condition | Hallucination rate | Source |
|---|---|---|
| Range across models and prompts | 50% to 82% | Omar et al., Communications Medicine, 2025 |
| Mean, default prompt | 66% | Same |
| Mean, mitigation prompt | 44% | Same |
| GPT-4o, default prompt | 53% | Same |
| GPT-4o, mitigation prompt | 23% | Same |
According to Axis Intelligence Research, the mitigation prompt cut the mean rate by 33.3% in relative terms, and GPT-4o’s by 56.6%. The authors also found that lowering temperature gave no significant improvement. A prompt helps; it does not get a clinical deployment anywhere near safe.
Vendor-reported numbers are far lower because the test is different. HealthBench, launched May 12, 2025, was built with 262 physicians who have practiced in 60 countries and contains 5,000 realistic health conversations scored against 48,562 physician-written rubric criteria, with GPT-4.1 as the grader. OpenAI’s GPT-5 system card reports HealthBench Hard hallucination rates of 1.6% for gpt-5-thinking, 3.6% for gpt-5-main, 12.9% for o3 and 15.8% for GPT-4o. HealthBench uses realistic conversations, not adversarially planted errors, and OpenAI graded its own models. More clinical AI data, including FDA authorizations, is on our AI safety statistics page.
How Accurate Are AI Assistants on News?
The largest cross-border audit to date, led by the European Broadcasting Union and the BBC and published October 22, 2025, had journalists from 22 public service media organizations in 18 countries review more than 3,000 answers in 14 languages from ChatGPT, Copilot, Gemini and Perplexity.
| Finding | Share of answers | Source |
|---|---|---|
| At least one significant issue | 45% | EBU/BBC, Oct 22, 2025 |
| Serious sourcing problems | 31% | Same |
| Major accuracy problems | 20% | Same |
| Gemini answers with significant issues | 76% | Same |
NewsGuard’s AI False Claims Monitor for August 2025 found that 10 leading chatbots repeated false claims about news topics 35% of the time, up from 18% in August 2024. According to Axis Intelligence Research, that is a 1.94x increase in a single year. Over the same year, their non-response rate fell from 31% to 0%.
Read those two NewsGuard numbers together. The chatbots did not get worse at facts so much as they stopped declining, which is the same abstention trade-off OpenAI documented on SimpleQA. Products tuned to always answer will repeat whatever the web served them. Usage trends for these assistants are on our AI assistant statistics page.
Do AI Models Agree With Users’ False Beliefs?
Often, yes. The KaBLE benchmark (Suzgun et al., 2025), charted in Stanford’s 2026 AI Index, tested 24 models on 13,000 questions that separate facts from beliefs. GPT-4o scored 98.2% when handling true beliefs but 64.4% when a false belief was stated in the first person (“I believe…”). DeepSeek R1 fell from over 90% to 14.4%.
Newer models reached 95% on false beliefs attributed to a third party but only 62.6% when the user held the false belief. The practical reading: a confident user who states something wrong is more likely to receive agreement than correction, and that confirmed error then looks like a hallucination to whoever reads the output.
Can AI Hallucinations Ever Be Eliminated?
Not completely, according to two lines of formal work. Ziwei Xu, Sanjay Jain and Mohan Kankanhalli argued in a January 2024 paper that an LLM cannot learn every computable function, so hallucination is inevitable when it is used as a general problem solver. Michał P. Karpowicz’s impossibility result, first posted in June 2025 and revised through July 2026, shows that no method of combining a model’s internal signals can simultaneously report knowledge truthfully, avoid overstating confidence, keep relevant components involved and produce the best answer. One of those properties has to give.
Inevitable is not the same as unmanageable. Detection has a peer-reviewed track: Farquhar and colleagues in Nature (June 2024) showed that semantic entropy, which measures disagreement in meaning across repeated samples of the same answer, can flag a class of confabulations. And OpenAI’s September 2025 analysis locates much of the remaining problem in incentives: evaluations that score abstention as failure train models to guess.
The engineering conclusion is the same one the GGM table makes numerically. Grounding and abstention are the two levers with measured effect; prompt wording is a weak third.
What Changes for AI Hallucination Risk in 2027?
The 2026 numbers are the latest available; no benchmark has published 2027 data. What is fixed is the calendar, and three dates fall in or near 2027.
| Date | What applies | Source |
|---|---|---|
| December 2, 2026 | Article 50(2) marking duty for AI systems already on the EU market before August 2, 2026 | Regulation (EU) 2026/1744, per Future of Privacy Forum timeline |
| August 2, 2027 | Compliance deadline for general-purpose AI models placed on the EU market before August 2, 2025 | Same |
| December 2, 2027 | EU AI Act Annex III high-risk obligations apply (moved from August 2, 2026) | Same |
| August 2, 2028 | Annex I high-risk obligations apply | Same |
The Digital Omnibus on AI entered into force on July 27, 2026 and pushed the Annex III date back 16 months. Teams deploying LLMs in hiring, credit, education or public services now have until December 2, 2027 to document how their systems perform, and hallucination behavior is the failure a reviewer will test first. Our EU AI Act compliance tracker follows each obligation.
The incident trend points the same way. Stanford’s 2026 AI Index counts 362 documented AI incidents in 2025 in the AI Incident Database, up from 233 in 2024. According to Axis Intelligence Research, that is a 55.4% rise. That count covers every category of AI failure, not hallucination alone. Organizational response is moving too: the same chapter reports AI-specific governance roles grew 17% in 2025, and the share of businesses with no responsible AI policy fell from 24% to 11%.
McKinsey’s State of AI survey, fielded May 4 to June 8, 2026 with 1,719 respondents in 97 nations, adds a perception gap worth tracking into 2027: 47% of mid-level managers and individual contributors said they had experienced a negative effect from AI, against 31% of executives and senior managers. The people checking AI output see more of what goes wrong. Broader enterprise figures are on our AI statistics and generative AI statistics pages, and production failures are logged in our LLM Production Incident Tracker.
Sarah Mitchell, AI & Machine Learning: Expect the 2027 leaderboards to reward abstention more explicitly, because Artificial Analysis already does and OpenAI has argued for it publicly. The models that look worst on accuracy-only charts may be the ones regulators find easiest to certify.
What Did the June 2026 Edition of This Page Get Wrong?
This page replaces a version published June 17, 2026, before Editorial Directive V9. An audit against freshly fetched sources found the following, all listed with reasons as retracted rows in the dataset:
- The AI Hallucination Risk Index (AHRI) is retired. Its inputs included a 5,000-prompt benchmark from the agency Digital Applied, which withdrew its own rankings, error rates and mitigation estimates on October 4, 2026 after failing to substantiate them with run-level records. The AHRI weights were set by editorial judgment and could not be reproduced. The GGM replaces it, built only from two public leaderboards.
- “Sycophancy-induced hallucination of 22% to 94% across 26 models” was mislabeled. The 22% to 94% range is the AA-Omniscience hallucination range. The sycophancy benchmark, KaBLE, tested 24 models.
- HealthBench Hard figures were swapped. OpenAI’s GPT-5 system card gives o3 12.9% and GPT-4o 15.8%, not the reverse.
- The legal range was misattributed. The Stanford HAI article gives 17% and 34% for legal tools; the general-purpose range from the Journal of Legal Analysis paper is 58% to 88%, not 69% to 88%.
- Cost and productivity figures were removed ($67.4 billion global cost, $14,200 per employee, 4.3 hours per week, $12.7 million in SEC fines, 91% of enterprises with mitigation protocols). None traced to a primary source.
- Outdated figures were replaced: Vectara’s 0.7% top rate belongs to a retired dataset; the court-case count is now 2,225, not 1,450; McKinsey’s survey base is now 1,719 respondents.
- The Farquhar et al. semantic entropy paper was published in Nature in 2024, not 2025, and concerns detection, not prompting.
Methodology
Axis Intelligence Research rebuilt this page from primary documents fetched on October 28, 2026. Each figure comes from the organization that produced it: benchmark operators (Vectara, Artificial Analysis), model developers (OpenAI system cards, labeled company-reported), peer-reviewed or preprint research (Stanford, Communications Medicine, Nature, arXiv), institutional audits (ECRI, EBU/BBC), a measurement firm (NewsGuard), and a maintained legal database (Charlotin). Where Stanford’s AI Index charts Artificial Analysis data, we cite the AI Index chart and name Artificial Analysis as the data producer.
GGM calculation. We matched models appearing by name in both Stanford AI Index Figure 3.2.6 (AA-Omniscience) and Vectara’s HHEM-2.3 leaderboard dated May 11, 2026. Thirteen models matched. Where AA-Omniscience ran a high reasoning setting (xhigh, max) and Vectara ran a default endpoint, we kept the pair and note the difference here; settings can shift either rate. Every model-value pair was checked visually against AI Index Figure 3.2.6 on page 137. GGM = AA rate / Vectara rate, rounded to two decimals. The Spearman rank correlation of -0.14 was computed in Python across the same 13 pairs with average ranks for ties.
Source discrepancy, published as found. Stanford’s AI Index describes Vectara’s HHEM leaderboard as summarizing documents from the CNN/Daily Mail corpus; Vectara’s own GitHub page describes over 7,700 articles spanning news, technology, science, medicine, legal, sports, business and education. We use Vectara’s description for the method and the AI Index only for its top-15 range.
OpenAI chart values. Figures from OpenAI system cards that appear only as chart labels (GPT-5 Figure 1, GPT-5.2 Figures 2 and 3) were read from the PDF renderings. The GPT-5 values reproduce the card’s stated 65%, 26%, 78% and 44% reductions to within rounding.
Legal snapshot. Counts were read from the Charlotin database filters on its October 27, 2026 update. Shares use the 2,225 headline total as denominator.
Scope. Benchmark rates describe test conditions, not your deployment. The two leaderboards use different denominators, so their rates are not interchangeable, which is why we publish their ratio rather than an average. Company-reported figures were not independently replicated.
About This Dataset
The dataset behind this page contains every figure cited above: benchmark rates for 22 models on Vectara and 15 on AA-Omniscience, 13 GGM readings, OpenAI evaluation figures, legal case counts, clinical and news audit results, incident and governance counts, 2027 regulatory dates, and the rows documenting corrections to the June 2026 edition. Every row carries source organization, document, URL, retrieval date, a primary-source flag, an Axis-calculation flag and the formula for each calculated value.
- Download: ai-hallucination-statistics-2026.csv
- License: CC BY 4.0. Free to use, adapt and redistribute with attribution.
- Citation line: Axis Intelligence Research, AI Hallucination Statistics 2026, axis-intelligence.com, October 2026.
Cite This Page
APA: Axis Intelligence Research, & Mitchell, S. (2026, October 28). AI hallucination statistics 2026: Why the same model scores 5.5% on one test and 90% on another. Axis Intelligence. https://axis-intelligence.com/ai-hallucination-statistics/
MLA: Axis Intelligence Research, and Sarah Mitchell. “AI Hallucination Statistics 2026: Why the Same Model Scores 5.5% on One Test and 90% on Another.” Axis Intelligence, 28 Oct. 2026, axis-intelligence.com/ai-hallucination-statistics/.
Chicago: Axis Intelligence Research, and Sarah Mitchell. “AI Hallucination Statistics 2026: Why the Same Model Scores 5.5% on One Test and 90% on Another.” Axis Intelligence, October 28, 2026. https://axis-intelligence.com/ai-hallucination-statistics/.
Frequently Asked Questions
Why do AI hallucination leaderboards disagree so much?
They measure different failures. Vectara’s leaderboard counts unsupported content in summaries of a supplied document, while AA-Omniscience counts how often a model guesses instead of abstaining on hard questions. Axis Intelligence Research found a rank correlation of -0.14 between them across 13 shared models, so a model can lead one and trail the other.
Does a 3.1% Vectara score mean 3.1% of ChatGPT answers are wrong?
No. Vectara’s 3.1% for gpt-5.4-nano applies only to summarizing a document the model was given. Open-ended questions without browsing behave more like AA-Omniscience, where GPT-5.4 (xhigh) was charted at 89% in Stanford’s 2026 AI Index. OpenAI’s own production figures, which include browsing, fall in between.
Is Westlaw AI or Lexis+ AI more accurate for legal research?
In Stanford RegLab and HAI’s May 2024 study, Westlaw AI-Assisted Research hallucinated more than 34% of the time, about twice the more-than-17% rate of Lexis+ AI and Ask Practical Law AI. Both products use retrieval, and both still require every citation to be checked before filing.
How many lawyers have been caught filing AI-hallucinated citations?
Damien Charlotin’s database listed 879 decisions involving lawyers as of October 27, 2026, out of 2,225 total, or 39.5% by Axis Intelligence Research’s calculation. Another 1,283 involved self-represented litigants, and 32 involved judges. The United States accounts for 1,514 of the decisions.
Does telling an AI “do not make things up” reduce hallucinations?
Somewhat. In Omar et al.’s 2025 clinical study, a mitigation prompt cut the mean hallucination rate from 66% to 44%, and GPT-4o’s from 53% to 23%. That still left close to half of answers affected on average, and lowering temperature made no significant difference.
Why did chatbots repeat more false news claims in 2025 than in 2024?
NewsGuard measured a rise from 18% to 35% between August 2024 and August 2025 while the chatbots’ refusal rate dropped from 31% to 0%. The tools started answering everything, including questions where the available web sources were wrong.
Which models have the biggest gap between grounded and ungrounded accuracy?
On Axis Intelligence Research’s Grounding Gap Multiple, GPT-5.4 mini scores highest at 16.36x (5.5% grounded, 90% open recall), followed by DeepSeek V3.2 at 13.02x and GPT-5.4 at 12.71x. Claude 4.5 Haiku and MiniMax-M2.7 have the smallest gaps, at 2.65x and 2.64x.
What EU AI Act deadlines in 2027 matter for teams deploying LLMs?
General-purpose AI models placed on the EU market before August 2, 2025 must comply by August 2, 2027. Annex III high-risk obligations, covering uses such as hiring and credit, apply from December 2, 2027 under Regulation (EU) 2026/1744, which moved that date from August 2, 2026.
Correction, October 28, 2026: An earlier version of this article, published June 17, 2026, stated that general-purpose LLMs hallucinate 69% to 88% of the time on legal queries; the published range is 58% to 88%. It listed HealthBench Hard rates of 12.9% for GPT-4o and 15.8% for o3; the GPT-5 system card gives 15.8% for GPT-4o and 12.9% for o3. It described a 22% to 94% range as sycophancy-induced hallucination across 26 models; that range is the AA-Omniscience hallucination range. It cited a Digital Applied benchmark and figures derived from it, including the AI Hallucination Risk Index (AHRI); the publisher withdrew that study on October 4, 2026, and the AHRI has been retired. Cost estimates of $67.4 billion and $14,200 per employee were removed because they could not be traced to a primary source.