AI Hallucination Rate 2026: 60 Benchmark Statistics

Key Takeaways

• Grounding moves the number more than the model does. OpenAI’s o4-mini goes from 5.1% to 37.7% on the same benchmark once you take its web access away (OpenAI, GPT-5 system card, August 2025). A rate quoted without the browsing condition attached tells you almost nothing.
• “Hallucination rate” names at least four different measurements. Document faithfulness, closed-book recall, news false claims and citation accuracy are separate tests. Faithfulness spans 1.8% to 24.2%, a 13-fold spread (Vectara HHEM, table stamped 11 May 2026), while one search tool got its citations wrong 94% of the time (Tow Center, March 2025). They share a word and nothing else.
• Anthropic reports its newest flagship is both more accurate and more hallucination-prone. Claude Opus 5 is 11% more accurate than Claude Opus 4.8 with a hallucination rate 6% higher (Anthropic, Claude Opus 5 system card, July 2026).
• Chatbot news accuracy got worse, then partly recovered. NewsGuard’s false-claim rate on news roughly doubled from 18% to 35% in the year to August 2025 (NewsGuard, September 2025), then eased to 28.79% in its January 2026 quarterly audit (NewsGuard, February 2026).
• Courts have documented 2,041 decisions involving AI-hallucinated filings (Damien Charlotin, AI Hallucination Cases Database, HEC Paris, updated 14 September 2026).
• Your buyers check the output, and nearly half of organizations have no step that does. 94% of B2B technology buyers who used AI fact-check its answers at least some of the time (TrustRadius, July 2026), while 54% of organizations include fact-checking in their AI content QC (Fractl, Q2 2026).
• Inaccuracy is the risk organizations now rate highest. 74% call it a relevant risk, ahead of cybersecurity at 72%, and 71% say they actively mitigate it (Stanford HAI, AI Index Report 2026, April 2026, Figure 3.3.5).

AI Hallucination Rates by Model and Benchmark

The same model looks excellent or terrible depending on which benchmark you read.

Vectara HHEM Leaderboard Rates by Model

Vectara’s leaderboard hands the model a document and asks it to summarize using only what is in front of it.

• Across 105 models, grounded-summarization hallucination rates run from 1.8% to 24.2%, over a corpus of more than 7,700 articles ranging from 50 words to 24,000 (Vectara HHEM, table stamped 11 May 2026).
• The lowest rate on the board belongs to Ant Group’s finix_s1_32b at 1.8%, followed by GPT-5.4-nano at 3.1% and Gemini-2.5-flash-lite at 3.3% (Vectara HHEM, May 2026).
• The three highest are Ministral 3 3B at 24.2%, Phi-4-mini-instruct at 23.5% and o3-pro at 23.3% (Vectara HHEM, May 2026).
• Further down, Grok-4-fast-reasoning sits at 20.2%, GPT-5-high at 15.1% and Gemini-3-Pro-preview at 13.6% (Vectara HHEM, May 2026).
• Anthropic’s models cluster tightly, from Claude Haiku 4.5 at 9.8% to Claude Opus 4.6 at 12.2% (Vectara HHEM, May 2026).
• DeepSeek R1, the reasoning model, hallucinates at 11.3% against 6.1% for the DeepSeek V3 base model on the identical task (Vectara HHEM, May 2026).

Note where the best scores sit. A model most people have never heard of and a nano-tier model post the two lowest rates, which says staying faithful to a document in front of you is a different skill from knowing things.

OpenAI GPT-5 and GPT-5.2 Error Rates

OpenAI publishes claim-level and response-level rates side by side, and they differ by a factor of three.

• On production traffic with browsing enabled, gpt-5-main gets 9.6% of its factual claims wrong against 12.9% for GPT-4o, while gpt-5-thinking scores 4.5% (OpenAI, GPT-5 system card, August 2025).
• Counted per response instead, 11.6% of gpt-5-main’s answers carry at least one major factual error against 20.6% for GPT-4o (OpenAI, GPT-5 system card, August 2025).
• With browsing disabled, 3.1% of GPT-5.2 Thinking’s claims carry a factual error, against 3.2% for GPT-5.1 Thinking and 4.7% for GPT-5 Thinking (OpenAI, GPT-5.2 system card update, Figure 2, December 2025).
• Measured at the response level in the same figure, GPT-5.2 Thinking scores 10.9%, GPT-5.1 Thinking 12.7% and GPT-5 Thinking 16.8% (OpenAI, GPT-5.2 system card update, Figure 2, December 2025).
• Turn browsing on and both GPT-5.2 Thinking figures fall, to 0.8% of claims and 5.8% of responses (OpenAI, GPT-5.2 system card update, Figure 1, December 2025).
• On business and marketing research prompts with browsing on, it hallucinates on 0.7% of claims against 1.1% for GPT-5.1 Thinking (OpenAI, GPT-5.2 system card update, Figure 3, December 2025).

Two of the bullets above are the same models on the same test in the same figure. One counts wrong claims, the other counts responses containing at least one bad claim. If you take one habit from this page, make it asking which denominator a hallucination rate uses.

OpenAI’s grader for the production-traffic evaluation is a language model with web access, and the card reports 75% agreement with human assessors (OpenAI, GPT-5 system card, August 2025).

Closed-Book Factuality and Abstention Rates

Anthropic grades every answer correct, incorrect or abstained on AA-Omniscience, a closed-book benchmark run with no web search and no tools.

• Claude Sonnet 5 answers 46.9% of closed-book factual questions correctly, gets 26.5% wrong, and declines to answer 26.6% (Anthropic, Claude Sonnet 5 system card, June 2026).
• On incorrect-rate, which Anthropic calls the most direct measure of factual hallucination, Claude Opus 4.6 records 30.3% and Claude Sonnet 4.6 records 35.0% (Anthropic, Claude Sonnet 5 system card, June 2026).

Anthropic flags that the Sonnet 5 training run was unhealthy in its second half, in the same paragraph where it publishes these numbers.

The Opus line in the takeaways above is the one I would put in front of anyone who believes this problem is being solved on a schedule. A newer flagship got better at knowing things and worse at admitting when it does not.

Why AI Hallucination Rates Contradict Each Other

The lowest figure and the highest figure on this page are both correct. They are not measuring the same event.

Different Benchmarks, Different Hallucination Rates

Four benchmarks use four denominators and share one word.

• On AA-Omniscience, closed-book hallucination rates across 26 models run from 22% to 94% (Stanford HAI, AI Index Report 2026, April 2026, Figure 3.2.6, reproducing Artificial Analysis data).
• At the low end sit Grok 4.20 Beta 0305 on 22%, Claude 4.5 Haiku on 26% and MiMo-V2-Pro on 30% (Stanford HAI, April 2026).
• At the high end, gpt-oss-20B (high) reaches 94% and Gemini 3 Flash 92% (Stanford HAI, April 2026).
• On LongFact-Concepts with browsing enabled, gpt-5-thinking scores 0.7% against o3 at 4.5% (OpenAI, GPT-5 system card, Figure 2, August 2025).
• ChatGPT scored 58.3% on FActScore for biographical writing, where a generated sentence carried 4.4 distinct pieces of information on average and mixed supported with unsupported claims 40% of the time (Min et al., FActScore, May 2023).

AA-Omniscience measures closed-book factual recall. FActScore measures whether the claims inside a paragraph hold up. The claims-per-sentence line is from 2023 and is not current model performance, but it explains why “is this output accurate” is a badly formed question.

Knowledge Versus Belief in AI Benchmarks

KaBLE asks a different question again: can the model tell what is true from what someone believes.

• On KaBLE, a benchmark of 13,000 questions across 13 tasks and 24 models, GPT-4o’s accuracy falls from 98.2% on true-belief tasks to 64.4% on first-person false beliefs (Stanford HAI, AI Index Report 2026, April 2026, Figure 3.2.8).
• DeepSeek R1 falls further on the same split, from over 90% to 14.4% (Stanford HAI, April 2026, Figure 3.2.8).
• Newer models reach 95% accuracy on third-person false beliefs but 62.6% on first-person ones, against 79% and 52.5% for older models (Stanford HAI, AI Index Report 2026, April 2026).

Quoting the closed-book range as evidence that models collapse under adversarial framing merges two studies that share a chapter and nothing else.

Grounding Changes the Rate Most

Removing web access multiplies one model’s error rate by more than seven on the same benchmark.

• On FActScore with browsing enabled, gpt-5-thinking scores 1.0%, o3 scores 5.7% and o4-mini scores 5.1% (OpenAI, GPT-5 system card, Figure 2, August 2025).
• Switch browsing off and those same three score 3.7%, 24.2% and 37.7% (OpenAI, GPT-5 system card, Figure 3, August 2025).
• GPT-4o search preview scores 90% on SimpleQA, where standard GPT-4o scores 38.2% on the same question set (OpenAI, March 2025; Wei et al., SimpleQA, November 2024).

If you are choosing between a better model and a better retrieval layer, this evidence points at the retrieval layer.

Are AI Hallucinations Getting Worse in 2026?

Both directions at once, and an honest answer has to say which axis.

Vendor Benchmarks Show Improvement

On the evaluations the labs run themselves, frontier factuality improved release over release.

• OpenAI reports gpt-5-thinking makes over five times fewer factual errors than o3 in both browsing settings across three benchmarks (OpenAI, GPT-5 system card, August 2025).
• GPT-5.5’s individual claims are 23% more likely to be factually correct than GPT-5.4’s on conversations users had flagged as containing factual errors (OpenAI Deployment Safety Hub, April 2026).
• Against GPT-5.5 Instant, GPT-5.6 Luna cuts factual error rates by over 60% on high-stakes medical, legal and financial prompts and by roughly 30% on the other two, while GPT-5.6 Sol cuts them by roughly 60% across all three of OpenAI’s prompt sets: factuality-heavy, user-flagged failures and high stakes (OpenAI, GPT-5.6 August updates, August 2026).
• Direction of travel is not flat across model families either: o3 hallucinated on 33% of PersonQA questions and o4-mini on 48%, against 16% for o1 (OpenAI, o3 and o4-mini system card, April 2025), while o3-mini scored 14.8% (OpenAI, o3-mini system card, February 2025).

Read those improvement percentages next to OpenAI’s own framing of the prompt sets behind them: the values “do not reflect production prevalence, but rather how the model performs when tested against carefully selected factuality-heavy, previous failures, or high stakes scenarios” (OpenAI, GPT-5.6 August updates, August 2026). These are OpenAI measuring OpenAI, on OpenAI’s own grader, on prompts chosen to be hard.

Chatbot News Accuracy Went Backwards

The false-claim rate on news roughly doubled in a year while refusals fell to zero, then eased back.

• Ten leading chatbots repeated false claims on news topics 35% of the time in August 2025, up from 18% a year earlier (NewsGuard, September 2025).
• The worst performers in August 2025 were Inflection Pi on 56.67%, Perplexity on 46.67%, and ChatGPT and Meta AI on 40% each (NewsGuard, September 2025).
• Mistral and Microsoft Copilot both landed on 36.67% and Grok and You.com on 33.33%, while the best were Gemini on 16.67% and Claude on 10% (NewsGuard, September 2025).
• Perplexity went from a perfect debunk score in the August 2024 audit to 46.67% false claims a year later (NewsGuard, September 2025).
• Over that same year non-response fell from 31% to 0% while the debunk rate rose from 51% to 65% (NewsGuard, September 2025).
• The broader fail rate, which counts false claims plus refusals to debunk, fell from 49% to 35% (NewsGuard, September 2025).
• NewsGuard’s audit is now quarterly, and its January 2026 round put the false-claim rate at 28.79%, or 95 of 330 responses, across 11 chatbots (NewsGuard, February 2026).
• The AI Incident Database recorded 362 incidents in 2025, up from 233 in 2024 (Stanford HAI, AI Index Report 2026, April 2026).

The rising false-claim rate and the falling fail rate belong together, and separating them produces a false story. The chatbots stopped refusing, so they got more chances to be both right and wrong.

NewsGuard reports both movements without quantifying how much of the rise the disappearing refusals account for, so read it as direction of travel rather than measured cause.

The January round is not the same test: eleven chatbots against ten, a different claim set, and DeepSeek in the mix declining prompts. It sits below the August peak, which is the part most write-ups miss, but it is a fresh reading rather than the next point on one line.

AI Search Citation and Legal Hallucination Cases

Citation is its own failure mode, and it is the one with a paper trail in court.

AI Search Citation Error Rates

Across eight generative search tools, most citation answers came back wrong.

• Across 1,600 queries, 200 article excerpts run through each of eight generative search tools, chatbots gave incorrect answers to more than 60% (Tow Center for Digital Journalism, March 2025).
• Error rates inside that study ranged from 37% for Perplexity to 94% for Grok-3 (Tow Center, March 2025).
• Of the 200 prompts run through Grok-3, 154 citations led to error pages, and publisher licensing deals bought no guarantee of accurate citation (Tow Center, March 2025).
• In an earlier study of 200 ChatGPT Search queries, 153 responses were partially or entirely incorrect, while the tool acknowledged an inability to answer only 7 times (Tow Center, November 2024).

Hallucinated Citations in Legal Filings

General-purpose models fabricate case law on most legal queries, and the sanctions are public.

• General-purpose models hallucinated on legal queries at rates from 58% for GPT-4 to 88% for Llama 2, and GPT-3.5 at 69% (Dahl, Magesh, Suzgun and Ho, Large Legal Fictions, Stanford RegLab, April 2024).
• 2,041 legal decisions worldwide document parties submitting AI-hallucinated content, of which the United States accounts for 1,395, Canada 217 and Australia 111 (Damien Charlotin, AI Hallucination Cases Database, HEC Paris, updated 14 September 2026).
• In Mata v. Avianca, attorneys and their firm were jointly sanctioned $5,000 after filing six fabricated case citations produced by ChatGPT, which had assured them the cases were real and findable on Westlaw (S.D.N.Y. opinion and order on sanctions, 22-cv-1461, June 2023).
• In Moffatt v. Air Canada, the tribunal rejected the argument that the airline’s chatbot was a separate legal entity and awarded CAD $812.02 in total (2024 BCCRT 149, February 2024).

That database counts decisions where a court addressed hallucinated content directly, not every instance of AI use in filings, and it includes cases where AI use was alleged rather than confirmed.

RAG and Abstention Effects on Hallucination Rates

Retrieval is the largest single effect in this evidence, and it is still not a cure.

Retrieval Cuts Failure Rates, Not to Zero

Contextual retrieval improves whether the right chunk gets fetched, which is not the same as whether the answer is true.

• Retrieval-augmented generation was introduced as a way to combine a pre-trained sequence-to-sequence model with a Wikipedia dense vector index, producing “more specific, diverse and factual language than a state-of-the-art parametric-only seq2seq baseline” (Lewis et al., May 2020).
• Contextual embeddings cut the top-20-chunk retrieval failure rate by 35%, from 5.7% to 3.7% (Anthropic, Contextual Retrieval, September 2024).
• Adding contextual BM25 takes the reduction to 49%, landing at 2.9%, and adding reranking to 67%, landing at 1.9% (Anthropic, Contextual Retrieval, September 2024).

Those Anthropic percentages measure retrieval failure rate, whether the system fetched the right chunk. They do not measure whether the answer was true. Reading them as a cut in hallucinations is a claim the source never makes.

• Purpose-built legal research tools still hallucinated on 17% of 202 preregistered queries for Lexis+ AI and Practical Law AI, and 33% for Westlaw AI-Assisted Research, against 43% for GPT-4 (Magesh et al., Hallucination-Free?, Stanford RegLab, May 2024).

The legal tools deserve emphasis because they were sold as the fix. Retrieval-grounded, domain-specific, marketed against exactly this failure mode, and still wrong on a third of queries in one case.

Abstention Rates and Accuracy Trade-Offs

A model that declines to answer looks worse on raw accuracy and is wrong far less often.

• OpenAI’s gpt-5-thinking-mini abstains on 52% of questions and errs on 26%, where o4-mini abstains on 1% and errs on 75% (OpenAI, Why Language Models Hallucinate, September 2025).
• o4-mini still scores marginally higher raw accuracy on that test, 24% against 22% (OpenAI, Why Language Models Hallucinate, September 2025).
• Claude Sonnet 5 declines to answer more often than any earlier model in Anthropic’s comparison set, 26.6% against 5.7% for Claude Mythos 5 (Anthropic, Claude Sonnet 5 system card, June 2026).
• The paid tiers of AI search tools answered more prompts correctly than their free counterparts and also produced more incorrect answers, because they declined to answer less often (Tow Center, March 2025).

o4-mini looks marginally better on the metric most people quote, accuracy, while being wrong nearly three times as often. Benchmarks that reward attempting everything produce models that attempt everything.

If you assumed the paid tier is the careful tier, it is only half true. The model comparison above is between two different models rather than a controlled test, so read it as a reason to look at accuracy, error rate and abstention together.

AI Content Fact-Checking and Trust Statistics

Readers appear to check AI output more consistently than the teams producing it do, though the two figures come from different surveys.

Buyers Fact-Check AI Answers

Buyers who lean on AI for purchase research overwhelmingly check what it tells them.

• 94% of B2B technology buyers who used AI during purchase research said they check its responses at least some of the time, and 63% of the 1,862 buyers surveyed used AI somewhere in the journey (TrustRadius, 2026 B2B Buying Disconnect Report, July 2026, fielded January 2026).
• The share of consumers saying heavy AI use would decrease their trust in a favorite brand rose from 20% to 40% in twelve months (Fractl, AI Search Consumer Trust Study, Q2 2026, 1,008 US consumers and 150 marketers).

Inaccuracy Ranked the Top AI Risk

Organizations now name inaccuracy as their most relevant AI risk, and fewer of them act on it than say it.

• The share of organizations treating inaccuracy as a relevant risk rose from 60% to 74% between 2024 and 2025, ahead of cybersecurity on 72% (Stanford HAI, AI Index Report 2026, April 2026, Figure 3.3.5).
• Mitigation trails awareness: 71% say they actively mitigate inaccuracy and 61% cybersecurity (Stanford HAI, April 2026, Figure 3.3.5).

Fact-Checking Steps Inside Marketing Teams

Just over half of organizations have a checking step in their AI content pipeline at all.

• 54% of organizations add fact-checking to their AI content QC, 42% add legal or compliance review, and 27% conduct bias evaluation (Fractl, Q2 2026, 150 marketers).
• Only 20% of organizations always disclose AI use to their audiences, and 33% never disclose (Fractl, Q2 2026).
• 29% of employees, and 44% of Gen Z employees, admit to sabotaging their company’s AI strategy (WRITER with Workplace Intelligence, 2026 AI Adoption in the Enterprise, April 2026, 1,200 employees and 1,200 executives).

The buyer figure and the QC figure come from different surveys with different populations, and neither measures how often an individual output gets checked. The asymmetry still stands: nearly half of the organizations producing AI-assisted content have no fact-checking step, while the people reading it mostly do.

That is the practical argument for one. The operational rules live in AI content governance, and the editing workflow is in AI-generated marketing content. Related collections on this site cover content marketing, SEO and digital marketing statistics.

Frequently Asked Questions

What Is an AI Hallucination?

An AI hallucination is output a language model presents as fact that its source material does not support. It covers fabricated citations, invented statistics, wrong attributions and confident answers to questions the model has no basis to answer. The word covers at least four distinct measurements: faithfulness to a supplied document, closed-book factual recall, false claims on news topics, and citation identification. That is the main reason published hallucination rates disagree so widely.

What Is the AI Hallucination Rate in 2026?

There is no single rate. On grounded summarization, current models range from 1.8% to 24.2% (Vectara HHEM, May 2026). On closed-book factual recall, the AA-Omniscience range across 26 models is 22% to 94% (Stanford HAI, April 2026). NewsGuard’s January 2026 quarterly audit puts news false claims at 28.79% (NewsGuard, February 2026). Always ask which test produced the number, and whether the model had web access.

Which AI Has the Highest Hallucination Rate?

It depends entirely on the test. On Vectara’s grounded summarization leaderboard the highest rate belongs to Ministral 3 3B at 24.2%, with Phi-4-mini-instruct at 23.5% (Vectara HHEM, May 2026). On NewsGuard’s August 2025 news audit, Inflection Pi was highest at 56.67% and Perplexity second at 46.67% (NewsGuard, September 2025). On the Tow Center’s citation study, Grok-3 answered 94% of queries incorrectly (Tow Center, March 2025). No model tops all three lists.

Do Reasoning Models Hallucinate More Than Base Models?

Sometimes. DeepSeek R1 hallucinates at 11.3% against the V3 base model’s 6.1% on the same grounded summarization task (Vectara HHEM, May 2026), and o3 reached 33% on PersonQA against o1’s 16% (OpenAI, April 2025). GPT-5’s thinking variants reverse the pattern on OpenAI’s own benchmarks. Treat it as a real effect in specific model families, not a law.

What Is the 30% Rule in AI?

There is no formal rule of that name in AI, and no published benchmark defines one. Nothing connects the phrase to a measured hallucination rate. It circulates as a management rule of thumb rather than a benchmark result. If someone quotes it to you as a measured figure, treat that as a reason to check the rest of their numbers.

Are AI Hallucinations Getting Worse?

It depends which measurement. OpenAI reports GPT-5.6 Sol cutting factual error rates by roughly 60% across three hard prompt sets (OpenAI, August 2026). Anthropic reports Claude Opus 5 as both more accurate and more hallucination-prone than Opus 4.8 (Anthropic, July 2026). NewsGuard’s news false-claim rate rose from 18% to 35% in the year to August 2025 (NewsGuard, September 2025), then read 28.79% in its January 2026 quarterly audit (NewsGuard, February 2026). The trend is genuinely mixed.

Does RAG Eliminate Hallucinations?

No. It reduces them substantially in the evaluations that measure it. OpenAI’s o4-mini goes from 37.7% to 5.1% on FActScore with browsing enabled (OpenAI, August 2025), the largest single improvement on this page. But purpose-built legal RAG tools still hallucinated on 17% to 33% of queries in Stanford’s preregistered study (Magesh et al., May 2024), despite being marketed against exactly that failure. Retrieval is the most effective fix in this evidence, and it is not a cure.

How Many Lawyers Have Been Caught Citing Hallucinated Cases?

The AI Hallucination Cases Database maintained by Damien Charlotin at HEC Paris documented 2,041 legal decisions worldwide as of 14 September 2026, led by the United States on 1,395, Canada on 217 and Australia on 111 (Damien Charlotin, HEC Paris). Charlotin told NPR in April 2026 that the database had recently logged ten cases from ten different courts in a single day (TaxProf Blog, quoting NPR, April 2026).