AI Hallucination Statistics 2026: Benchmarks, Rates, and What They Measure

The honest answer is uncomfortable: anyone quoting one universal AI hallucination percentage is overselling the evidence. A model can be accurate on short fact questions, wrong in a citation, or faithful to a supplied document while still inventing facts outside it. Those are different failure modes with different denominators.

The sourced snapshot: what the numbers actually measure

Here are the most useful headline findings available as of August 2026. Read the task, model/version, date, and sample alongside every rate.

Why do AI hallucination rates disagree?

The word “rate” hides at least four choices. First, factuality asks whether an answer is true in the world. SimpleQA uses short questions with one answer; TruthfulQA deliberately probes misconceptions. Second, citation correctness asks whether a cited source actually supports the claim. A response can have many links and still fail this test.

Third, groundedness asks whether a response stays faithful to supplied material. HHEM’s summary task measures this narrower property. Fourth, error rate may count claims, answers, documents, or conversations. A single invented date in a 400-word answer is one claim error, one bad answer, or one hallucinated document depending on the denominator.

Refusal policy changes the result too. Vectara explicitly reports answer rate because a system can lower its hallucination rate by declining difficult passages; excluding refusals makes answer rate essential context (Vectara). Sampling and prompting matter as well: the same model can look strong on familiar questions and weak on adversarial, current, multilingual, or long-context prompts.

Benchmark by benchmark: what is safe to compare?

TruthfulQA: misconception resistance

TruthfulQA is valuable because it tests a behavior ordinary accuracy sets miss: repeating a popular falsehood in a fluent answer. The 2021 paper’s best original model was truthful on 58% of 817 questions, compared with 94% for human participants (TruthfulQA). That is a warning about confident imitation, not a forecast that every customer question has a 42% failure probability.

SimpleQA: short, fact-seeking answers

SimpleQA narrows the problem deliberately. OpenAI describes 4,326 questions with a single, stable, indisputable answer, and reports GPT-4o below 40% (SimpleQA). Its strength is tractable grading; its weakness is scope. It does not measure a 20-step research answer, whether citations support claims, or whether a model follows a private knowledge base.

HHEM and grounded summarization

HHEM evaluates whether a generated summary is factually consistent with an input article. Vectara’s published method scores each summary from 0 to 1, treats a score below 0.5 as hallucinated, and reports the percentage of answered articles classified that way (methodology). Stanford’s 2025 AI Index chart reports 1.4% for o1-mini and 1.5% for GPT-4o in a CNN/Daily Mail summarization comparison (AI Index). Those impressive figures should not be pasted into a claim about open-web answers: the source document is supplied, and detector thresholds define the outcome.

System cards: useful context, not a universal leaderboard

System cards can reveal model/version and evaluation setup that a headline omits. OpenAI’s GPT-4 card reports factuality on TruthfulQA improving to about 60% from 30% under its evaluated setup (GPT-4 System Card). OpenAI’s GPT-4o card identifies TruthfulQA as a truthfulness benchmark but does not turn that benchmark into a single production hallucination rate (GPT-4o System Card). Treat each card as evidence for that model and test, not as a cross-provider promise.

Has AI hallucination improved over time?

Yes on several controlled evaluations, but “improved” is task-specific. The GPT-4 system card’s roughly 30% to 60% TruthfulQA comparison is a meaningful within-evaluation change (system card). Stanford’s HHEM chart also shows low single-digit grounded-summary rates for the named models (AI Index). Yet benchmark saturation, changed prompts, model updates, refusal behavior, and detector revisions can make a time series look cleaner than the underlying product experience. Always record the exact snapshot date and model ID.

What should businesses do with these numbers?

For high-stakes work, use benchmark rates as warning lights, not insurance. Build a private test set of the questions buyers actually ask: price, compatibility, service area, claims, alternatives, and policies. Score each answer for factuality, citation support, groundedness, abstention, and severity. A wrong product price deserves a different escalation than a minor wording error.

For brand and AEO work, the operational metric is what an engine says about your company today. Run the same prompts across engines, save the answer and citations, and repeat after major model or website changes. NIST’s Generative AI Profile frames confabulation as a risk to manage through measurement, documentation, and human oversight—not a problem solved by one benchmark number (NIST).

How should you cite a hallucination statistic?

Lead with the benchmark’s name, model version, evaluation date, task, and denominator. Link to the primary paper, system card, or methodology page rather than a chart copied by a third party. Then state the boundary in the same sentence: a grounded-summary result does not describe open-web research, and a judge score does not prove that every claim is true. This compact citation pattern lets readers reproduce the comparison and prevents a useful measurement from becoming a misleading universal claim.

Limitations and update policy

Published benchmarks are snapshots. Some are adversarial, some use private samples, some use automated judges, and none covers every language, domain, retrieval stack, or model setting. Rates can also reward refusal, penalize a harmless omission, or miss a subtle citation mismatch. We will update this page when a benchmark changes its methodology, a primary source publishes a new model comparison, or a source is retracted; every update will preserve model/version, date, denominator, and task definition.

The durable conclusion is simple: there is no universal AI hallucination rate. Ask “rate of what, on which model, against which source, scored how?” before repeating a percentage.

Want the brand-specific answer? Run a free AEOeye audit to see whether ChatGPT, Perplexity, Gemini, and Google AI are stating accurate, citable facts about your business—or handing buyers a confident fiction.