What Is the AI Hallucination Rate? Real Numbers from 2024-2026

There is no single AI hallucination rate. Published results measure different tasks, model versions, and units of error. A percentage of inaccurate summaries cannot be treated as a percentage of false claims, failed answers, or unsafe decisions.

Source check: September 16, 2026. The figures below retain each source's own publication or update date. Checking a study today does not turn its historical results into a test of today's products.

AI hallucination rates at a glance

Sources: Vectara leaderboard, HalluHard, Tables 3 and 5, Stanford HAI, npj Digital Medicine, and HALoGEN's published paper.

Do not average these percentages. They answer different questions. None establishes the chance that your next chatbot answer will be wrong.

What counts as an AI hallucination?

In source-grounded work, two useful distinctions are contradiction and unsupported addition. If a document says revenue fell and the summary says it rose, the summary contradicts the source. If it invents a revenue figure absent from the document, it adds unsupported information. A claim can even be true elsewhere while still being unsupported by the supplied source.

For open-ended questions, evaluations may instead check facts against external evidence. Some count individual claims; others flag an entire response when any part fails. Refusals and omissions also matter: saying less can produce fewer false claims without producing a more useful answer.

Before using a rate, check the model version, task, scoring unit, evaluator, search access, and treatment of unanswered questions.

ChatGPT, Claude, and Gemini: what can you actually compare?

The Vectara board checked for this refresh still identifies May 11, 2026 as its last update. It evaluates summaries against supplied documents using an automated judge. Selected rows are:

These are named model configurations, not universal accuracy ratings for the ChatGPT, Claude, or Gemini apps. The board also reports answer rate and summary length. Consider those alongside error rate when choosing a model. A short summary that omits useful details is a different product from a thorough one.

The current dataset contains more than 7,700 documents and differs from earlier leaderboard versions. Comparing an old percentage with a new one without checking the dataset can create a misleading improvement or decline. Source and methodology: Vectara.

Does web search reduce hallucinations?

HalluHard tests three-turn conversations about legal cases, research, medical guidelines, and coding. Its reported averages are:

The first three domains score extracted, verifiable claims. Coding scores whole responses. The averages combine these domain scores, so they are not a single per-answer probability.

For Opus 4.5 in the research domain, search reduced reference failures from 38.6% to 7.0% and content-grounding failures from 83.9% to 29.5%. Both improved substantially; neither disappeared. A real source can still fail to support the attached claim. Source: HalluHard, Tables 3 and 5.

The practical check has two parts: open the source, then locate the passage that supports the claim. A working link alone is insufficient.

What the legal and medical studies tell us

Stanford's 2024 evaluation found hallucinations in more than 17% of answers from Lexis+ AI and Ask Practical Law AI, and more than 34% from Westlaw AI-Assisted Research. These are historical results for the tested products and queries, not verified September 2026 rates for their current versions. The study shows why retrieving legal sources does not by itself guarantee a correct answer. Source: Stanford HAI.

The 2025 clinical summarization study examined 12,999 annotated sentences across 18 experimental configurations. It reported hallucinations in 1.47% of sentences, with 44% classified as major because of potential effects on diagnosis or management. These are error-severity judgments, not a measured rate of patient harm. The denominator is sentences, not patients or complete notes. Source: Asgari et al., npj Digital Medicine.

HALoGEN provides another reason to avoid a single headline rate. Its published ACL paper evaluated roughly 150,000 generations from 14 models using 10,923 prompts across nine domains. It reports GPT-4 task scores ranging from 3% to 86%. That spread describes the evaluated tasks and historical model, not today's general-purpose chatbot error rate. Source: Ravichander et al., ACL 2025.

Court cases: a count of incidents, not a hallucination rate

Damien Charlotin's AI Hallucination Cases database listed 2,041 cases when checked on September 16, 2026. The page displayed a last-updated date of September 14, 2026. It tracks court and tribunal decisions involving hallucinated content, including some cases where AI use was alleged but not confirmed. Source: AI Hallucination Cases database.

This is a documented incident collection, not the total number of AI mistakes in legal work. It lacks the denominator of all AI-assisted filings or research tasks, so its count cannot establish a failure percentage. Growth can reflect adoption, detection, reporting, or changes in coverage.

This refresh replaces the earlier July count and removes the previous claim about a record sanction. Comparing penalties requires checking what each order includes, such as fines, fees, or other costs.

Can comparing several AI models catch mistakes?

Comparing answers can surface disagreements and identify claims worth checking. But agreement between models is not independent confirmation: models can repeat the same misconception or rely on the same weak source.

HalluHard measures the effect of web search in the configurations above. It does not establish an accuracy improvement for AskThree or prove that a three-model consensus eliminates hallucinations. No product-specific error rate is claimed here.

Use comparison to guide verification. When models disagree, investigate. When they agree, still verify the claims that determine your decision.

A practical verification workflow

1. Identify the claims that matter. Prioritize names, dates, quantities, quotations, citations, and anything that changes the decision.
2. Request traceable evidence. Ask for the source and the supporting passage, then open it yourself.
3. Check scope and currency. Confirm that the source concerns the right model, product version, jurisdiction, population, or time period.
4. Use disagreement as a review signal. A second answer can expose a gap, but another confident answer does not settle it.
5. Leave unresolved claims unresolved. Remove them, label the uncertainty, or get qualified review before relying on them in consequential work.

For more detail on prompts and grounding, see how to stop AI hallucinations.

Try AskThree to compare AI research answers, then verify the evidence behind the claims that matter.