How often do AI models hallucinate, and can it be stopped?

AI language models still produce fabricated answers, and none of the mitigations tested here removes them entirely. Each AI hallucination study summarised below measures one part of the problem, from error rates on specific tasks to the way users respond once they learn an answer is false. Together they show where errors are most likely and what reduces them.

• Across seven platforms, the average hallucination rate fell from 31.7% in Q1 2024 to 16.0% in Q3 2026.
• Questions containing a false premise produced the highest hallucination rate, at 46.7%.
• Retrieval-augmented generation cut one model's rate on Bangla diabetes questions to 11.3%, but no configuration was free of hallucination.
• Participants told a dialogue contained fabricated information reported lower trust in generative AI and lower dependence on it.

How often do leading AI platforms hallucinate?

A September 2026 test ran 4,182 prompts across seven AI platforms and measured hallucination at the level of whole responses.^† Claude had the lowest rate, at 8.7%, and DeepSeek the highest, at 24.6%.

The same tester reported a long-run decline. The all-platform average hallucination rate fell from 31.7% in Q1 2024 to 16.0% in Q3 2026.

Which questions are most likely to produce errors?

The type of prompt made a large difference to the result. Questions containing a false premise produced the highest hallucination rate, at 46.7%. Summaries of documents supplied by the user produced the lowest, at 4.3%.

Subject matter also mattered. Topics where facts change often, such as local business information, current events and pricing, produced hallucinations at three to four times the rate of stable topics such as historical facts.^†

AI hallucination study results in medicine and hardware

An evaluation of question answering on diabetes in Bangla found wide gaps between models.^† The rates ranged from under 3% to more than three-quarters of answers.

Grounding Qwen3.5-9B in a Bangla diabetes guideline through retrieval-augmented generation lowered its hallucination rate to 11.3%. That is a reduction of 66.3 percentage points. Even so, no configuration in the evaluation was free of hallucination.^†

Separate work examined faults that originate in hardware. Researchers built LLMEI, a tool that injects hardware soft errors into language models during inference. They used it to evaluate the effect of soft errors on hallucination across several model architectures and tasks, including mathematical reasoning and code generation.^†

Does retrieval stop AI models from hallucinating?

Retrieval lowers the rate without bringing it to zero. In the seven-platform testing, web retrieval cut the response hallucination rate from 16.0% to 10.9%. Web retrieval combined with extended reasoning cut it to 8.6%.

A 50-question pilot comparison recorded 49 correct responses out of 50 without retrieval-augmented generation and 50 out of 50 with it. Its authors caution that the dataset is small and simple, so the result does not prove that retrieval eliminates hallucination.^†

Some researchers argue the limit is built into the technology. Kalai et al. (2026) and Banerjee et al. (2025) describe hallucinations as structural. In their view, retrieval, reasoning modules and alignment may reduce how often hallucinations occur but cannot eliminate them.^†

How do users react when told an answer is fabricated?

A between-subjects experiment with 546 participants tested how labels on a generative AI dialogue shape attitudes. One group saw the dialogue labelled as containing fabricated information, and the other saw it labelled as verified accurate.^†

The labels worked as intended. In the manipulation check, perceived error averaged 6.26 in the hallucination-labelled condition and 2.07 in the accuracy-labelled condition.

Participants who saw the fabrication label reported lower trust and lower perceived controllability. They also reported greater cognitive dissonance and frustration, and lower dependence on generative AI.

Literacy shaped only one of the relationships tested. Higher generative AI literacy weakened the positive association between controllability and dependence. The other moderation hypotheses, covering literacy and self-efficacy, were not supported under the adjusted criteria.^†

What falling error rates mean for users who check answers

Error rates are lower than in Q1 2024, and retrieval, extended reasoning and guideline grounding each brought them down further in testing. False-premise questions and fast-changing topics remain the conditions with the highest rates.

Lower error rates bring a separate risk. Li et al. (2026) describe an "accuracy paradox", in which increasingly accurate AI systems may strengthen user confidence while reducing users' willingness to verify outputs.^†

Frequently asked questions