One of the biggest challenges facing large language models (LLMs) is hallucination: generating information that is false but presented as fact. Moreover, these fabrications tend to be plausible and convincing. For example, in a legal context, LLM-generated court documents regularly hallucinate precedent court cases. Many submissions by lawyers have been caught, and law firms and lawyers sanctioned by judges, but more and more continue to occur. One database of legal AI hallucination cases lists well over a thousand examples. LLMs make up quotes and plausible sounding facts when they don’t actually know the answers. Because LLMs generate the statistically most likely continuation of text rather than retrieving verified facts, they can produce answers that sound convincing despite being incorrect. Large language models are prediction engines. They generate the most likely next word rather than retrieving verified facts from a database. Even with perfect training data, there will always be situations where they have to guess when they see ambiguous questions, novel situations not seen during their training, conflicting sources, missing information, or requests that have no correct answer. Academic research suggests that AI hallucinations can never be eliminated.
How have hallucination rates changed over time? One problem in answering that is that there is no single agreed benchmark for all hallucinations. Hallucination rates vary between simple factual lookups and more difficult reasoning cases. The Vectara HHEM hallucination leaderboard has been running since 2023 and ranks leading models by their hallucination rates, based on errors made when summarising documents. In July 2026 this was showing rates of 7% for GPT 5.4, 5% for Mistral and 3.3% for Gemini 2.5 Flash. However, this index is measuring the simplest possible task for an LLM: simply recalling something where the answers are already present in the supplied document. For these basic tests, LLM fabrication rates have improved over time. In 2024 the best models were scoring roughly 1-3% on the original, simple HHEM dataset. By early 2025 the best model (Gemini-2.0-Flash) had fallen to a 0.7% hallucination rate on that same benchmark. So, for simple lookups, LLMs are getting better. However, the picture is different as the complexity of questions increases, as we shall see.
A more challenging test is the AA-Omniscience benchmark. This uses 6,000 questions across different domains and topics, spanning Business, Humanities & Social Sciences, Health, Law, Software Engineering, as well as Science, Engineering & Mathematics. Sub topics range from Python data libraries to public policy and taxation. Questions are derived from authoritative academic and industry sources. They are generated to ensure the questions are unambiguous, scalable, and factually precise. The questions are deliberately difficult enough that only an expert in the field would know the answer, so the benchmark doesn’t saturate with easy questions. The latest results of this index are quoted in the influential Stanford AI Index 2026 report. This shows that the hallucination rates for these more complicated questions are much higher. Across 25 LLMs tested, Grok 4.2 had the best hallucination rate, and that was still 22%. ChatGPT 5.4 had a rate of 89%, Claude Opus max had 61%, and a Gemini 3 Flash had 92%. On specific factual-recall tests like OpenAI’s own PersonQA, newer “reasoning” models haven’t improved and indeed are getting worse: OpenAI’s o3 hallucinated 33% of the time versus the older o1’s 16%.
These findings are quite worrying. For real processes in industry, error rates of even 1% are generally unacceptable. You do not want 1% of your deliveries going to a fabricated address, 1% of your car engine parts being misplaced or 1% of your bank transfers going to an imaginary account. Nobody would accept 1% of airline reservations disappearing or 1% of pharmaceutical prescriptions containing invented drugs. In the Six Sigma approach used in industry for many industrial processes, the target error rate is 3.4 errors per million, vastly below 1%. Yet we can see that LLMs hallucinate at more than this even for the simplest questions. For more specialist queries, hallucination rates are at least 22% and up to 92%.
It is troubling that this reliability issue does not seem to be universally understood by company executives or politicians, who are enthusiastically pushing AI use into enterprises and government bodies. Stanford research and other studies show hallucination rates to be far higher in specialized domains than in general chat use. ChatGPT has been documented inventing non-existent cancer treatments in a meaningful share of tested medical cases. This is why the ECRI (a health-technology safety non-profit) named AI chatbot misuse in healthcare its top health-technology hazard for 2026, given that tens of millions of people now consult AI for health information daily.
On December 9, 2025, The US Financial Industry Regulatory Authority (FINRA) published its 2026 Annual Regulatory Oversight Report. For the first time in the report’s history, it devoted a dedicated section specifically to AI risks, including hallucinations. The report makes clear that a hallucinated fact in an AI-drafted client communication is a compliance violation in just the same way a human misstatement would be.
It is good that regulators such as these are beginning to wake up to the risks associated with hallucinated AI content. There is no doubt that LLMs have many genuine use cases, from image generation to translation to coding software. But as AI spreads further and further into everyday life, it is vital that people understand the risks as well as the opportunities. We are used to relying on computer systems for accurate, consistent answers. LLMs are fundamentally unable to provide either consistency or complete reliability.
