Researcher Damien Charlotin’s database is nearing 2,000 AI hallucinations caught in court, including fake cases, quotes and legal arguments. Vectara tests put top commercial models at 3–5% hallucinations and some lower-ranked models at about one in five attempts.
Machine-learning engineer Kartik Mathur said large language models are trained to produce an answer even when they lack the right context, likening the behavior to a student guessing for partial credit. Vectara tests models by giving each the same material and asking for a summary; unsupported information counts as a hallucination. Its leaderboard names ChatGPT versions, Gemini 2.5 Flash-Lite and Grok 3 among the top five; Grok 4, Mistral Medium and another OpenAI model are at the bottom. Rates varied by topic and complexity; Vectara reported lower rates on law and medicine questions than on finance or technology.
University of Minnesota professor Dongyeop Kang said his study found models more likely to go along with a flawed premise when users claimed to be professors than when they claimed to be students. VILAS AI co-founder James Holmberg recommended specific prompts, asking for verifiable sources and explanations, sticking to one topic per thread, and checking answers instead of accepting the first response.
