AI Models Show Hallucination Rates From 22% to 94% Across Benchmarks

A 2026 Stanford HAI AI Index report measured hallucination rates ranging from 22% to 94% across 26 top AI models, depending on the benchmark and use case. The same study found that models performed better when a false statement was presented as something another person believes, but performance dropped when the statement was presented as something the user believes.

A BBC and European Broadcasting

Union study of AI answers drawn from 22 media organizations found that at least 45% contained at least one significant issue. Within that group, 31% had sourcing problems and 20% included major accuracy issues such as hallucinated details or outdated information.

Dr. Fara Kamangar, founder of DermGPT, estimated that error rates on complex professional queries may fall between 20% and 40%. Aleshia Hayes, a clinical associate professor at Southern Methodist University, said even the lower end of that range warrants concern.

AI errors include misinformation, outdated information, hallucinations, duplication, omitted information, false citations, and mixtures of true and false information. A top Wall Street law firm submitted court filings containing fabricated case citations generated by AI, according to a New York Times report.

Awasthi, an assistant teaching professor at Drexel University, said AI models generate text by predicting statistically probable word sequences based on patterns learned during training. Jan Liphardt, an associate professor at Stanford University and CEO of OpenMind, said addressing inaccuracies involves checking sources, verifying claims through direct observation when possible, and consulting multiple people or multiple AI systems.

Techniques listed in the guidance include lateral reading, pushing back on responses, repeating prompts across different models, and checking for timeliness by asking about changes since the model’s last training cutoff.