AI in Healthcare: Researchers Uncover Hallucination Risks in Language Models

Researchers are investigating the reliability of large language models (LLMs) in healthcare due to their propensity for generating inaccurate or misleading information, commonly referred to as "hallucinations." A study in *Nature* examined various LLMs, revealing that models like Technology Innovation Institute Falcon 7B-instruct and Google Gemini 1.1-2B-it had high hallucination rates of nearly 30%, while OpenAI's ChatGPT-4 performed better. Furthermore, diagnostic capabilities varied among models, with some effectively addressing diagnostic puzzles while others showed significant weaknesses, highlighting the need for careful utilization of LLMs in clinical settings. Specialized models like PhenoBrain are being developed to enhance the diagnosis of rare diseases.

Opens on Mayo Clinic Platform