LLMs struggle to distinguish between facts and beliefs

Stanford researchers tested 24 popular large language models, including GPT-4o, on 13,000 questions and found they struggle to distinguish between facts and personal beliefs. The models were 34-39% less likely to identify false beliefs compared to true beliefs, though they performed much better at recognizing factual statements with over 90% accuracy. These limitations raise concerns about deploying AI systems in critical fields like medicine and law.