Artificial intelligence tools such as ChatGPT are struggling to differentiate between belief and fact, according to a new study by researchers at Stanford University, raising concerns over their reliability in high-stakes fields like law, medicine, and journalism.
The study found that all major AI chatbots — including ChatGPT, Claude, DeepSeek, and Gemini — failed to consistently identify when a belief was false, increasing their likelihood of producing hallucinations or spreading misinformation.
Read More: Experts warn of rising ‘AI psychosis’
“As language models increasingly infiltrate high-stakes domains such as law, medicine, journalism and science, their ability to distinguish belief from knowledge, and fact from fiction, becomes imperative,” the researchers said. “Failure to make such distinctions can mislead diagnoses, distort judicial judgments, and amplify misinformation.”
The research team evaluated 24 large language models (LLMs) using 13,000 questions designed to test how well they distinguished between knowledge, belief, and fact. All models tested demonstrated weaknesses in recognising false beliefs and connecting information to objective truth.
Read More: Hey chatbot, is this true? AI ‘factchecks’ Pakistan-India war information
Pablo Haya Coll, a researcher at the Computer Linguistics Laboratory of the Autonomous University of Madrid, who was not involved in the study, said the findings reveal a “structural weakness” in AI systems. “Their difficulties in robustly distinguishing between subjective conviction and objective truth have critical implications in fields where this distinction is essential,” he said.
Coll suggested that training models to be more cautious in their responses could help reduce hallucinations — though it might also limit their overall usefulness.