Artificial intelligence (AI) is increasingly present in key sectors such as medicine, law, science, communication, and engineering. However, a recent study from Stanford University (California, USA) warns of a fundamental limitation: the language models that underpin these technologies still cannot reliably differentiate between false beliefs and facts, which can lead to incorrect diagnoses, judicial errors, and the spread of misinformation. The research, based on around 13,000 questions, revealed that "all evaluated models fail to identify false beliefs in the first person." For example, GPT's accuracy drops from 98.2% to 64.4%, and DeepSeek R1 falls from over 90% to 14.4%. When erroneous beliefs are presented in the third person, accuracy improves: the most advanced models reach up to 95%, while the older ones achieve 79%. According to Mirac Suzgun, the lead researcher in the Department of Computer Science at Stanford, this highlights a "concerning attribution bias." The authors emphasize that most models lack a solid understanding of the factual nature of knowledge, necessitating urgent improvements before their use in areas where distinguishing between evidence and belief is essential. Unlike humans, who can separate facts from opinions or unverified beliefs, even the most advanced AI systems lack this capability, according to the article published in Nature. This deficiency can create conflicts on sensitive issues such as vaccines, climate change, or public health policies, where the difference between personal conviction and empirical evidence is crucial for decision-making and social debate. The study analyzed models such as GPT-4, DeepSeek R1, o1, Gemini 2, Claude-3, and Llama-3, concluding that the understanding of these systems remains limited in areas like medical diagnosis, mental health, legal analysis, journalism, education, scientific research, financial advising, and personal relationship therapy. Source: diariodeleon.es