ChatGPT can’t tell the difference between beliefs and facts
New study exposes critical flaw of ChatGPT
- A new study by Stanford University revealed that artificial intelligence tools, including major chatbots like ChatGPT, struggle to differentiate between belief and fact.
- The research found that all 24 large language models tested failed to consistently recognise false beliefs and statements, making them prone to hallucination and spreading misinformation.
- This critical flaw has significant implications for the deployment of AI in high-stakes sectors such as law, medicine, and journalism, where distinguishing truth from falsehood is paramount.
- Researchers evaluated the LLMs using 13,000 questions to assess their ability to discern beliefs, knowledge, and facts, exposing a structural weakness in their understanding.
- The study, published in Nature Machine Intelligence, urges tech companies to urgently improve these models before their widespread use in critical domains.