Why AI chatbots ‘hallucinate’ and what researchers are doing about it

People increasingly rely on AI for answers, but it can mislead or spread harmful falsehoods. Here’s how experts are tackling the challenge AI has made it easier than ever to find information: Ask ChatGPT almost anything, and the system swiftly delivers an answer. Publishing confidence scores along with a model’s answers could help people to think more critically about the veracity of information that these tools provide. My lab has also shown that confidence scores can be used to help AI models generate more accurate answers. Most of these approaches assume that the information needed to correctly evaluate an AI’s accuracy can be found on Wikipedia and other online databases.