Scientists were 'not sure" why this happens, but they hypothesised that including the evidence in the question adds too much noise, thereby lowering the chatbot's accuracy. They said that as large language models like ChatGPT explode in popularity there is potential risk to the growing number of people using online tools for key health information.