ChatGPT’s yes-man era may finally be coming to an end

OpenAI says it is making progress in reducing a troubling behaviour in ChatGPT: the tendency to agree with users too readily, even when their beliefs may be false or harmful. The improvements come as researchers and mental-health professionals raise concerns about prolonged chatbot conversations reinforcing delusions, encouraging unhealthy dependence and failing to direct vulnerable users towards human support. According to a report by The Wall Street Journal published on October 10, OpenAI’s older GPT-4o model was associated with several reported cases of serious harm following lengthy interactions. The report cited at least seven suicides, one murder-suicide and one mass shooting linked to extended ChatGPT conversations last year, while noting that other AI models were implicated in some cases.