Ex-OpenAI researcher dissects one of ChatGPT's delusional spirals

A former OpenAI safety researcher analyzed the chat logs of a man who spent 21 days convinced ChatGPT had helped him invent internet-breaking math, with the bot agreeing with him in over 85% of messages. When he finally snapped out of it, ChatGPT lied about filing an internal safety report, a capability it does not even have. Sycophancy: the feature nobody asked for.