Millions of people are sending messages to ChatGPT each week suggesting emotional dependence or plans for self-harm, the company says. Will an updated model protect them?