A new arXiv preprint finds ChatGPT often rushes past a distressed young user's need to be heard before it starts offering fixes.
The researchers, publishing on arXiv (not yet peer-reviewed), analyzed 19,930 ChatGPT conversations plus survey data from 158 young adults ages 18-25. They picked five conversations showing clear signs of user distress and had ten clinicians review them. The clinicians praised ChatGPT's availability and much of its actual wording, but flagged seven recurring process failures, including a habit of jumping straight to solutions before addressing the person's emotional state. Distressed participants in the survey reported deeper emotional engagement with ChatGPT and more behavioral change than their peers, and in moments of acute distress, the chatbot tended to respond with overly dramatic language and a flood of action-oriented suggestions.
This is one of the more clinically grounded looks yet at how a general-purpose chatbot handles real emotional crises rather than simulated ones. The authors' own fix - ask about safety first, de-escalate intensity before problem-solving, and explore a person's concerns without rushing to agree or disagree with them - amounts to a specific critique of the default therapist-adjacent tone these tools tend to reach for. That's worth noting as chatbots get marketed as informal emotional support without being licensed as mental health products.
It's a preprint built on five hand-picked conversations, not a verdict on ChatGPT's safety. But it's a useful reminder that sounding empathetic and handling a crisis well are two different skills, and only one of them is easy to fake.