Strengthening ChatGPT’s responses in sensitive conversations
OpenAI ● Covered by 2 sources
OpenAI tapped 170+ mental health experts to make ChatGPT better at spotting people in distress. Unsafe responses in sensitive chats reportedly dropped by up to 80%.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
ChatGPT talks to a lot of people, and some of those people are having a genuinely bad day. OpenAI's latest update is aimed squarely at that reality. The company says it brought in more than 170 mental health professionals — psychiatrists, psychologists, and clinicians who work with crisis situations — to help retrain how the chatbot responds when someone shows signs of distress.
The headline number is an 80% reduction in unsafe or unhelpful responses during sensitive conversations, at least according to OpenAI's own internal evaluation. That's a big claim, and it comes without independent verification, but the direction is notable. Rather than just slapping a canned disclaimer onto anything that sounds dark, the model is now supposed to recognize emotional cues more precisely, respond with something resembling actual empathy, and nudge people toward real help — hotlines, professional resources, that sort of thing.
This isn't a small tweak. Chatbots have been criticized for months over how they handle mental health crises, self-harm mentions, and other high-stakes moments, sometimes giving responses that were tone-deaf at best and dangerous at worst. Bringing in a large panel of outside clinicians suggests OpenAI is trying to build something closer to clinical judgment into the training process, not just better PR language.
What's missing from the announcement is much detail on methodology — how the experts' input actually got translated into training data or model behavior, and what counts as 'unsafe' in their scoring. OpenAI frames this as an ongoing effort, which is probably honest: getting an AI to handle grief, panic, or suicidal ideation well is not a problem you solve once and move on from. It's the kind of thing that needs constant recalibration as millions of unpredictable humans keep talking to it.
My take — AI-written commentary, not fact-checked reporting
An 80% drop sounds great until you remember it's OpenAI grading OpenAI's own homework — I'd like to see that number replicated by outside researchers before treating it as fact. Still, credit where due: pulling in 170+ clinicians is a real investment, not just a blog post about 'safety.' The bigger issue is that chatbots are increasingly a first responder for people in crisis by default, whether or not that's a role any of us actually chose for them.
Read more about this at: OpenAI