AI Trained on Real Reddit Support Gives Better Mental Health Replies

Lots of people can't get or afford a therapist, so they post about their struggles on forums like Reddit, hoping a stranger who's been through it will respond. The problem: a large share of those posts get no reply at all. A team of researchers asked whether AI could fill that empty space — not by acting like a doctor, but by responding the way a supportive community member would.

To find out, they built a dataset called COPES, made of real peer-support conversations from mental health forums, and a scoring system with three parts: does the AI give the type of help the person asked for, does the emotional tone fit, and does it match what that community considers respectful. Then they tested several AI models — some off the shelf, some specially trained on this data. The specially trained ones did much better. For general-purpose AI models, the match with what the person actually needed improved by over 50 percent, and their emotional tone got closer to the mark too.

But the results came with real warnings. The improvement wasn't evenly spread: the AI did noticeably better in some online communities than others, and better for some kinds of requests than others. More troubling, training shifted the AI toward 'fix-it' mode — offering practical suggestions and solutions — while quietly pushing out the gentler, emotion-focused responses, like simply validating someone's feelings. For a person in distress, that's often the part that helps most.

The takeaway isn't that AI therapists are here. It's that AI can be measurably improved by learning from real communities — and that it still can't be trusted to support everyone equally. If you or someone you know leans on online support, this suggests AI could one day help fill the silence, but only if it learns to listen before it starts solving.