OpenAI says ChatGPT just got a lot better at recognizing when people are in distress. The company teamed up with more than 170 mental health professionals to train the AI to respond more compassionately, de-escalate crisis moments, and guide users toward real-world help. In testing, these new updates cut unsafe or insensitive replies by as much as 80 percent.
It sounds like a win for safety, but it also raises a deeper question: is ChatGPT slowly becoming a kind of digital therapist?
OpenAI’s newest model, GPT-5, has been taught to spot early signs of self-harm, psychosis, mania, and emotional dependence. When someone says something like “I feel like people are targeting me,” the AI now offers calm reassurance, grounding exercises, and gentle nudges toward mental health professionals or trusted friends. The company even added new prompts encouraging users to take breaks and connect with real people during long chat sessions.
That may sound comforting, but it also blurs a line. As millions of people turn to ChatGPT for advice, comfort, and even companionship, OpenAI is now deciding what empathy looks like in a machine. And that’s both fascinating and unsettling.
The improvements are rooted in detailed behavioral guides known as “taxonomies.” These describe what healthy and unhealthy model responses should look like in sensitive mental health conversations. By refining those definitions and testing them with psychiatrists and psychologists, OpenAI says GPT-5 now responds appropriately over 90 percent of the time, a major jump from earlier versions.
Still, the company admits that mental health-related chats are rare, accounting for just fractions of a percent of total usage. But those tiny percentages still represent millions of messages. For self-harm or suicidal topics, clinicians found GPT-5 reduced unsafe answers by 52 percent compared to GPT-4o. For emotional reliance, when users start treating ChatGPT like a substitute for human relationships, unwanted responses dropped by 80 percent.
That last metric is especially interesting. OpenAI has essentially built a filter to detect when users might be getting too attached to the chatbot. It now replies with subtle reminders that AI cannot replace human connection. “I’m here to add to the good things people give you, not replace them,” one example reads. It’s a strange paradox, a machine gently warning you not to love it too much.
OpenAI insists these efforts are not about replacing therapy but about keeping users safe. It has expanded access to crisis hotlines and rerouted certain sensitive messages to safer versions of ChatGPT. Still, some might see this as a first step toward AI-mediated counseling, an area that could quickly become controversial. If people already confide in chatbots more openly than in humans, where does the line between assistance and therapy begin?
For now, OpenAI’s approach seems grounded in responsibility. The company says it is continuing to work with global experts to refine its safety standards and will add emotional reliance and non-suicidal mental health emergencies as part of its regular model testing going forward.
Still, the broader issue lingers. As ChatGPT learns empathy, it is not just getting better at conversation, it is quietly taking on a role that once belonged only to humans. Whether that makes it more helpful or more dangerous may depend on how much people start to depend on it.
