OpenAI Builds Crisis Router, Taps 250 Doctors for AI Safety

OpenAI is overhauling how its models handle mental health crises, outlining a 120-day sprint to deploy safeguards backed by a newly formalized network of more than 250 physicians. The company isn’t waiting for a major launch event to preview the work; it’s detailing a real-time routing system that will shift users showing signs of acute distress to more capable reasoning models like GPT‑5‑thinking, regardless of which model they initially selected. The logic is straightforward: deliberative alignment training makes these reasoning models better at following safety guidelines and resisting prompts designed to jailbreak them, so they should handle sensitive moments more reliably.

Behind the technical plumbing is a significant investment in human expertise. OpenAI is convening an Expert Council on Well-Being and AI, pulling in specialists in youth development and human-computer interaction to define what well-being even means in the context of a chatbot. That council works alongside a Global Physician Network—a pool of clinicians from 60 countries, including over 90 psychiatrists and pediatricians who are already shaping how models behave in mental health contexts. The company says it’s actively recruiting more specialists in eating disorders and adolescent health, signaling that this isn’t a one-and-done consultation but an ongoing build-out of clinical infrastructure.

The most tangible changes are coming for teens, a demographic the company openly acknowledges are the first “AI natives.” Within the next month, parents will get tools to link their account to a teen’s, set age-appropriate model behavior rules by default, and disable features like memory. More controversially, the system will also push notifications to parents when it detects a moment of acute distress—a feature OpenAI insists will be shaped by expert input to preserve trust, though the tension between safety and surveillance is obvious and unresolved.

It’s a pragmatic admission that chat models are already doing emotional triage, whether anyone planned for it or not. The company is betting that swapping in a more computationally intensive model during a crisis, guided by actual doctors, is better than leaving it to a standard chat endpoint. The unanswered question is whether users will feel that shift happening, and if knowing a reasoning model is suddenly parsing their panic attack feels helpful or deeply strange.