OpenAI updates ChatGPT safety tracking

OpenAI said ChatGPT has been updated to recognize risk signals that build across a conversation, extending its safety checks beyond single-message review in sensitive exchanges. The company said the work focuses on acute cases including suicide, self-harm and harm-to-others, and was informed by mental-health experts. The change was described in recent OpenAI safety materials and was reported Tuesday, May 19, 2026, by EdTech Innovation Hub. OpenAI framed the update as part of a broader effort to make ChatGPT respond more carefully when warning signs appear gradually rather than all at once. ### What exactly changed inside ChatGPT’s safety system? OpenAI said in a post published in May that it updated model policies and training so ChatGPT can “recognize warning signs that emerge over the course of a conversation” and use that context in later replies. The company said the work applies to “acute scenarios including suicide, self-harm, and harm-to-others.” (openai.com) The shift is notable because OpenAI described it as moving beyond isolated message-by-message handling in some sensitive situations. In practice, that means the system is intended to connect earlier signals — such as escalating distress or references that become more explicit over time — before deciding how cautiously to answer. That description is based on OpenAI’s own explanation of the update. (openai.com) ### Who helped shape the update? OpenAI said it worked with more than 170 mental-health experts on related improvements to sensitive-conversation handling. In an October 2025 post, the company said those experts helped train ChatGPT to recognize distress more reliably, de-escalate conversations and guide users toward professional care when appropriate. (openai.com) OpenAI said those earlier changes reduced responses that fell short of its desired behavior by 65% to 80%. The company did not, in the newer context-tracking post surfaced this week, break out a separate performance figure specific only to the across-conversation update. ### Which kinds of conversations are affected? (openai.com) OpenAI said the update is aimed at uncommon but “critically important” cases involving suicide, self-harm and harm-to-others. The company said its goal is to connect relevant signals “without overreacting in ordinary conversations.” OpenAI has separately said ChatGPT is designed to avoid facilitating harmful acts, surface localized crisis resources and encourage people to contact trusted people or mental-health professionals in moments of distress. (openai.com) In recent months, the company has also introduced an optional “Trusted Contact” feature for adults in cases where serious self-harm concerns are detected after automated systems and trained human review. (openai.com) ### How does this fit with OpenAI’s broader safety push? OpenAI has been publishing a series of safety updates tied to mental-health and emotionally sensitive use cases. Those include posts on strengthening ChatGPT’s responses in distress, broader community-safety commitments, and an update on mental-health-related work that said the company was developing evaluation methods simulating extended conversations. (openai.com) EdTech Innovation Hub’s May 19 report presented the new context-tracking change as another step in that sequence. The outlet said the update was meant to improve detection of high-risk signals that may only become clear across a session over time. ### Where did OpenAI publish the details? (openai.com) OpenAI published the most direct description in a post titled “Helping ChatGPT better recognize context in sensitive conversations,” which was available on the company’s website this week. Related background appears in OpenAI posts on sensitive-conversation handling, mental-health-related work and GPT-5 system-card materials. (edtechinnovationhub.com) OpenAI said it will continue sharing more as safety updates roll out in ChatGPT. The company’s recent mental-health-related update said further details would be posted as those changes are deployed. (openai.com 1) (openai.com 2)