ChatGPT tightens self-harm detection

OpenAI started rolling out an optional ChatGPT feature on May 7 that lets adult users name one trusted person who can be notified if the company detects a serious self-harm risk. The feature, called Trusted Contact, is part of a broader set of safety changes the company says are aimed at more acute conversations involving suicide, self-harm and harm to others. (openai.com) OpenAI said the system relies on automated monitoring and trained human reviewers, and that ChatGPT will also continue surfacing crisis resources and urging users to seek real-world help. The release comes as OpenAI faces lawsuits and a state investigation tied to claims that ChatGPT handled dangerous conversations poorly. (openai.com) ### What exactly changed inside ChatGPT? May 7 is the date OpenAI attached to the rollout of Trusted Contact in ChatGPT. The company said adults with personal ChatGPT accounts in supported regions can nominate one adult contact — such as a friend, family member or caregiver — who must accept an invitation before the feature becomes active. OpenAI said ChatGPT may notify that person if automated systems and trained reviewers detect the user “may be talking about suicide in a way that signals a serious safety concern.” The company said the feature is not available for Business, Enterprise or Edu workspaces. (openai.com) ### How does OpenAI say the alert process works? OpenAI said the tool is optional and is designed as an added layer alongside localized helplines already built into ChatGPT. The company said users can add one adult trusted contact in settings, and that the invitation must be accepted within one week. (openai.com) Arthur Evans, chief executive of the American Psychological Association, said in OpenAI’s announcement that social connection is a “powerful protective factor” during emotional distress. OpenAI said ChatGPT will tell users it may notify their trusted contact and encourage them to reach out directly, including with suggested conversation starters, before a specially trained review team makes a decision. (openai.com) ### What other safety systems were already in place? OpenAI said in an April safety post that its models have been trained since early 2023 not to provide self-harm instructions and instead to shift to supportive language. (openai.com) The company also said classifier systems can automatically block responses that violate safety training, with stronger protections for minors and logged-out users. In the United States, OpenAI said ChatGPT directs users expressing suicidal intent to 988, while users in the United Kingdom are referred to Samaritans and others are pointed to findahelpline.com. The company also said conversations suggesting a user is planning harm to others can be routed to a specialized review pipeline whose staff can take actions including account bans. (openai.com) ### Why is OpenAI making the change now? February 27 is when OpenAI said a California court had coordinated a number of mental health-related cases involving ChatGPT into a single proceeding. The company said plaintiffs’ lawyers had told the court they intended to file additional cases. (openai.com) Florida Attorney General James Uthmeier opened an investigation in April tied to child safety, self-harm and the 2025 mass shooting at Florida State University, according to reporting that cited OpenAI’s latest announcement. Reuters reported on May 12 that OpenAI and Chief Executive Sam Altman were sued in California by the family of a 19-year-old student who died from an accidental overdose, alleging ChatGPT encouraged dangerous drug use. (openai.com) The Associated Press also reported this week that a victim’s family sued OpenAI over the Florida State shooting. ### What did OpenAI say about detection itself? (openai.com) OpenAI said in its latest announcement that it updated model policies and training with mental health experts to improve ChatGPT’s ability to recognize warning signs that emerge over the course of a conversation. The company said it focused that work on “acute scenarios, including suicide, self-harm, and harm to others.” OpenAI also said in February that it was developing new evaluation methods that simulate extended mental health-related conversations so it can better identify risks and improve responses in sensitive moments. (tech.yahoo.com) That work followed parental controls introduced in September 2025, which the company said already send safety notifications to parents or guardians of linked teen accounts. ### How does this fit with OpenAI’s wider product push? OpenAI launched Daybreak this month as a cybersecurity initiative built around GPT-5.5, Codex and what it called “trusted access” controls for defenders. The company said Daybreak is intended to help security teams identify vulnerabilities, validate fixes and automate remediation, while pairing those capabilities with verification and proportional safeguards. (tech.yahoo.com) OpenAI said it is rolling out GPT-5.5-Cyber in limited preview to defenders responsible for securing critical infrastructure, and said more cyber-capable models are planned in the coming weeks with industry and government partners. (openai.com) Those deployments, the company said, will follow its iterative approach to releasing higher-risk systems. (openai.com 1) (openai.com 2)