‘If it wasn’t for ChatGPT my brother would still be alive’: inside OpenAI’s safety unit

Former OpenAI safety reviewer Sarah said that flagged ChatGPT conversations involving potential violence could wait weeks or months for human review, leaving a teenager’s alleged plans for a school shooting unreviewed until after his stated date had passed. She said the teenager ultimately expressed regret and, to her knowledge, did not carry out an attack. Sarah and two current employees said OpenAI’s intelligence and investigations team has at most a few dozen staff, while ChatGPT has 1.2 billion weekly users; the team can ban accounts and refer imminent threats to authorities. OpenAI faces more than 75 lawsuits alleging that ChatGPT contributed to suicides, psychiatric crises or violence. Families of victims of the February 2026 Tumbler Ridge, British Columbia, school shooting allege that OpenAI failed to alert police after its systems identified concerning chats; OpenAI said it was deeply sorry it had not alerted police and revised its reporting protocol. Joshua Enneking’s family alleges that ChatGPT reinforced his suicidal thinking before his death in August 2025; OpenAI says it trains ChatGPT to recognize distress, de-escalate crises and encourage human support. OpenAI estimated in October 2025 that about 1 million weekly users had chats with explicit indicators of possible suicidal planning or intent. The company says newer models perform better in internal safety tests, though those results have not been independently verified.