ChatGPT’s new violence rules and 200+ police referrals put OpenAI’s safety judgments on the record

OpenAI announced a recurring Community Safety Report on violence prevention and how well its safeguards work, plus new restrictions on violence and weapons content, especially for users it believes are under 18. The first report is due in early 2027, covering the second half of 2026, with updates every six months. The company also disclosed more than 200 law enforcement referrals in the third quarter of 2026. The new safeguards roll out over the next few weeks.

For a product this widely used, rules about when an AI steps in, ends a conversation or alerts a parent have become public policy in all but name. The announcement follows lawsuits accusing OpenAI of aiding the Tumbler Ridge shooting (our coverage) and Common Sense Media calling ChatGPT’s teen mode an unacceptable risk (coverage).

ChatGPT will limit reconstructions of past mass attacks, including weapons used, how they were obtained and comparisons of scale that could feed an attacker’s notoriety. People can still ask how such events happened, how communities were affected and how to spot warning signs. Violence researchers have long worried about copycat dynamics, so an AI that won’t rank atrocities is a reasonable place to land. The limits apply to everyone regardless of age or stated reason, closing the claim-to-be-a-novelist loophole.

ChatGPT will respond more conservatively when recent chats show warning signs of possible violence and can automatically end a chat when someone appears fixated on harming another person. Multiple ended chats may lead to a temporary account pause. For teens, OpenAI restricts practical information about weapons and security systems even without signs of violent intent, and says its age prediction now continually reassesses signals like behavior patterns to move people into the under-18 experience.

Each step is defensible alone. Together they describe a system constantly judging who users are and what they might do, based on signals most people never see. Misclassification will happen. Adults wrongly placed in the teen experience can verify through Persona, OpenAI’s age verification partner, which still means handing identity information to a third party. No comparable appeal is described for someone whose chats are cut off for perceived fixation on violence, which could catch writers, researchers, survivors processing trauma or people venting.

OpenAI says referrals help disrupt violent plots and that mental health and behavioral experts help assess hard cases. If AI can surface a credible threat early, that can save lives. But more than 200 referrals in a quarter is meaningful volume, and the public doesn’t know the thresholds, error rates or outcomes. Those belong in the promised report.

Parental controls already link parent and teen accounts, sending alerts when reviewers identify serious self-harm concerns or disordered-eating behaviors; in July they expanded to notify parents if a teen’s account is deactivated for violence violations. OpenAI is working on notifications for pauses over repeated violent discussion.

This is a responsible direction, and the transparency commitment is the most important part. The first report should include false positives, account pauses, referrals and reversals, not a curated set of anonymized success stories. Without that, the result is a private, opaque safety regime governing hundreds of millions of conversations. OpenAI is also funding outside research on children’s safety (our analysis). More in our OpenAI coverage.