OpenAI said on Friday that it is tightening what ChatGPT can tell users it judges to be teenagers about weapons, the latest step in its effort to keep the chatbot from playing any part in real-world violence. The changes, announced in a company blog post, land amid growing scrutiny of how AI chatbots handle younger users and what they do when a conversation turns toward harm.

What changes for teens

Under the new rules, ChatGPT will hold back practical information about weapons and security systems that could be misused when it is talking to a user it estimates is under 18 — even when the question looks harmless and shows no sign of violent intent. Teens were already refused help finding or obtaining weapons; the new policy extends the restriction to practical detail that could be put to harmful use.

Who counts as a teen is decided by OpenAI's age-prediction system, which reads signals such as behaviour patterns and keeps assessing them over time. Users it estimates are under 18 are moved into the under-18 experience, while adults who are wrongly classified can verify their age to have the restrictions lifted. OpenAI also said it is working on notifying parents if a teen's account is paused for repeatedly discussing violence.

The move builds on ChatGPT for Teens, the version for 13- to 17-year-olds that OpenAI launched in August with extra safety measures and parental controls. OpenAI said earlier this month that teens spend under 15 minutes a day on ChatGPT on average, but a report from the non-profit Common Sense Media around the same time found that the teen version often alerted parents late or not at all when accounts held conversations about suicide, self-harm or eating disorders.

New limits for every user

The changes are not limited to teenagers. For all users, ChatGPT will give less detail about past mass attacks, including the weapons and equipment used, how they were obtained, and comparisons that could amplify a perpetrator's notoriety. OpenAI said the aim is to avoid turning its answers into a reference for future attacks.

ChatGPT will also be able to automatically end a chat in which a user appears to be fixated on seriously harming another person, and OpenAI may temporarily pause accounts after multiple such chats. The company said the safeguards will begin rolling out over the coming weeks and will keep being refined.

A public safety report, and 200-plus referrals

Alongside the new limits, OpenAI promised its first Community Safety Report in early 2027, covering the second half of 2026, with updates every six months. The report will explain how the company detects, investigates and responds to potentially harmful activity, and will include anonymised examples of violence-related risks, changes to safety measures and data on how well they work.

As a preview of that transparency, OpenAI disclosed that it made more than 200 referrals to law enforcement authorities during the third quarter of 2026 over potential threats of violence, saying the referrals had helped disrupt violent plots. The company gave no details of the individual cases or their outcomes.

The announcement follows a week in which OpenAI rewrote its Model Spec, the document setting out how ChatGPT should behave, and comes as regulators move in the same direction: the European Union has been preparing rules that would restrict children under 15 from independently accessing social media and AI chatbots, including ChatGPT. For OpenAI, the message is that teen safety is no longer a side feature of ChatGPT — it is becoming part of the product itself.