OpenAI is narrowing what ChatGPT will say about violence and weapons, and promising to publish regular data on how those safeguards perform. In a post dated October 30, the company announced a package of policy changes that apply to all users, stricter rules for teenagers and a new transparency report, with the changes due to begin rolling out over the next few weeks.
The broadest change concerns past mass attacks. ChatGPT is being directed to give less detail about such events and the people who carried them out, even when that information is available elsewhere online. Users will still be able to ask how an attack affected victims and communities or how to recognize warning signs, but the model is meant to decline reconstructions of the attack, explanations of how weapons and equipment were obtained and comparisons of death tolls that could feed an attacker's notoriety. OpenAI says these limits apply whatever the user's age or stated reason.
The company is also making ChatGPT react to patterns across conversations rather than single prompts. Building on a feature described in May, in which the assistant keeps short factual notes about earlier warning signs, ChatGPT will answer more cautiously when recent chats suggest possible violence. If a conversation shows someone fixating on seriously hurting another person, the system will be able to end the chat on its own, and an account that triggers several such endings may be paused temporarily. Accounts found to have broken the violence rules can still be banned.
Teenagers get a further layer. OpenAI says the under-18 experience will now hold back hands-on details about weapons and about security systems if those details could help someone cause harm, even when a question looks harmless and shows no sign of violent intent, while still allowing teens to seek help, de-escalate conflicts or reach a trusted adult. Those stricter limits will also apply when the company's age prediction system estimates that a user is a minor. OpenAI says that system, launched last year, now keeps reassessing signals during use, and adults who are misclassified can confirm their age through its verification partner, Persona. Parents with linked accounts, who since July have been alerted when a teen's account is deactivated for violence-related violations, are due to receive notices when an account is paused for repeatedly discussing violence.
The post also offered a rare figure on escalation: OpenAI says it referred more than 200 cases involving imminent threats of violence to law enforcement in the third quarter of 2026, and that mental health and behavioral experts help assess difficult cases. To make this work more visible, it committed to a Community Safety Report covering violence prevention and the effectiveness of its safeguards, with anonymized examples and supporting data. The first edition, covering the second half of 2026, is promised for early 2027, followed by updates every six months.
The announcement comes as AI companies face growing scrutiny over how chatbots respond to vulnerable users and to people expressing violent intent, and it moves OpenAI toward the kind of periodic reporting that social media platforms already publish. The substance, however, is still largely a set of commitments. OpenAI gave no data on how often its classifiers misfire, how many accounts have been paused or banned, or what share of the referrals led to action, and it acknowledged that recognizing warning signs reliably remains an open problem.