In the past, users had to rein themselves in; now AI grabs the microphone first. OpenAI recently announced that in the coming weeks it will add stricter safeguards against violent content to ChatGPT: when a conversation keeps circling around violence or thoughts of seriously harming others, the model can end the conversation on its own; if the same account is cut off multiple times for such issues, a ban will also be imposed. This is not a rule being created from scratch—ChatGPT already prohibited providing guidance on planning or carrying out violence. This upgrade adds "active cutoff" to enforcement, moving from "passive refusal to answer" to "active interruption," and the platform's logic for handling dangerous content has clearly moved upstream.

Same rules for all ages

The new rules extend restrictions to all users, making no exception for adults. The adjustment focuses on how it answers questions about "historical attack incidents"—from a safety standpoint, the model will reduce or even refuse to provide details, covering the weapons, equipment, and acquisition methods used in attacks, and it will also limit side-by-side comparisons of casualty figures that could amplify the perpetrator's notoriety. In other words, adults and minors face the same tightened boundaries; the difference is only in depth: adult users will also be stopped, but without the extra layer that blocks practical information about weapons and security systems.

Minors get an extra layer

For accounts identified as under 18, OpenAI additionally blocks practical information about weapons and security systems that could be used to commit violence, even if the question does not appear on its surface to have violent intent. Teenagers are still allowed to use ChatGPT to seek help—how to prevent or de-escalate violence, or how to contact a trusted adult, are not among the blocked topics. The accompanying improved age prediction will combine multiple signals, such as user behavior, to continuously assess whether an account may belong to a minor, and automatically apply teen protections when necessary; if a user is misclassified, they can also proactively submit identity information for verification.