ChatGPT is used by so many people that small changes to how it handles dangerous questions reach a very large audience, which is why OpenAI's latest safety tightening is worth reading closely. The company says ChatGPT will give less detail about past mass attacks, grow more cautious when a recent chat shows warning signs, end conversations that fixate on violence, and restrict practical weapons information for anyone it judges to be a teenager.
The concrete changes are specific. ChatGPT is being directed to provide less detailed information about past mass attacks and the people who carried them out, even where that detail is available elsewhere online, with the stated aim of not helping reconstruct an attack or amplifying an attacker's notoriety. These limits apply regardless of a user's age. Separately, the model will respond more conservatively when recent chats show warning signs of possible violence, and if a conversation shows someone fixating on seriously harming another person, ChatGPT can automatically end the chat; repeat it and the account may be paused. OpenAI says it made over 200 law enforcement referrals in the third quarter of 2026.
For teenagers the restrictions go further. OpenAI already bars helping anyone plan violence or find weapons, and it is now extending that to practical information about weapons and security systems that could be misused, even with no sign of violent intent. Teens can still use ChatGPT to seek support or help de-escalate a situation. The most useful commitment is a twice-yearly Community Safety Report, the first due in early 2027 covering the second half of 2026, which is the part to hold OpenAI to, because referral counts without context are hard to judge on their own.
The mechanism behind the teen rules is where the trade-offs live. These protections apply not only to known under-18 accounts but to users OpenAI predicts are under 18 from behaviour signals, and prediction is not exact. Some adults will be placed in the teen experience by mistake, and the route back is an identity check through Persona, OpenAI's age-verification partner. That fixes the error but asks the user to hand over identity data, which not everyone will accept, so the convenience of an unverified account and the stricter teen limits are now in tension.
The honest open question is effectiveness. OpenAI gives no figures on how often its detection catches real threats or how many warnings turn out to be false, so the referral count is a number without a denominator. Tighter rules on mass-attack detail and teen weapons information are reasonable steps, but whether they measurably reduce harm, and how often they wrongly restrict legitimate questions, will only show once the promised safety report starts putting data behind the policy.