OpenAI recently officially announced that over the coming weeks it will fully upgrade and strengthen ChatGPT's restrictions on violence-related content, while also introducing stricter safeguards for minor users regarding weapons and security system information.

The core of this policy adjustment is to take more decisive intervention measures for high-risk conversations. According to the newly announced measures, if the system detects and assesses that a user's conversation shows a sustained focus on violence or ideas of seriously harming others, ChatGPT will have the ability to proactively end the conversation. More severe still, if a user has had conversations terminated by the system multiple times for such issues, their account may face suspension or even an outright ban.

In terms of specific content-filtering boundaries, the existing policy continues, as always, to prohibit providing any operational guidance that helps users plan or carry out violent acts. In this adjustment, however, the scope of restrictions has been further expanded to the historical dimension—the model's response details to questions about "past mass attacks and their perpetrators" will be substantially reduced, or it may even choose to refuse to answer outright. The restricted content specifically includes not only the weapons, equipment, and ways to obtain them involved in recreating an attack, but also limits on users' queries seeking detailed comparisons of the scale of harm intended to amplify an attacker's notoriety. It should be noted that this series of restrictions applies to all users, regardless of age group or purpose of the question; even adult users will be subject to the same constraints when making related queries.

In addition to general restrictions for all users, OpenAI has further stepped up targeted protections for the group identified as under 18 years old. The system will specifically restrict practical information about weapons and security systems that could be used to carry out violence; this protection will be automatically triggered even if the related question does not on its surface directly reveal violent intent. However, OpenAI emphasizes that teenagers can still normally seek help through ChatGPT, such as learning how to prevent or de-escalate violence, or obtaining effective guidance on how to contact a trusted adult.

To ensure that the teen-protection mechanism is accurately implemented, OpenAI has introduced an improved age-prediction mechanism. This mechanism can combine multidimensional signals such as users' daily behavior to continuously conduct dynamic assessments of whether an account belongs to a minor, and automatically apply the corresponding protective measures at the appropriate time. If a user's account is unfortunately misjudged by the system as belonging to a minor, it can also be verified and adjusted by proactively submitting identity information.