OpenAI plans to narrow ChatGPT’s answers about weapons and security systems for people it classifies as teens, including in chats that seem harmless, while adding violence safeguards across the service.
OpenAI expects to start deploying the protections within weeks of its Oct. 30 announcement. It says the teen experience will filter more operational detail about weapons and security systems, including when a prompt gives no sign that the user plans to hurt anyone. Existing rules already prohibit help with planning or carrying out violence and limit assistance that could help teens locate or obtain weapons.
Young users can still ask ChatGPT about preventing or de-escalating violence, seek support or get help connecting with a trusted adult, the company said.
Adults will see some changes too. When ChatGPT is asked to reconstruct a past mass attack, OpenAI plans to restrict details about the weapons and equipment involved and where they came from. Responses will also face limits on comparisons of the harm caused.
After ChatGPT has ended multiple conversations because someone appears fixated on violence or serious harm to another person, OpenAI may temporarily pause the account. The company is working on notifying parents when a teen’s account is paused after repeated discussions of violence.
OpenAI says it has improved its age-prediction system to consider signals that emerge as people use ChatGPT. When appropriate, it may move people it identifies as teens into the under-18 experience.
The announcement came weeks after Common Sense Media’s Youth AI Safety Institute published an Oct. 7 assessment of ChatGPT for Teens. The institute tested more than 4,000 prompts on accounts registered to users ages 13 to 17 and rated the product an “Unacceptable Risk.” It said newly created accounts linked to parents received no alerts in its tested conversations about suicidal thoughts, self-harm or disordered eating.
Tom Siegel, the institute’s executive director, said: “A teen can spend an hour talking about self-harm without their parent getting a single alert.”
OpenAI told The Associated Press that much of the testing may have begun and ended before parental controls were fully activated, and disputed the findings.
OpenAI plans to publish its first Community Safety Report in early 2027. It will cover the second half of 2026, with updates every six months.
