OpenAI has rolled out a suite of safety features aimed specifically at teenage users of ChatGPT, in one of the most significant moves by an AI company to address youth safety concerns.
In a blog post titled 'Teen safety, freedom, and privacy,' CEO Sam Altman laid out three competing principles the company is trying to balance: protecting privacy, preserving user freedom, and safeguarding minors. For teens, Altman wrote, safety takes precedence over the other two.
Key features include: - Age-prediction system: An AI-based system that estimates whether a user is under 18 based on how they interact with ChatGPT. When in doubt, the system defaults to the under-18 experience. - Parental controls: Parents can link their accounts to their teen's account and set content restrictions. - Content restrictions for minors: ChatGPT will refuse to engage in flirtatious conversations with teens, and will not discuss suicide or self-harm even in creative writing contexts. - Crisis intervention: If an under-18 user expresses suicidal ideation, OpenAI will attempt to contact the user's parents and, if unable, will contact authorities. - ID verification: In some cases or countries, users may be asked for ID to confirm age — a privacy trade-off Altman acknowledged but called 'worthy.'
The announcement comes amid growing regulatory pressure on AI companies to protect young users, and follows a wave of lawsuits against social media platforms for allegedly harming children's mental health. Meta is currently facing a federal trial brought by 29 US states seeking up to $1.4 trillion in damages.
Altman emphasized that OpenAI treats conversations with AI as potentially the 'most personally sensitive accounts you'll ever have,' comparing them to privileged conversations with doctors or lawyers. However, he noted that automated systems will still monitor for serious misuse.
The age-prediction system represents a novel technical approach to the challenge of identifying minors online without requiring upfront identity verification, though critics have raised questions about its accuracy and potential for bias.
