OpenAI proposes six-pillar youth safety blueprint for Australia

OpenAI published a six-pillar Australian Youth Safety Blueprint on September 18 as a proposed framework for safer AI use by young people. It is a policy proposal for policymakers and AI providers—not enacted policy, and not evidence that every measure is already in place.

The blueprint asks Australian policymakers and AI providers to build around six areas: responsible AI adoption in education; privacy-preserving age assurance; under-18 safety policies; protection from manipulative and deceptive outputs; research-informed crisis response; and accessible parental controls. Age assurance means estimating or verifying whether a user belongs in an age group so a service can apply age-appropriate rules.

The education pillar puts AI literacy alongside safeguards. It follows OpenAI and AARP’s separate ChatGPT workshops for older adults, a US program that is not part of the Australian blueprint.

The document also calls for companies to maintain separate policies for users under 18 and test safeguards against manipulation and deceptive outputs. Its proposed crisis measures include documented protocols, default alerts to linked parents when a system detects signs of suicidal intent, referrals to Australian crisis services, external research and expert advisory councils.

Those recommendations go beyond the protections OpenAI says it has already deployed. The company said it began rolling out ChatGPT for Teens in Australia in August as the default experience for accounts identified as belonging to users ages 13 through 17. Australia’s eSafety Commissioner describes the same automatic treatment when ChatGPT predicts an account belongs to someone under 18 or a user states an age in that range.

According to OpenAI, the teen experience includes default interventions around self-harm, violence, eating disorders, dangerous activities and explicit content, along with break reminders and cues identifying ChatGPT as an AI system. The available source material does not provide independent performance data for those safeguards.

OpenAI says its age-prediction system considers factors including account age, active hours, usage patterns and a user’s stated age. It says uncertain cases default to the safer experience, while users placed there incorrectly can seek selfie-based verification through Persona. The blueprint’s privacy language aligns with Australian privacy guidance, which calls for necessary and proportionate checks, lower-intrusion methods where possible, data minimisation, review mechanisms and destruction or de-identification of inputs after use.

OpenAI’s existing parental controls separately let linked parent and teen accounts set quiet hours, disable voice, memory or image generation, and opt out of model training. OpenAI says trained staff review acute-distress alerts before they are sent to a linked parent unless the parent has opted out. The company acknowledges that false alarms can occur and users may bypass guardrails.

On accountability, the blueprint proposes continuing risk assessments, enforceable requirements, independent assessments against common standards and public summaries written in plain language, while withholding details that could help people circumvent safeguards. OpenAI also says it will share findings, support independent research, seek feedback and work with Australian stakeholders. Those are forward-looking commitments, not completed outcomes. The document does not identify who would conduct the proposed assessments or set a timetable for industry-wide adoption.

The proposal comes as Australia consults on an exposure draft of a Digital Duty of Care bill, released September 8. Separate Age-Restricted Material Codes already cover applicable generative AI services; relevant provisions for designated internet services took effect March 9 and require age-assurance measures to comply with privacy law. Consultation on the proposed duty-of-care bill is scheduled to close September 22.