The news

Primaries:

OpenAI frames the Blueprint as a roadmap for protecting young people who use AI and as a practical contribution to Australian policy. The company page says the Blueprint's six pillars cover AI literacy and age-appropriate safeguards, privacy-protective age assurance, connections to real-world crisis support, and accessible parental controls.

Product note on the company page and PDF: in August OpenAI began rolling out ChatGPT for Teens in Australia, a default experience for users identified as aged 13 to 17, building on parental controls, under-18 safety policies, and age assurance.

Four commitments (OpenAI's words)

The PDF lists four commitments that guide OpenAI's teen work:

  1. Put teen safety first (age-appropriate safeguards in ChatGPT).
  2. Encourage access to real-world support and trusted relationships.
  3. Treat teens like teens (developmental stage without adult defaults).
  4. Be transparent about how systems work and should behave.

Referenced product/policy mechanisms: parental controls; Under-18 Principles in the Model Spec; age prediction; ChatGPT for Teens rollout in Australia.

Six pillars policymakers should consider

Summarized from the PDF (OpenAI's recommendations):

  1. Recognise positive uses and encourage responsible adoption in education. Treat AI literacy as opportunity and safety strategy. Teacher-led school adoption. Notes state/territory education responsibility and Productivity Commission interest in a national ed-tech / AI approach. Claims nearly 9 in 10 teens who use ChatGPT turn to it for learning, information, skill-building, or productivity in a given week (OpenAI figure; independent public replication: UNKNOWN).

  2. Privacy-preserving age assurance. Distinguish teens from adults with risk-based age estimation that minimizes sensitive data; prefer OS/app-store signals where possible. OpenAI says Australian under-18 identification uses its global age-prediction approach; low-confidence cases default to a safer experience. Users can confirm age via Persona (selfie); OpenAI says it does not see the selfie/ID and Persona deletes verification data within seven days, citing consistency with the Privacy Act 1988.

  3. Under-18 safety policies. Age-appropriate defaults; pre-deployment testing and post-deployment monitoring; protocols for self-harm, exploitation, grooming, and other high-risk interactions; publish robust child-safety policies.

  4. Protect teens from manipulative and deceptive outputs. Mitigate unhealthy emotional dependency and inappropriate relationship dynamics; avoid outputs that encourage overreliance, compulsive engagement, secrecy from parents, or confusion about machine vs human; avoid initiating/reinforcing anthropomorphic behaviours that imply consciousness, romantic partnership, or human authority; wellbeing interventions (breaks, helplines, offline support).

  5. Crisis response protocol. Documented protocols and embedded features. PDF examples include: notify linked parents by default if a teen expresses suicidal intent; refer to Australian resources such as Triple Zero (000), Lifeline (13 11 14), Kids Helpline (1800 55 1800), or 13YARN; support crisis organisations; long-session break reminders; external research; expert advisory councils (OpenAI cites its Expert Council on Well-Being and AI).

  6. Accessible parental controls. Defaults first, then parent/educator layers. PDF says controls should allow: link parent and teen accounts (teen 13+) via email invite; manage memory/chat-history/location privacy settings; auto-alerts on self-harm intent; quiet hours and study hours; notice when a child modifies or disables a parent-configured safety setting.

Legislative frame OpenAI wants

The PDF argues for a risk-based, outcomes-focused legislative framework rather than fixed technical prescriptions. It situates Australia's existing eSafety Codes and Standards (class 1 and class 2 material, including certain AI-generated content) and notes the government is developing a Digital Duty of Care framework via the draft Online Safety Amendment (Digital Duty of Care) Bill 2026, which would replace Codes and Standards if enacted. OpenAI says it supports requiring providers to identify and address design/operation risks, with proportionate obligations tailored to service type, and that it is consulting with the Australian Government.

The PDF also asks for practical independent assessments of teen-safety safeguard implementation, common audit standards, and plain-language public summaries (with protection for misuse-sensitive detail).

Who is bound

  • OpenAI: bound by the product commitments it ships (ChatGPT for Teens, parental controls, age prediction, crisis features) under its own terms and applicable Australian law. The Blueprint itself is advocacy.
  • Australian Parliament / eSafety / government: own the Digital Duty of Care bill process. Enactment status and final text: UNKNOWN / follow the official exposure draft and consultation. Treat OpenAI's summary as industry comment only.
  • Schools and state/territory education authorities: decide classroom AI policy; OpenAI does not control curriculum.
  • Parents/guardians: optional control linkage under OpenAI's described product.

What's new

  1. A jurisdiction-specific OpenAI Blueprint PDF aimed at Australian youth-AI policy, dated September 2026, announced 18 September.
  2. Explicit mapping of six pillars onto Digital Duty of Care / eSafety reform talks.
  3. Operational detail on Australian age-assurance fallback (Persona selfie, seven-day deletion claim) and Australian crisis numbers in the crisis pillar.
  4. Confirmation that ChatGPT for Teens rollout in Australia started in August for ages 13-17.

What it does not settle

  • Whether Digital Duty of Care passes, and in what form. UNKNOWN.
  • Independent audit results of ChatGPT for Teens safeguards in Australia. OpenAI proposes such assessments; published third-party results: UNKNOWN.
  • Accuracy of the "nearly 9 in 10 teens" weekly learning/productivity usage claim. Source methodology beyond the PDF: UNKNOWN.
  • How age prediction error rates compare to document-based age checks in practice. UNKNOWN from this package.
  • Cross-border applicability (other countries copying the Blueprint). Not claimed as law.

What to do now

  • If you operate in Australia with under-18 AI users: read the PDF pillars as OpenAI's bargaining position and map them against your own age-assurance, crisis, and parental-control stack.
  • If you track Online Safety Amendment (Digital Duty of Care) Bill 2026: pull the official exposure draft and consultation materials; treat OpenAI's summary as industry comment.
  • If you run schools or ed-tech procurement: separate teacher-led classroom policy from consumer ChatGPT for Teens defaults; confirm data-retention and parent-link behaviour in writing.
  • If you evaluate teen safety claims: demand measurable false-positive/false-negative rates for age prediction and crisis detection; those rates are UNKNOWN here.
  • Watch for OpenAI publishing independent assessment summaries of the kind the Blueprint recommends.