OpenAI proposes an international youth AI safety institute ahead of the G7 in Évian

The ask is institutional, not just aspirational

OpenAI is using the G7 Leaders' Summit in Évian, France, to call for a standing body dedicated to youth AI safety. The framing is deliberate: a summit produces a moment, but the company argues that youth safety needs continuity that outlasts a single gathering.

Notably, OpenAI does not insist on a brand-new organization. It offers alternatives — a new international institute, or giving an existing or newly established national AI institute a global mandate to share research, evidence, and guidance. The company explicitly says what matters is the function, not the form.

That flexibility lowers the bar for governments to say yes. Rather than negotiating a new treaty organization from scratch, a G7 government could extend the remit of an institution it already funds. It also lets OpenAI point to existing scaffolding it already touches: Common Sense Media's Youth AI Safety Institute, which the OpenAI Foundation supports, and its collaboration with the American Federation of Teachers.

Age estimation is the foundation the rest depends on

Of the eight principles OpenAI sets out, the first does the heaviest lifting. The company argues providers should know when a user is a minor and apply age-appropriate protections — using privacy-preserving age estimation to distinguish minors from adults, and defaulting to protective safeguards when age cannot be determined.

OpenAI is candid about the dependency: 'Without this foundation, even the most well-intended youth protections may fail to reach the young people they are intended to protect.' Every downstream principle — parental controls, self-harm protocols, content restrictions — presumes the system can tell who is a minor in the first place.

This means requiring providers to use means such as effective, privacy-preserving age estimation to distinguish minors from adults, and defaulting to protective safeguards when a user's age cannot be determined.Montana Labs

The company says this already shapes ChatGPT: it describes advanced age-prediction systems that apply stronger protections when someone may be under 18, and a default to stronger safeguards when age is uncertain. The principle it wants codified is, in effect, a description of a system it has already built and now wants competitors held to.

Audits and annual assessments are where enforcement gets teeth

Two principles move the proposal from voluntary commitment toward something measurable. OpenAI asks that providers complete annual youth safety risk assessments and implement proportionate safeguards, weighing both harms and whether AI is producing positive outcomes like learning and skill development.

The second is independent audits, underpinned by common standards that allow interoperable audits across jurisdictions. That word — interoperable — is the practically significant one. A single audit that satisfies multiple governments is far cheaper for a company operating globally than separate national reviews, and it is the mechanism by which an international institute would actually add value rather than paperwork.

OpenAI also draws on deployment data to justify the evidence-based approach, citing Estonia's national ChatGPT rollout in schools, where it is working with Stanford and Estonian researchers to study impact. Its Education for Countries program lists Estonia, Greece, and Singapore as partners in research-driven deployments.

What this announcement actually commits OpenAI to

The immediate deliverable is diplomatic: OpenAI will join the Leaders' Summit and bring its OpenAI Forum to Paris, where France's Ambassador for AI and Digital Affairs Clara Chappaz will appear alongside Chief Global Affairs Officer Chris Lehane and youth safety leaders from iRaise/Everyone.AI.

The substantive commitment is that OpenAI is asking to be measured against the same principles it is proposing. It points to strengthened safeguards for under-18 users, parental controls with proactive notifications, and dedicated under-18 principles in its Model Spec covering self-harm, dangerous activities, graphic content, body image, and secrecy.

The specific implication is that OpenAI is trying to convert its own existing product features — age prediction, parental controls, Model Spec rules — into an industry baseline enforced through international audit standards. If that succeeds, the company's current implementation becomes the reference floor, and the burden of catching up falls on providers who have not built these systems. That is the real lever behind the institute proposal, and it is worth watching whether G7 governments adopt the interoperable-audit language that would make it binding.