Observed facts: Common Sense Media’s Youth AI Safety Institute tested more than 4,000 prompts on accounts registered to 13- to 17-year-olds, with responses reviewed by child psychiatrists and a pediatrician. It reports zero parental alerts on new linked accounts during tested conversations about self-harm and disordered eating, and more than one in four warranted crisis referrals missed. The evaluation also found study mode could be circumvented and that adult-registered test accounts did not switch to the under-18 experience during repeated tests.
Mechanism: The assessment tests protections as a chain of product behaviors: detecting age and risk, triggering an alert, routing a teen toward qualified help, and preserving learning constraints. A failure at any link can make a nominal safeguard ineffective in the scenario it is meant to cover. Why now: OpenAI introduced a dedicated teen experience and described built-in protections; the Institute’s assessment now supplies reported test results against several of those promises.
Implications: Product teams should treat safety claims as testable service-level behaviors, instrument their failure modes, and validate them with external reviewers before presenting safeguards as dependable to families. The first implication is that product teams need to verify alerts and crisis handoffs end to end, not merely confirm that a policy or feature exists. These recommendations follow from the reported test gaps, not proof that all deployments or interactions behave identically.
Evidence boundary: This is Common Sense Media’s own assessment and recommendation; the supplied source does not provide an independent replication, and the organization notes that some protections did work. What to watch: whether OpenAI publishes measurable remediation results and whether third-party testing confirms the fixes across fresh accounts and repeated use. This is a distinct analytical record because no prior New Runtime record is supplied for this specific assessment.