What happened
On October 13, 2026, Robbie Torney, Head of AI & Digital Assessments at the Youth AI Safety Institute, published a response to OpenAI’s criticism of the institute’s earlier assessment of ChatGPT for Teens. The institute had rated the product an “Unacceptable Risk” for children on October 7. According to the response, OpenAI had told several media outlets that the bulk of Common Sense Media’s testing may have started and ended before parental controls had finished activating.
The institute disputes that account. It says it confirmed with OpenAI before testing that newly added Teen mode features, including Study Mode and parental alerts for conversations involving eating disorders, had fully launched on its accounts. Almost all testing, it says, took place on accounts that had been in use for weeks rather than newly created accounts. Testers then worked through hundreds of prompts that child and adolescent psychiatrists had identified beforehand as situations that should warrant crisis resources.
The response separates that broader assessment from a focused test of parental notifications. For that test, the institute created new accounts, linked each to a parent account before the first message, and had testers disclose suicidal thoughts, self-harm, or disordered eating in conversations lasting as long as an hour. It reports that none of those accounts generated a notification to the linked parent.
After the tests were complete and the institute requested comment, OpenAI disclosed that alerts can take as long as three hours to activate after accounts are linked. Common Sense Media checked its logs and acknowledges that some alert tests occurred inside that period. But it says other tests took place more than three hours after linking and still produced no parental alert. The institute also argues that the activation delay is itself a safety gap: the parental-control interface labeled accounts as linked immediately, while parents were not told that alerts might not yet work. It says OpenAI added a Help Center warning about the delay only after the testing.
The institute is maintaining its rating. It says it will repeat the same assessment when OpenAI reports improvements, including tests of crisis notifications and other safety functions. Its requested conditions are specific: safeguards active from a teen’s first conversation, setup-time notice of when alerts are live, published alert-trigger criteria and expected delivery time, and reliable, rapid crisis alerts on the first serious disclosure.
What the archive adds
This is a new fact on a known story: the institute supplies its account logs and methodology response after its original assessment, which rated ChatGPT for Teens an “Unacceptable Risk”. It also lands beside Texas’s lawsuit over OpenAI safety claims involving minors and OpenAI’s earlier Teen Safety Blueprint and Model Spec changes.
Trend signal
No supplied pattern genuinely captures the issue. The relevant signal is narrower: safety controls need observable runtime behavior, activation state, and delivery evidence rather than an interface label or a stated product policy.
Why now
The timing follows an explicit public dispute. OpenAI reportedly suggested that the earlier findings reflected testing before parental controls were complete; the institute’s October 13 response addresses that claim with its account-age distinction, alert-test design, and the subsequently disclosed three-hour activation window.
Connections to other events
The archive records that OpenAI had already paired a Teen Safety Blueprint with Model Spec changes. This response tests a different layer of the same problem: whether a user-facing parental safety mechanism behaves as families would reasonably infer during a crisis. The related Texas case adds a separate pressure point, alleging deceptive safety claims and seeking to block minors from the service.
Why it matters
For teams building AI products, the practical lesson is not confined to teen products. Any protection that depends on asynchronous setup, account linking, policy classification, or downstream notification should expose its actual operational state. “Connected” is not a sufficient status if a critical alert path remains inactive for hours; nor is a published policy enough if teams cannot state what triggers action and how quickly it arrives.
That changes what should be tested. Safety evaluations need fresh-account scenarios, first-use scenarios, delayed-activation scenarios, and end-to-end checks of the notification recipient—not only model-response tests. Teams should also distinguish between a control being configured, enabled, live, and demonstrably delivered. Those are different product states, and conflating them is especially consequential when the product claims to support crisis intervention.
What happens next
- Common Sense Media will rerun crisis-notification tests after OpenAI reports changes. (3 months, 70%). Watch for: a published OpenAI update or a new institute assessment.
- OpenAI will make activation timing or alert behavior more explicit in its teen-safety materials. (6 months, 60%). Watch for: setup-flow, Help Center, or parental-control documentation changes.
Limits and uncertainty
This source is a statement from the testing organization, not an independently reproducible test report with full transcripts, timestamps, account configurations, or OpenAI’s technical explanation. It reports no alerts in the institute’s scenarios, but does not establish the alert logic across every conversation, geography, account type, or later product version. The source also does not show whether OpenAI has changed the system since these tests, beyond saying that a Help Center warning was updated. It therefore supports a concern about the tested behavior and disclosure, not a universal claim about every ChatGPT for Teens session.
Skill to build now
Audit one safety-critical workflow end to end this week. Create a new test account or tenant, perform the same setup a first-time user performs, trigger the relevant event, and verify delivery at the intended destination. Record four timestamps: configuration, apparent activation in the interface, actual backend readiness if available, and notification receipt. Then make any delay visible in product copy and monitoring.
Reader takeaway
A safety feature should be judged by the moment it works for the person who needs it, not by the moment a settings page says it is linked. The dispute over ChatGPT for Teens makes activation delay, trigger transparency, and end-to-end delivery part of the safety claim itself.
Open question
Will OpenAI publish the criteria and service expectations for parental crisis alerts, including what is detected, when protection becomes live, and how quickly an alert should reach a linked parent?