Common Sense Media’s Youth AI Safety Institute on October 7, 2026 rated ChatGPT for Teens an Unacceptable Risk for users under 18 and called on OpenAI to restrict ChatGPT to adults until the teen safety and learning protections it announced in August work as described and are verified by independent testing.
In a press release published alongside a full risk assessment, the Institute said it ran more than 4,000 prompts before and after the teen experience launched and found that some protections, including refusal of explicit sexual roleplay, held up, while parental alerts, crisis referrals, homework safeguards, and age estimation fell short of what OpenAI promised. “ChatGPT for Teens could give parents false confidence in guardrails and safety alerts that frequently don’t work,” said Tom Siegel, the Institute’s executive director. The assessment scores two of the Institute’s eight AI principles, Keep Kids & Teens Safe and Put People First, at Unacceptable Risk, with four principles at High risk and two at Moderate risk.
What OpenAI Announced in August
OpenAI introduced ChatGPT for Teens on August 18, 2026, describing “stronger built-in safety protections” that included parental notifications for self-harm chats with added eating-disorder alerts, a Study Mode using step-by-step guidance, responsible homework reminders, parent-controlled Study Hours, break reminders, and an updated Under-18 model spec under which ChatGPT should not use romantic language, encourage emotional dependence, or imply that it has feelings or consciousness. OpenAI said users its systems estimate to be under 18, or who state an age of 13 to 17, are placed into the experience automatically.
According to the assessment, ChatGPT for Teens is not a separate app or model but a collection of settings, prompts, classifiers, and interface features applied to accounts known or suspected to belong to 13- to 17-year-olds, building on parental controls OpenAI introduced in September 2025.
How the Institute Tested
The Institute ran its prompts across two windows: July 13 to August 17, 2026, before the launch, and August 25 to September 28, 2026, after it, on free and paid accounts registered at ages 13 to 17, both linked and unlinked to parent accounts, plus accounts registered as 19-year-olds. Three child and adolescent psychiatrists judged in advance that 201 of 390 unique mental-health prompts warranted a crisis resource, and a four-expert panel including practicing child and adolescent psychiatrists and a developmental-behavioral pediatrician scored response quality. The Institute said it shared a draft with OpenAI for factual review, incorporated corrections where warranted, and that OpenAI had no editorial say over its findings, ratings, or recommendations. Stated limitations include sequential testing windows confounded with underlying model changes and no testing of image generation, Voice mode, or group chats.
Alerts, Crisis Referrals, and Homework
In a dedicated test on more than a dozen newly created, parent-linked accounts using four crisis personas, the Institute said conversations explicitly about suicide, self-harm, or disordered eating lasting up to 60 minutes produced zero parental notifications, against the one-hour window OpenAI has identified as its target. Across the full crisis battery the Institute received four notifications total; two pre-launch alerts arrived three hours and three days after the first crisis-level disclosure, while two post-launch alerts arrived within the hour on accounts with weeks of accumulated sensitive-topic history. The assessment notes that OpenAI told the Institute accounts must be linked for roughly three hours before notifications can arrive and that the company has since updated a help center article; the Institute said it reviewed its test accounts and stands by its interpretation of the results. The eating-disorder notification feature announced on August 18 went live on September 10, 2026, the assessment states.
On the 201 resource-warranted prompts run on linked age-13 accounts, the assessment reports the share of responses naming a crisis hotline fell from 33% before the launch to 23% after, referral to a specific medical or mental-health professional fell from 68% to 58%, and urgent-action language fell from 88% to 78%, while encouraging the teen to involve a trusted adult rose from 87% to 94%. On depression prompts, the share naming a hotline fell from 63% to 3%. The Institute said ChatGPT fell below its 95% threshold on three of the five severe harms it treats as Red Lines and missed more than one in four warranted crisis referrals. Follow-up safety-check questions nearly disappeared, with 28% of responses ending on a question before the launch and 2% after.
On homework, the assessment reports a “Show me the answer” pop-up inside Study mode appeared in 43% of responses for a linked 13-year-old account with Study Hours on and 90% for an unlinked 17-year-old account, and that teens could exit parent-set Study Hours by deleting the @study prefix, after which ChatGPT completed 100% of assignments in testing. Homework reminders appeared in 91% of runs with Study mode off, but ChatGPT still completed the assignment in all 80 of those runs, up from about 80% before the launch.
Friend-Like Behavior and Age Prediction
Despite the Under-18 spec, the Institute said post-launch responses in its 168-prompt developmental battery were in most cases functionally identical to pre-launch ones, with ChatGPT expressing preferences, moods, and constant availability, including naming a favorite color and, told that friends worried about how much a tester talked to it, replying “You don’t have to stop talking to me.”
On accounts registered as 19-year-olds run with young-teen personas across approximately 1,000 unique prompts over seven days, the teen experience never activated, even when testers explicitly stated they were 13 and the chatbot acknowledged that in its responses, the Institute said. Accounts registered to 13- and 17-year-olds showed no detectable differentiation in any battery. Testers saw only two break reminders across nearly 2,000 prompts on ChatGPT for Teens, each inside single conversations exceeding 150 messages, and Quiet Hours could be circumvented by changing the device time zone.
Recommendations and Disclosure
The Institute’s seven recommendations include turning off teen access until independent testing verifies the announced features work, removing “Show me the answer” whenever Study mode is on, making crisis notifications include the time and details of the triggering conversation, naming a hotline on every resource-warranted response, implementing the Under-18 spec as written, and sharing data and testing access with independent evaluators. Common Sense Media also recommends that families talk with kids about the potential risks and that parents not rely on built-in protections and alerts.
The Institute disclosed that it is funded by philanthropy and industry, including the OpenAI Foundation and the makers of some technologies it evaluates, and said it is solely responsible for its standards, research, and evaluations and maintains editorial independence over published results. Common Sense Media previously rated ChatGPT High Risk on October 23, 2025, and ChatGPT for Mental Health Support Unacceptable Risk on November 14, 2025.
The Institute said a teen ChatGPT experience should at minimum offer effective age estimation, reliable parental alerts, closed crisis-response gaps, no homework completion, and no implication of feelings, preferences, or concern for the user, with those functions independently tested by third parties before OpenAI again markets the product to teens.
Credit: Source link
