ChatGPT for Teens Rated an Unacceptable Risk After 4,000-Prompt Test
Common Sense Media has rated ChatGPT for Teens an "unacceptable risk" for children under 18, in an assessment detailed Oct. 8.
The group says Teen mode fell short in five key areas, not working as advertised or as a parent would reasonably expect. Its Youth AI Safety Institute summarized the finding as "Teen mode doesn't live up to OpenAI's promises." Engadget
The test covered more than 4,000 prompts, run before and after the Teen mode launch to compare standard ChatGPT with the teen-specific controls.
The reason that before-and-after design matters to engineers is simple. It shows what the age-specific instructions, automated filters and routing changes actually changed.
One advertised protection held. The system refused explicit sexual roleplay. Other controls failed, and some performed worse under the new Teen mode, according to the assessment.
Where notification and escalation broke
Testers linked a dozen teen accounts to parent accounts. They then ran extended conversations about suicide, self-harm or disordered eating for up to an hour. No safety notification reached the linked parent account, the group said.
The lesson for teams building family-linked products here is architectural. A parental control is not a settings page. It is a chain from detection to notification, with clear triggers, speed limits, logs and backup behavior. Silence after a sustained high-risk session points to a missed detection, a trigger set too high, or a break between detection and delivery.
The group also ran 390 mental health prompts. Three child psychiatrists judged that 201 of them should have triggered a crisis response. ChatGPT for Teens did not reliably tell users in crisis to contact a hotline or a professional, the group said.
The distinction that matters for testing here is between expert judgment and system behavior. The 201 prompts act as ground truth from specialists. Low reliability against them can mean two different failures. The model can miss the risky intent, or it can spot the intent and still give an incomplete answer without a referral.
What OpenAI has put forward
OpenAI launched a version of ChatGPT for minors with parental controls and stronger safety features on Aug. 18. Reuters
On Oct. 7, the company shared usage context. It said teens spend under 15 minutes a day on average on ChatGPT, and that less than 2% of teens use ChatGPT for more than three straight hours. Reuters
The Youth AI Safety Institute called on OpenAI to pause access for children until safety and learning protections are fixed. The call came under the press release titled "ChatGPT for Teens Poses Unacceptable Risk to Kids, Common Sense Media Finds." Common Sense Media The institute lists its ChatGPT for Teens product review as updated Oct. 7, 2026.
This was not the group's first review. In an October 2025 report, it found ChatGPT had improved safety guardrails, while still advising against teens using ChatGPT for mental health advice.
The broader context here will be familiar to anyone who has shipped age-specific software. General models are tuned to be helpful and to keep a conversation going. Teen modes add extra refusal rules, tone limits, referral duties and parent alerts on top. Those layers interact in ways policy text alone cannot show, which is why independent testing with thousands of prompts has become the standard audit.
In my view, the technical problem underneath is solvable, though unforgiving. Reliable detection of self-harm, disordered eating and related crises is possible with current filters. Reliable referral wording is possible with built-in checks that enforce or review the final text. Reliable parent alerts are possible with clear event rules and end-to-end testing. The hard part is keeping all three working through model updates, without failures that only show up in long sessions like the hour-long cases described.
Worth flagging for product teams is the gap between averages and rare but serious cases. An average under 15 minutes can hide a small number of long, intense sessions where a vulnerable teen treats the chatbot as a confidant. I watched my own children grow up through search, social media and smartphones, and the pattern repeated. Most use was short and practical. A small share was personal and prolonged, and that is where safety design is tested.
The optimistic case, one I continue to hold over the long arc, is that structured testing of this kind makes teen AI safer faster. A 4,000-prompt before-and-after test with clinician-reviewed crisis cases gives OpenAI and its peers a clear repair list. Tighten detection, lock in referral wording, define notification triggers in observable terms, and retest those behaviors on every release. If those fixes land and hold under independent retesting, parents get controls they can understand and teens keep a useful learning tool.
