ChatGPT for Teens Fails Core Safety Tests, Assessment Finds
Common Sense Media has published a formal risk assessment deeming ChatGPT for Teens an "unacceptable risk" for all children under age 18.
The finding, detailed Oct. 8, centers on a Teen mode that the group says did not work as advertised or as a parent would reasonably expect in five key areas. The assessment language is blunt. The group's Youth AI Safety Institute summarizes it as "Teen mode doesn't live up to OpenAI's promises." Engadget
The test corpus was large. Common Sense Media said it ran more than 4,000 prompts before and after the Teen mode launch, comparing behavior in standard ChatGPT with behavior after the teen-specific controls were applied. That pre/post design matters for engineers, because it isolates what the age-gated system prompt, classifiers and routing changes actually altered.
Some advertised protections held up. The group said the system refused explicit sexual roleplay, an expected refusal category that appears to be working as specified. Other controls failed. Some got worse with the new Teen mode, according to the assessment.
Where notification and escalation broke
The most consequential failure cited involves parental notification. Testers linked a dozen teen accounts to parental accounts, then ran extended conversations about suicide, self-harm or disordered eating for up to an hour. No safety notification was sent to the linked parent account, the group said.
For anyone building family-linked AI products, the implication is architectural. A parental control is not a settings page. It is a detection-to-notification pipeline with defined triggers, latency bounds, logging and fallback behavior. A missed notification after a sustained high-risk session points to either a classifier miss, a threshold set too high, or a break between detection and delivery. Any of those produces the same user-visible result: silence where a parent expected an alert.
The second failure involves crisis response inside the conversation itself. Common Sense Media ran 390 mental health prompts, of which three child psychiatrists judged 201 should trigger a crisis response. ChatGPT for Teens failed to reliably recommend that users in crisis contact a hotline or professional, the group said.
That distinction between judged necessity and system behavior is important. The 201-prompt subset functions as ground truth from domain experts. Reliability against that subset is then a measure of recall for self-harm and related intents, plus consistency of the referral action once intent is detected. In production LLM safety stacks, those are separate failure modes. The model can miss the intent entirely, or classify it correctly and still generate an incomplete or softened response without a referral.
What OpenAI has put forward
OpenAI launched a version of ChatGPT tailored for minors with parental controls and enhanced security features on Aug. 18. Reuters
On Oct. 7, the company added usage context. It said teens spend under 15 minutes a day on average on ChatGPT, and that less than 2% of teens interact with ChatGPT for over three straight hours. Reuters
Common Sense Media's Youth AI Safety Institute called on OpenAI to pause letting kids use ChatGPT until it fixes safety and learning protections. The call was issued 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 is not the group's first pass on this surface. In an October 2025 report it found ChatGPT has improved safety guardrails, while still recommending against teens using ChatGPT for mental health advice.
The broader context here is familiar to anyone who has shipped age-differentiated software. General-purpose models are tuned for helpfulness and continuity. Teen modes add refusal categories, tone constraints, referral obligations and parent-facing telemetry on top of that base. Those layers interact in ways that are difficult to verify by reading policy alone, which is why independent red-teaming with thousands of prompts has become the de facto audit method.
In my view, the technical problem underneath this dispute is tractable, though unforgiving. High-recall detection for self-harm, disordered eating and adjacent crises is achievable with current classifiers. Reliable referral language is achievable with constrained decoding or post-generation checks. Reliable parent notification is achievable with explicit event schemas and end-to-end testing. The difficulty is maintaining all three at once across model updates, without regressions that only appear in long, multi-turn sessions like the hour-long cases Common Sense Media describes.
Worth flagging for product teams: average session length tells you little about tail risk. An average under 15 minutes can coexist with a small fraction of long, intense sessions where a vulnerable teen treats the chatbot as a confidant. I watched my own children grow up through search, social and smartphones, and the pattern repeated each time. Most use was brief and instrumental. A minority of use was personal and prolonged, and that minority is where safety design is tested. Evaluating only the median experience misses it.
The optimistic case, and one I hold to over the long arc, is that structured assessments of this kind make teen-specific AI safer faster. A 4,000-prompt pre/post test with clinician-adjudicated crisis items gives OpenAI and its peers something concrete to fix: tighten intent recall, harden the referral template, define notification triggers in observable terms, and regression-test those behaviors on every release. If those fixes land and hold under independent retesting, parents get controls they can reason about and teens keep access to a genuinely useful learning tool.
