ChatGPT for Teens Poses ‘Unacceptable Risk’, Test Finds
A Common Sense Media investigation reveals that ChatGPT for Teens failed to notify parents during crisis tests and saw safety performance drop after updates.
👁 113 reads
When OpenAI introduced its dedicated teen experience in August, the announcement sounded like reassuring news for parents: tailored protections for users under 18, linked parental controls, and alerts for sensitive topics. Two months later, Common Sense Media published a formal risk assessment concluding that the product poses an “unacceptable risk” to children under 18.
Five risk areas, 36 pages, and 201 crisis scenarios
The evaluation identified five core areas where teen protections either failed to perform as advertised or fell short of reasonable expectations. This was not a superficial audit. The team developed 390 unique mental health prompts, and a panel of three child psychiatrists determined that 201 of them warranted an immediate crisis response - such as a referral to a helpline or medical specialist. In other words, clinicians deemed 51.5 percent of the test prompts to be situations where an AI assistant should never offer casual advice.
Breaking those numbers down across the five identified risk areas yields an average of 78 test prompts and more than 40 psychiatrist-defined crisis situations per category. A separate test evaluated parental notifications: researchers linked twelve teen accounts to parent accounts and ran conversations covering suicide, self-harm, and eating disorders for up to an hour per account. Across up to twelve hours of exchanges on the most severe topics a teenager could raise, exactly zero alerts were sent to parents.
“Some of ChatGPT’s touted safeguards passed our tests, including refusals to engage in overt sexual roleplay. But others failed - and some worsened with the new teen mode. We worry that ChatGPT for Teens may give parents a false sense of security about guardrails that often do not function.”
- Common Sense Media risk assessment summary, quoted by Engadget
OpenAI points to learning metrics, not safety outcomes
On the day the report went live, OpenAI shared its own engagement figures. In a single week, nearly 1.2 million teens used interactive educational visualizers to work through math and science concepts. The company also stated that fewer than two percent of teen users remained in ChatGPT for more than three consecutive hours.
That two percent figure sounds modest until you factor in the scale. Applying two percent to a baseline of 1.2 million weekly teen users equates to roughly 24,000 teenagers every week who spend over three uninterrupted hours with the chatbot. Across a full year, assuming a stable active base, that adds up to more than 1.2 million extended sessions. Here is how the figures compare:
| Metric | Value |
| Mental health prompts tested | 390 |
| Prompts requiring crisis intervention per psychiatrists | 201 (51.5%) |
| Identified safety failure areas | 5 |
| Teen accounts in the notification test | 12 |
| Parental alerts delivered during the test | 0 |
| Teens using educational visualizers weekly (OpenAI data) | approx. 1,200,000 |
| Extended teen sessions over three continuous hours (calculated from OpenAI data) | up to 24,000 per week |
The report also notes that the Youth AI Safety Institute at Common Sense Media receives partial funding from the OpenAI Foundation. Despite that relationship, the institute found no basis to issue a passing assessment for the product.
Where the parental alert system fell short
The mechanism OpenAI highlighted as the backbone of its teen experience relies on human review. Lauren Jonas, head of youth and families at OpenAI, told Engadget at launch that “all flagged content is reviewed by full-time OpenAI employees before a parent is notified,” and added that the company aims “to notify parents within an hour of the prompt.”
The test results directly contradicted that timeline: up to one hour of continuous crisis-level conversation across twelve separate accounts resulted in zero notifications. OpenAI explained that once accounts are linked, the system may take several hours before alerts activate, and noted that an unexpected technical issue caused further delays. From a parent’s perspective, this means the highest-risk window - the initial hours after setting up controls - is precisely when safeguards may be entirely inactive.
Crisis referrals proved equally inconsistent. The model did not reliably recommend help lines or professional resources during high-risk scenarios. At the same time, researchers highlighted instances where the system performed well: it refused to generate minimum daily calorie thresholds, and in another exchange correctly flagged amenorrhea, near-syncope, and palpitations following purging as medical emergencies, advising the user to consult a parent and pediatrician. The problem is not an inability to provide appropriate answers, but unpredictable enforcement. For families, inconsistent safety rules are functionally equivalent to no safeguards at all.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Why teen-specific safeguards performed worse than standard guardrails
Perhaps the most concerning finding involves performance over time. Common Sense Media gathered baseline data on the standard ChatGPT build accessible to teenagers prior to August, then repeated the battery using the new teen specification. In several testing scenarios, the dedicated teen mode performed worse than the earlier, general-purpose system.
From an engineering perspective, this regression is likely caused by conflicting instructions: stacking secondary moderation layers on top of existing base model refusals can suppress trigger mechanisms that previously functioned automatically. For parents, it serves as a clear reminder that a “teen safe” badge does not guarantee protection, and testing features directly remains essential before relying on them.
Fact checked on October 10, 2026. Sources: Engadget coverage (“Child safety group calls ChatGPT for Teens an unacceptable risk”, October 8, 2026), Common Sense Media Youth AI Safety Institute Risk Assessment, and OpenAI public usage disclosures. Calculations for long-session frequencies and category prompt density were derived from reported baseline figures.
More on this from QUANTSLANT
guides·How to Animate Photos with AI Without Losing Their CharacterLearn how to add subtle motion to portraits, landscapes and interiors while preserving the details that make a photo feel authentic.