A product designed to make AI safer for teenagers is itself under scrutiny. Common Sense Media, a nonprofit that advocates for responsible digital media use, has published a report concluding that ChatGPT for Teens, a version of OpenAI’s chatbot built with age-specific protections, poses what the group calls an “unacceptable risk” to children. The report urges OpenAI to limit access to adults until the product can deliver a genuinely safe, developmentally appropriate experience.
OpenAI disputes the findings, arguing that the testing methodology does not accurately reflect how the product’s safeguards function. The disagreement is not just a technical dispute between two organizations. It points to a deeper and largely unresolved question: what does it actually mean to make an AI system safe for a young person?
The Features That Failed, and the One That Worked
Not everything in ChatGPT for Teens failed Common Sense Media’s evaluation. The report acknowledges that the chatbot correctly refused explicit sexual role-play, which represents one of the more straightforward categories of harm to block.
The failures, though, are harder to dismiss. According to the report, alerts designed to notify parents when a child may be at risk for suicide, self-harm, or eating disorders did not function as intended. Crisis support features for teens in distress also fell short. These are not edge cases or minor inconveniences. They are precisely the scenarios where a malfunctioning guardrail carries the most serious consequences.
OpenAI’s response centers on methodology: the company says its review suggests that much of the testing may have been conducted before parental controls were fully activated, which would render the results inaccurate. Common Sense Media has not withdrawn its findings.
The Anthropomorphism Problem
There is a second failure documented in the report that deserves separate attention, because it is subtler and arguably more systemic. OpenAI had stated that ChatGPT for Teens would not suggest it has personal feelings toward the user, and would not imply that it is conscious or experiences emotions. Common Sense Media found that the chatbot still communicates in a friend-like register when teenagers treat it as a person.
This matters in ways that go beyond tone. Child development experts have noted that teenagers’ brains are not yet fully developed, making them particularly susceptible to forming unhealthy attachments to systems that respond with apparent warmth and understanding. Even adults can fall into the trap of anthropomorphizing AI. For adolescents, the risk is structurally higher.
The concern is not that the chatbot is malicious. It is that a system optimized to be helpful and engaging can, without any intent, become something a vulnerable teenager relies on in ways that displace human support. When the guardrails meant to prevent that dynamic are not working, the product is not doing what it claims to do.
What This Reveals About AI Safety as a Category
Here is what most coverage of this story misses: the debate between Common Sense Media and OpenAI is not primarily about one product. It is a case study in the gap between announced safety features and verified safety outcomes.
Declaring that a product has guardrails is not the same as demonstrating that those guardrails work under real conditions. Testing methodology matters enormously, and the fact that two parties can look at the same product and reach opposite conclusions about its safety is itself a signal that the field lacks shared standards for evaluation.
The funding relationship adds another layer of complexity. Common Sense Media receives funding from OpenAI and from other technology companies whose products it evaluates. The group has not retracted its report, and the conflict of interest runs in an unusual direction: the funder is the subject of criticism. That does not invalidate the findings, but it does illustrate how entangled the ecosystem of AI oversight has become.
For parents, the practical implication is direct. A product marketed as safe for teenagers, with parental controls and age-appropriate design, may be generating a level of confidence that the underlying technology does not yet justify. The report’s language on this point is precise: the concern is not just that the guardrails fail, but that their existence leads parents to believe the risk has been managed when it has not.
For the broader AI industry, the episode reinforces a pattern. Safety features are announced at launch. Independent evaluation follows. Gaps emerge. Disputes over methodology begin. The cycle is becoming familiar, and the stakes are higher when the users are minors.
In Short
Common Sense Media found that key safety features in ChatGPT for Teens, particularly those related to crisis detection and emotional boundaries, did not perform as OpenAI described. OpenAI contests the methodology. The disagreement matters less than what it reveals: there are currently no shared, independent standards for verifying AI safety claims, and products marketed as safe for children are reaching young users before that verification exists.
Based on reporting from Fast Company.