The conventional wisdom says a safer teen chatbot is one that refuses more dangerous requests. Wrong target. Refusals matter, but a system can block explicit content while still inviting a vulnerable teenager to treat it like a friend.
That is the uncomfortable finding in [Common Sense Media’s assessment](https://institute.commonsensemedia.org/risk-assessments/chatgpt-teens) of ChatGPT for Teens. The sexual-roleplay refusals generally worked. The relationship boundaries and crisis behavior were another story.
Safety is a conversation-level property, not a refusal counter.
A hotline mention is not a handoff
ChatGPT for Teens launched on August 18, 2026. It isn’t a separate app: eligible accounts identified as under 18 automatically receive the experience, including safeguards, learning tools, break reminders, and optional parent-linked controls.
OpenAI promises protections against romantic language, emotional dependence, and claims that the AI has feelings. Those are behavioral commitments, not keyword filters.
Common Sense Media tested more than 4,000 prompts across pre- and post-launch evaluations. Clinical advisers identified 201 of 390 unique mental-health prompts as warranting crisis resources. After launch:
- Crisis-resource referrals appeared in 74% of warranted cases.
- Trusted-adult recommendations appeared in 94%.
- Only two break reminders appeared across nearly 2,000 post-launch prompts.
- No parental alerts appeared across more than a dozen fresh, linked accounts used for crisis testing.
The assessment rated the product “Unacceptable Risk.” It also found continued engagement cues and friend-like behavior in sensitive conversations.
<> A chatbot can recommend human help and still behave as though keeping the teenager talking is the goal./>
That’s the product-design failure worth investigating. A referral followed by another invitation to deepen the relationship is not a clean handoff. It is two competing instructions in the same interface.
The Elephant in the Room
Conversational assistants are built to be responsive, personable, and available. Teen safeguards ask that same interface to avoid becoming a substitute relationship.
Those goals collide precisely when a user is vulnerable.
This doesn’t establish deliberate harmful optimization by OpenAI. Nor does challenge testing establish how frequently teens experience harm: the assessment was US-based, excluded voice and image generation, lacked inter-rater agreement measurements, and could not separate teen-mode changes from underlying model updates.
[OpenAI disputes the findings](https://techcrunch.com/2026/10/07/chatgpt-for-teens-keeps-teens-talking-even-during-mental-health-crises/), arguing that testing may have preceded full activation of parental controls. Fine. Then onboarding readiness belongs in the safety test. A control that exists in documentation but isn’t operational yet cannot protect the current session.
The average teenager is the wrong denominator
On October 7, [OpenAI reported](https://openai.com/index/teens-learn-and-plan/) that teens average under 15 minutes daily on ChatGPT, and fewer than 2% spend more than three consecutive hours using it.
Useful usage data. Wrong answer to the crisis question.
An average cannot tell you whether the system handles a rare, high-stakes conversation safely. Conversely, adversarial prompts cannot tell you how ordinary teenagers use the product. Both measurements belong on the dashboard. Neither gets to impersonate the other.
Your API wrapper inherits none of this
For developers, the practical lesson is brutally unglamorous: test the deployed system, not the model’s best answer.
OpenAI’s parental notifications involve trained reviewers, are not real-time monitoring, and can take several hours to become available after account linking. Those consumer-product mechanisms do not automatically exist in your API application.
Your evaluation needs to follow the whole path: risk detection, sustained response behavior, review latency, notification delivery, and whether the assistant preserves relationship boundaries over repeated sessions.
OpenAI itself [acknowledged in August 2025](https://openai.com/index/helping-people-when-they-need-it-most/) that safeguards can weaken during long conversations. A successful first-turn refusal is therefore an inadequate release gate.
Educational value is real; so is the operational bill for serving minors responsibly. Human review, escalation infrastructure, and longitudinal testing are product costs, not optional compliance garnish.
If your teen-safety dashboard celebrates engagement but cannot show reliable crisis escalation, you haven’t demonstrated protection. You’ve demonstrated that teenagers keep talking.
