OpenAI put users aged 13 to 17 into a separate policy bucket and said uncertain ages will default to the under-18 experience. My read is simple: this is not just a safety statement. It is a product-liability redesign. The company is formalizing two operating modes — “treat adults like adults” and “protect teens first” — then using age prediction as the switch. Most of the real risk now sits inside that switch: how accurate it is, how users appeal it, and who pays for false positives.
The article gives a few hard boundaries. ChatGPT is intended for ages 13 and up. Under-18 users will not get flirtatious dialogue. They also will not get suicide-themed creative writing help. If a minor appears to face imminent self-harm risk, OpenAI says it will try to contact parents, and if that fails, contact authorities. That is a major escalation path. The problem is that the post does not disclose the metrics that matter: age-prediction accuracy, country-by-country differences, human review rates, evidence thresholds for parent outreach, or remediation when the system gets it wrong. Without those details, nobody outside the company can tell whether this is narrowly scoped or overbroad.
I’m skeptical of behavior-based age inference in a chat product. Yes, age estimation from language, usage patterns, device signals, and session behavior is already common in trust-and-safety systems. Social platforms have been doing versions of this for years. But chat is messier than feed ranking. Users role-play. They write fiction. They imitate other voices. They do homework. They test prompts. A 29-year-old writing YA fiction can look younger than a 16-year-old Olympiad student. Once age classification controls not just recommendations but content permissions and crisis escalation, the error cost changes completely. It stops being “wrongly categorized content” and becomes “creative work blocked, account constrained, or external contact triggered.”
The broader context is easy to see. Meta, Instagram, and TikTok have all spent the last two years tightening teen defaults, adding parental controls, and leaning on age assurance. UK and EU regulators have also pushed platforms toward stronger age checks and child-safety duties. I can’t verify that this OpenAI post maps to one specific regulatory deadline, but it reads like preemptive compliance architecture. Chatbots used to defend themselves as information tools. This post quietly concedes that when users treat the model as companion, counselor, or emotional outlet, the platform will be judged more like a high-risk interactive system.
Altman is also trying to hold two positions at once. In the first half, he frames AI conversations as highly sensitive, close to doctor- or lawyer-level privacy in spirit. In the second half, he carves out automated monitoring, human review for severe cases, and parent or authority contact for minors in imminent danger. I understand the logic. I don’t fully buy the analogy. Doctor and lawyer privilege sits on top of legal doctrine, licensure, and clear remedies. Model providers do not have that institutional scaffolding. If OpenAI wants the field to accept stronger privacy claims, it needs to publish much sharper operational criteria: what counts as imminent harm, whether a human must review before outreach, how cross-border escalation works, and how an adult wrongly placed into the teen bucket restores access. The article does not provide that.
There is also a product cost here that people tend to underestimate. Once adults get broader latitude for fictional self-harm contexts or romantic dialogue, while teens get tighter restrictions, OpenAI has to maintain at least two behavior stacks across model policy, account controls, and safety routing. That gets brittle fast. Anthropic, Character.AI, and Meta AI have all run into adjacent problems over the last year: once policy varies by age, region, and risk state, the system stops behaving like one model and starts behaving like a policy router wrapped around a model. The same prompt can produce sharply different outcomes depending on account state. That is rough for user trust and rough for benchmarking, because you are no longer evaluating only model capability. You are evaluating policy segmentation.
My pushback is not on the goal. Protecting minors more aggressively is defensible. My issue is that OpenAI disclosed the principle and not the audit surface. If even 2% to 5% of adults are misclassified into the under-18 path, the friction will be very visible. If crisis detection lacks precision, outreach to parents or police will burn trust fast. I don’t have those numbers, and the post does not either. So I’m not ready to treat this as a mature safety framework yet. Right now it is a strong policy claim attached to an undisclosed classifier. Whether it holds up will come down to false-positive rates, appeal latency, and human-review transparency.