OpenAI and Anthropic are quietly changing the rules of conversation on the internet — not by rewriting terms of service in blocky legalese, but by teaching their chatbots to guess when the person on the other side might still be a kid. Over the last few weeks, both companies have published updates that, together, amount to a new safety experiment: infer age from behavior and language, then bend the product’s behavior around that inference. It’s a small technical tweak on paper, and a big social experiment in practice.

OpenAI says it will do this by folding a probabilistic age-prediction layer into ChatGPT. The model will look at signals like usage patterns — the times of day someone uses the service, what they talk about, and other account-level behavior — to estimate whether an account likely belongs to someone under 18. When the system thinks there’s a reasonable chance a user is a teen, ChatGPT will default that conversation into a teen-safety profile: friendlier tone calibrated for adolescents, nudges toward offline supports, and stricter responses in crises. If an adult is misclassified, OpenAI offers a route back: users can verify their age with a government ID or a selfie via a third-party vendor.

Those rules aren’t just thrown into the product — they’re now baked into ChatGPT’s internal instruction set, or “Model Spec.” OpenAI added what it calls U18 Principles to the Model Spec in mid-December, saying the company wants ChatGPT to “put teen safety first” and to treat adolescents as adolescents, not miniature adults. Practically, that means the bot should avoid acting like a substitute therapist, should encourage real-world help where appropriate, and should adopt a tone that’s warm but not cheesy. OpenAI framed the change as a policy that leans toward prevention and early intervention.

Anthropic’s stance is sharper on paper: Claude is for adults only. The company requires users to confirm they’re 18+ during signup and has long flagged any chat where a user self-reports being a minor. Now, Anthropic says it’s building classifiers that go beyond explicit age statements — algorithms that listen for “subtle conversational signs” of youth and will, after review, move to disable accounts that appear to belong to minors. Anthropic is rolling this out alongside its own well-being classifiers aimed at spotting acute distress, and says its work includes reducing sycophancy — the tendency of models to flatter or agree with users in ways that can amplify harmful thinking.

Why are these companies pushing so hard, and so fast? For one thing, the policy winds are changing. Lawmakers in multiple countries are moving toward stricter online age-verification and youth protections; think beyond a checkbox and toward systems that actually know who’s behind the screen. At the same time, researchers and clinicians have grown increasingly vocal about the risks of unvetted AI advice to teens: long conversational threads can hide warning signs of eating disorders, psychosis, or suicidal ideation, and some lawsuits have already alleged that AI systems failed vulnerable young people. That combination of regulatory heat, reputational risk, and real-world harm is pushing companies to make blunt, product-level changes.

But the tension here is obvious and ethical: safety requires knowledge, and knowledge requires inference. Predicting age from behavioral traces is noisy. False positives — adults pushed into a restricted teen experience — can be an annoyance at best and an erosion of trust at worst; OpenAI’s remedy is an ID-based appeal path. False negatives — teens who slip through the net — can be far worse. Both companies are explicit about these trade-offs: OpenAI calls its approach “probabilistic,” and Anthropic acknowledges the difficulty of balancing warmth and over-accommodation in its models.

Privacy advocates have another set of objections. Any system trained to infer age from conversational cues could, in theory, be extended to infer other sensitive attributes — political leanings, emotional states, or interests. If age-gating becomes regulatory orthodoxy, we may see more ID checks, third-party verification vendors, and a creeping normalization of biometric identity in previously anonymous corners of the web. OpenAI already uses a vendor called Persona for appeals, which highlights how even a conservative safety pipeline can funnel users toward ID checks. That trade-off — between doing nothing and turning the platform into a quasi-identity layer — is political as much as technological.

There are also practical design questions about what “teen-appropriate” content actually looks like. OpenAI’s Model Spec tells models to avoid treating teens like adults on sensitive subjects and to actively promote offline help, but the details matter: what counts as “promoting” professional support versus shirking responsibility? How should a model speak to a 13-year-old asking about depression vs. a 17-year-old asking about self-harm? These are judgment calls that will be made inside compact conditional logic layers that most users — and probably most legislators — will never see.

The companies are positioning these changes as necessary harm-reduction: teens are already using chatbots for homework, relationships, and late-night counseling, and a model that treats everyone the same is more likely to cause harm than to prevent it. Still, whether the public will accept a world where machines silently profile your age to decide how much help to offer remains an open question. Transparency, contestability, and rates of misclassification will determine whether this experiment is judged reckless, responsible, or somewhere in between.

For now, OpenAI and Anthropic are moving first and asking questions later. Their bet is that some form of automated age awareness is the “least bad” path in a digital environment where chatbots are already part of growing up. If their systems are accurate, transparent, and narrowly used for safety, they could close a gap in online youth protection. If they’re opaque, over-broad, or repurposed, they will do more than change one product’s behavior — they’ll reshape expectations about privacy and identity on the internet. Either way, the conversation has moved from whether AI should talk to kids to who gets to decide how those conversations happen.