AI safety looks different at 8 than it does at 16

We serve kids from 8 to 16. That's an enormous range — an eight-year-old and a sixteen-year-old are not the same kind of person using the same kind of tool, and we don't treat them like they are. "Age-appropriate AI use" isn't a single rule we apply uniformly with the volume turned down for younger kids. It's a genuinely different posture at different ages, and it's worth explaining what actually changes, because "we take safety seriously" is a sentence every program says and none of them define.

What stays constant across the whole range

Some things don't move regardless of age, because they're not about maturity — they're about the room itself. Every session is supervised. Every mentor has been through a background check. The building is insured the way any program working with minors has to be. And the standing rule that we build things that help rather than things that surveil applies to an eight-year-old's project exactly as much as a sixteen-year-old's. None of that scales down for younger kids or loosens for older ones. See /safety for the full version.

What actually changes: the distance between the kid and the model

The real variable isn't "how much supervision" — it's how directly a kid interacts with an AI system, and how much of the judgment about what it says gets delegated to the kid versus handled by the design of the activity.

At the younger end (8–10ish), activities are built so the AI's role is narrow and the mentor is doing more of the interpreting. A younger kid might be directing an AI to generate an illustration for a story they wrote, or asking it to suggest three ways to fix a piece of code that's broken — bounded prompts, visible output, a mentor right there to talk through anything surprising the model produces. The kid is making creative decisions; the model is a tool inside a structure an adult built.

At the older end (13–16), kids work more directly with the raw tool — writing their own prompts, iterating on them, debugging why an AI-generated function doesn't do what they expected, sitting with the fact that a model can be confidently wrong. This is deliberate, not a laxer policy. Part of what we're teaching older kids is exactly that judgment: how to evaluate what a model gives you instead of accepting it, which is a skill you can only build by being trusted with more direct contact.

In the middle (11–12), it's genuinely a gradient, and it depends on the kid as much as the birthday. We adjust per kid, not per fixed cutoff — some eleven-year-olds are ready for more direct work than some thirteen-year-olds, and mentors are the ones making that call in the room, not a form.

Why we don't just apply the strictest rule to everyone

We could simplify our lives by running every age group under the eight-year-old's ruleset. We don't, because it would waste the most valuable thing the older end of our range is actually here for. A sixteen-year-old who's never allowed to write their own prompt, debug their own model output, or sit with an AI's mistake isn't being kept safe from anything real — they're being kept from the actual skill. Over-protecting the kids who are ready to do the harder thing is its own failure mode, just a quieter one than under-protecting the kids who aren't.

The part that doesn't change with age: nothing leaves the room without a reason

Regardless of a kid's age, we don't build the kind of AI product that pretends to be a friend, remembers everything a kid says, or nudges them to keep coming back out of guilt rather than interest. We've written about why elsewhere — that rule isn't graded by age either. It's a floor, not a dial.

If you're trying to figure out where your specific kid lands on this, that's exactly the conversation a mentor should have with you before their first session, not something you have to guess from a policy page.