You can't teach a tool you've never used

There's a conversation happening in kitchens and staff rooms right now that has the shape of a rule but the substance of a guess. Is she allowed to use it for homework? Should I turn it off at night? Is this cheating? The people asking are careful, well-intentioned adults. Almost none of them have used the thing for more than twenty minutes.

That's the actual problem. Not the technology, and not the kids. The gap is that we're trying to govern a tool from the outside — from headlines, from other parents, from a district memo — and it doesn't work, because every useful judgment about AI is a judgment you can only make from having been wrong with it yourself.

The gap is measured, not vibes

Gallup and the Walton Family Foundation surveyed 2,069 U.S. public K-12 teachers between February and March of this year. Eighteen percent said they get formal guidance on using AI at work. Forty-eight percent get informal guidance — meaning a colleague, a hallway conversation. Thirty-four percent get nothing at all.

Set that next to the adoption number from the same research program a year earlier: across the 2024-25 school year, six in ten teachers used AI for their work, three in ten at least weekly. So a clear majority of teachers are using this tool, and a clear majority of them are doing it without a single institutional sentence telling them how. The training didn't arrive. The use did anyway.

Parents are in the same position with less scaffolding and no colleague down the hall. If you're reading this and feeling behind — you are not behind relative to some competent peer group. There isn't one yet.

What "proficiency" actually means here

I want to be careful, because this argument gets overstated into something false and a little insulting. The claim is not that you need to understand transformers, or be able to explain attention heads, or hold an opinion about model architecture before you're allowed to parent or teach. That's gatekeeping dressed up as rigor, and it's exactly the thing that keeps good adults out.

The claim is narrower and much more achievable. You need to have used it enough to have been burned. Specifically, you need to have personally experienced:

A confident, fluent, completely wrong answer. Not read about hallucination — watched it happen to you, on a topic where you knew enough to catch it. Until that lands in your gut, you will either trust the output too much or distrust it in a vague, unhelpful way. Neither is a position you can teach from.

The difference between a bad prompt and a bad tool. Most of the "AI is useless" verdicts adults reach are the result of one lazy question. Most of the "AI is magic" verdicts are the result of one lucky one. Getting past both requires enough reps to know when the failure is yours.

The moment you noticed you'd stopped thinking. This is the one that actually matters for kids, and it can't be secondhand. There's a specific, slightly uncomfortable feeling when you accept a paragraph you didn't quite read because it looked right and you were tired. If you know that feeling in your own body, you can name it for a fourteen-year-old and she'll recognize it instantly. If you don't, everything you say about "using it as a tool, not a crutch" is a slogan she'll nod at and ignore.

What it's genuinely bad at. Long-horizon reasoning. Anything where the answer depends on knowing what happened last week. Anything where being confidently plausible is worse than saying nothing. Adults who've hit these edges give much better advice than adults working from a list.

That's the whole curriculum. Forty hours of ordinary use — plan the trip, draft the awkward email, argue with it about something you're an expert in — gets you all four. It is not a certification. It's just enough to have opinions you earned.

Why the outside-in approach keeps failing

The instinct when a technology arrives is to reach for a perimeter: a filter, a rule, a time limit, an honor code. Perimeters are appealing because they can be implemented by someone who doesn't understand the thing being contained. That's also why they fail.

We make a version of this argument in the output is the rubric — judge the work, not the clock — and the failure mode is the same here. A rule about how much is a rule you can enforce without knowing anything. A judgment about whether this particular use made her smarter or did her thinking for her requires you to be able to read the work — and increasingly, to read the work you need to know what the tool would have produced on its own.

Consider the most common real-world version of this: AI detectors. Schools bought them because they promised to restore a perimeter. An adult with forty hours of personal use would have predicted the outcome. A Stanford group ran 91 TOEFL essays written by non-native English speakers through seven widely used detectors: 61% were flagged as AI-generated on average, 89 of the 91 were flagged by at least one detector, and 18 were flagged by all seven. The same detectors were near-perfect on US eighth-grade essays written by native speakers — about a 5% false-positive rate. The study also showed that simple prompting defeats the detectors outright.

So the tool systematically accuses the students least able to defend themselves, while failing to catch anyone who actually tries to evade it. You did not need a research paper to see that coming. You needed to have watched the model produce five wildly different registers of the same paragraph on request, and to have noticed that "sounds like a machine wrote it" is not a property of machines. It's a property of careful, slightly stiff, second-language prose.

The order of operations

So the sequence we'd argue for, in a family or in a building, is:

1. The adults use it for a month, for real work, without kids involved. Not "evaluating an ed-tech product." Actually using it on your own tasks, where you care about the output.
2. The adults compare notes. This is the step everyone skips and it's where most of the value is. Two teachers comparing where the model lied to them is worth more than any vendor webinar.
3. Then you set expectations with the kid — and they'll be specific, defensible, and mostly about process rather than permission. Show me what you asked it. Tell me what you changed. Where do you think it's wrong?
4. Then you sit next to her while she uses it, which is the only supervision that has ever worked for anything.

Steps three and four are the ones everyone wants to start at. They don't hold weight without one and two underneath them.

The honest counterargument

There's a version of this thesis that's wrong, and it's worth naming so you can hold the right one.

Teachers taught internet literacy without being network engineers. They taught kids to evaluate sources without being librarians. Demanding tool mastery as a precondition for teaching has historically been a way of stalling. The dominant framework in this area — TPACK — makes the point directly: technological knowledge on its own doesn't produce good teaching, because it only becomes useful where it intersects with what you know about the subject and about how students learn it. A teacher who deeply understands how a student learns to write, with two hours of AI experience, will probably do better than an AI power user who has never taught.

That's true, and it's why the bar here is forty hours and not a credential. But there's something different about this tool that the internet analogy misses: a search engine returns sources you can evaluate. A language model returns a finished artifact that looks like the student's own thinking. The evaluation problem moved inside the work product. That's a change in kind, not degree, and it's the reason secondhand understanding runs out faster here than it did with the web.

What to do this week

If you accept any of this, the next step isn't a book or a course. It's picking one real task you were going to do anyway and doing it with the model in the room — then noticing, honestly, whether the result was better and whether you thought less.

If you want a structured version of that, our AI stack guide walks parents and teachers through building something actual — a quiz app, a reading tracker, a class site — using tools that didn't need you to learn to code first. It takes an afternoon. It's free. It will teach you more about what these systems can and can't do than a year of reading about them, because you'll be the one holding the thing when it breaks.

If you'd rather have the whole thing in order — the argument, the build, what to read while your opinions are forming, and how to sit down next to your kid afterward — we put it on one page: for grown-ups. It's short, it's free, and there's no certificate at the end, because a certificate would rather badly miss the point.

And when it breaks, you'll finally have something worth telling a kid.