Will my kid actually learn anything from AI tools?

The honest answer is conditional.

Parents Google this question because the marketing copy around kid AI tools is a mess. Some products promise “your kid will learn to code in 20 minutes.” Others promise “AI does the hard part so your kid can focus on creativity.” Neither claim survives contact with how kids actually learn. Both can be true of the same product depending on how the kid uses it.

The honest answer: a kid learns from an AI tool when the AI proposes and the kid decides. They do not learn when the AI executes and the kid receives. The mechanism is forty-five years old. Seymour Papert called it constructionism: knowledge becomes durable when the learner builds something they care about, sees it in front of them, and revises it.¹ Papert was writing about LEGO and turtle graphics, but the principle is the same when the partner is an AI. The kid has to be the one making the choices.

So “will my kid learn?” resolves into a different question: does the tool make the kid choose, or does it choose for them? You can answer this in five minutes by watching them use it. Look for the decision points. If the kid is pressing a button and getting a thing, no learning is happening. If the kid is seeing a proposal, comparing it to what they had, and saying yes or no, learning is happening in front of you.

What “not learning” looks like.

Picture a kid using a typical chatbot to make a story. They type “write me a story about a dragon.” A paragraph comes back. They read it. Maybe they like it. Maybe they ask for another. They paste it into a doc. They are done. Total decisions made: one prompt, maybe a re-roll, maybe a copy-paste.

This is not learning. It is delivery. The kid acquired a story-shaped artifact, but they did not build the understanding that produced it. If you asked them a week later why the dragon is green, or what changes when you make the wizard older, they would not know. The artifact has no felt provenance for them. They cannot revise it because they did not author the choices that made it the way it is.

Albert Bandura’s work on self-efficacy predicts what this kid is left with: nothing.² Self-efficacy is built through what Bandura called mastery experiences: small, successful, attributable acts. Pressing a button on a finished output is not a mastery experience because the kid cannot attribute the success to themselves. The AI did it. The kid watched. Repeat the cycle a hundred times and you have a child who treats AI as a slot machine for finished things.

Carol Dweck’s thirty years of work on mindset says the same thing from a different angle.³ Kids develop a growth mindset by experiencing the loop of effort, feedback, and revision. A finished-answer chatbot strips effort and revision out of the loop. What is left is feedback the kid cannot act on, because there is no version of the work that is theirs to revise.

What learning looks like: Theo’s Achilles.

Theo is nine. Achilles is the second game he shipped in the studio. The full project is live at god-games.vercel.app and you can play it. The interesting thing about Achilles for this question is not the game; it is the change log. Across the project, Theo made thirty AI decisions. Seventeen he kept. Nine he revised. Four he undid.

Hold that distribution in your head. It is not a kid pressing a button and getting a thing. It is a kid spending hours in front of a screen with an AI partner, listening to proposals, looking at what each one would change, and saying yes, sort of, or no. The seventeen keeps are decisions. The nine revises are decisions plus a counter-proposal in the kid’s own words. The four undos are decisions about what does not belong.

The seventeen keeps are decisions. The nine revises are decisions plus a counter-proposal in the kid’s own words. The four undos are decisions about what does not belong. That distribution is the learning. A change log from Achilles, the second game Theo shipped

This is what Papert was describing in 1980. The artifact is the proof of the learning. You can pick it up. You can run it. You can ask the kid why the arrow trap fires when it does, and he can tell you, because he chose to keep that proposal and revise the timing on it. The understanding is durable because it is entangled with the thing he built.

Mitchel Resnick’s Lifelong Kindergarten describes the same loop in different vocabulary.⁴ Resnick names projects, passion, peers, play as the four conditions for kid-creative environments to teach anything. Achilles satisfies all four. It is a project Theo cares about. He is doing it from passion, not assignment. The deployed URL is share-ready, so peers see the work. And he plays his own builds between sessions, which is where most of the revising originates.

Four outcomes you can watch for.

If you want to know whether a kid AI tool is teaching anything, watch for four specific signs during a session. These are not test scores or skill checklists. They are observable behaviors a parent can spot in twenty minutes of looking over a kid’s shoulder.

1. The kid says no to the AI. A learning kid rejects proposals. They do it casually, without ceremony, because they have a felt sense of what fits the project and what does not. A kid who never says no to the AI is treating the AI as a teacher. A kid who says no several times an hour is treating the AI as a collaborator. The studio surfaces this as the Undo button on every change card, and we watch for the press rate in cohort observation.

2. The kid can explain why. Stop them mid-session and ask “why did you keep that one?” A learning kid has a reason, even if it is “because the music was scarier and I wanted scarier.” The reason might be aesthetic, might be functional, might be inconsistent across sessions. What matters is that there is one. Duri Long and Brian Magerko’s 2020 framework on AI literacy lists this capacity to articulate AI decisions as a core competency.⁵ Kids who can name what they kept and why have started building a working model of how AI behaves.

3. The kid revises their own work. The next session, look at what they touch first. A learning kid often returns to something they shipped last time and changes it. Not because anyone asked them to. Because they thought about it overnight and want it different. This is the iteration habit forming. Dweck’s growth-mindset research finds that kids who experience revision as a normal part of the work, rather than as evidence of failure, develop more durable confidence in unfamiliar domains.³

4. The kid teaches someone else. The strongest signal. A kid who has actually learned the loop will, unprompted, show a sibling or a parent the move they figured out. They are not parroting Inkie. They are explaining the choice they made and why it worked. Papert called this the moment the artifact becomes thinkable — the kid can hold their own work as an object of reasoning, not just an output of activity.¹

A test you can run at home.

Here is the smallest test for whether any AI tool is teaching your kid. Sit next to them for ten minutes. Count two numbers. How many proposals did the AI make? How many decisions did the kid make? If the ratio is roughly one to one, the tool is teaching. If the AI made ten proposals and the kid pressed one button to accept all of them, the tool is not teaching. It is delivering.

In a typical Tell and Show session, the ratio is closer to one decision per proposal. Sometimes higher: a kid will look at a proposed change, ask Inkie to try it differently, look at the new version, ask for one more variation, and then keep version three. That is three decisions on one proposal. Across an hour, a kid might make sixty to a hundred decisions of this shape. The decision count is the learning rate.

This is also the test you can run on the studio itself. Open Theo’s game in one tab. Open the gallery walkthrough in another. Watch the change log. The ratio is what we are designed to produce. Every wizard renders as a proposal card. Every chat reply renders as a proposal, not an answer. Every change ships through the ChangeDisclosure card with Keep, Review, and Undo buttons. The architecture is designed to keep the kid in the decision seat.

So the answer to “will my kid actually learn anything from AI tools?” is yes, conditional on what kind of tool, used how. The studio is one example of a tool designed for the yes side of that condition. There are others. The thing to look for is whether the kid is choosing, every minute, what to keep. If they are, they are learning. If they are not, they are watching a slot machine.