What kids learn when they make a game (not just play one).

Twenty years of research, in one paragraph.

Yasmin Kafai started studying what happens when kids design video games in the early 1990s, as a graduate student at the MIT Media Lab under Seymour Papert. The work moved with her to UCLA and then to the University of Pennsylvania’s Graduate School of Education, where her lab has been running studies on game-making and learning continuously since. In 2015 she and Quinn Burke synthesized roughly two decades of that work into a survey paper in Educational Psychologist, titled Constructionist Gaming: Understanding the Benefits of Making Games for Learning.¹ They expanded the argument into a book the next year, Connected Gaming, from MIT Press.²

The synthesis is careful and specific. It does not claim that designing games turns kids into engineers, that it raises test scores, or that any single intervention generalizes neatly. What it does claim, and document across many studies, is that kids who design games develop a recognizable cluster of capabilities that kids who only play games don’t. The capabilities are partly cognitive, partly social, and partly about identity. Kafai and Burke call the broader frame constructionist gaming, an explicit descendant of Papert’s constructionism.³

The research base predates the AI moment by a long way. That matters. It means the pedagogical case for kid-designed games rests on findings collected when the only authoring tools were Logo, Scratch, GameMaker, and similar. AI partners are new. The reason game-making teaches what it teaches is not.

What game-making teaches that game-playing doesn’t.

The findings sort into four buckets. None of them are mysterious. All of them are hard to fake.

Systems thinking. Every game is a small system. Rules, state, feedback loops, win conditions. A kid playing a game operates inside the system; a kid designing one has to hold the whole thing in their head at once. What state changes when the player jumps? What happens when the enemy reaches the edge of the platform? When does the level end? Kafai and Burke document this shift across multiple studies as a move from procedural fluency to systems fluency.¹ The kid stops asking “how do I beat this” and starts asking “how does this work.”

Audience reasoning. A designer has to imagine someone else playing. Will my friend understand the controls without me explaining? Will this puzzle feel fair, or just annoying? Game-making is empathy practice with a deadline: the playtest. Kafai’s earlier studies on kids designing games for younger students at the same school showed this audience-orientation emerging clearly, and showed it transferring beyond the game.⁴

Failure modeling. A kid-player who gets stuck quits or asks for a walkthrough. A kid-designer who watches a friend get stuck has a different problem. They have to ask why. Is the puzzle confusing or just hard? Is the difficulty curve broken? Is the feedback wrong? Salen and Zimmerman’s formal game-design pedagogy in Rules of Play frames this as the distinction between meaningful and frustrating difficulty, and treats the ability to tell them apart as the central craft skill of the discipline.⁵ Kid-designers learn to make that distinction earlier than most adults using AI tools today.

Identity-as-maker. This is the finding Kafai writes about most directly. Across her longitudinal work, kids who design games over time start to describe themselves differently. They’re not kids who play games; they’re kids who make them. The shift is small and durable, and it transfers to other computational domains. A kid who has shipped a game is a kid who knows, in a way that’s hard to undo, that software is a thing people make.²

Why the line between playing and making matters more with AI.

The playing-vs-making distinction was an interesting research question in 2005. It’s a load-bearing one in 2026. Here’s why.

Generative AI can produce playable games on demand. A kid with a chatbot and an hour can have something runnable. So the act of playing a novel game is no longer scarce. What’s scarce, and what AI cannot yet do well, is the act of designing one. Articulating, at nine years old, “I want the enemies to feel scary but not unfair” is the thing a kid does that the model can’t. The model can implement a hundred variants of scary-but-fair. It cannot decide which one is the game.

This is what Papert was reaching for in 1980 when he called the computer an “object to think with.” The point of a maker tool isn’t to produce more software in the world. It’s to give the kid a medium they can press their ideas into and watch what comes back.³ A kid who plays a hundred AI-generated games has watched a hundred ideas come back. They were just someone else’s ideas.

Mitchel Resnick’s long-running work at the MIT Media Lab’s Lifelong Kindergarten group reaches the same conclusion from a different angle: the creative tools that build durable learning are the ones with a low floor and a high ceiling, where the kid is the source of the project.⁶ AI that lowers the floor without changing who’s deciding is constructionist. AI that delivers a finished game on request is not.

The model can implement a hundred variants of scary-but-fair. It cannot decide which one is the game. On what stays scarce after AI

What this looks like in the Game track.

The Game track is a constructionist gaming environment with a modern partner inside it. The loop is the same one every kid-designer has run for forty years, with the friction in different places. A kid picks a genre. Defines the rules. Prompts Inkie for a specific mechanic. Playtests the result. Keeps the parts that work, revises the parts that don’t, undoes the parts that broke the game. Then does it again.

The four competencies show up because the loop forces them to. Systems thinking shows up when the kid has to name the change before Inkie can make it. Audience reasoning shows up the first time they hand the game to a sibling who doesn’t know the controls. Failure modeling shows up when they playtest their own work and have to decide why the third level isn’t fun. Identity-as-maker shows up the first time the URL works on someone else’s phone.

Theo’s Achilles is the canonical example we point parents at, because the loop is fully visible in the artifact. The decision tally for the project is observable: 30 AI proposals reviewed across the build, 17 of them kept, 9 revised, 4 undone. Every one of those decisions is a moment where a 9-year-old was doing the work AI can’t do.

The numbers are not the point. The shape of them is. A kid running through a constructionist-gaming loop is making roughly that many small calls per project. They’re the calls Kafai and Burke would recognize from the studies they ran in the 1990s with Scratch and GameMaker, just with the tedium of typing factored out.¹

The four other game projects in the Game track gallery show the same loop on different surface areas. A platformer is a different system than a top-down shooter; the systems-thinking work transfers anyway. That transfer is what Kafai’s longitudinal work documents most clearly.²

Playing-vs-making is a real dichotomy. The studio is built around it. The short version of the difference looks like this:

That last row is the one parents notice first. The first row is the one that compounds.

What parents should watch for.

None of the four competencies are testable in any easy way. They show up in how a kid talks about their own work. Four signals, in roughly the order they appear.

The kid can name the design choice they made. Not “I made the boss harder.” Something closer to “I made the boss harder by giving him a second attack but a longer wind-up, so it’s fair.” That sentence is systems thinking and failure modeling running together. The vocabulary will be the kid’s, not a designer’s. The structure of the thought is what to listen for.

The kid revises based on what they saw. They played the third level themselves, saw the player keep dying at the same jump, and changed the jump distance. That sequence (test, observe, change) is the loop Kafai’s lab has spent twenty years documenting. If the kid is doing it without being told to, the loop has taken.

The kid can explain why their game is hard in the way it’s hard. There’s a difference between hard because the controls are confusing and hard because the boss is supposed to be a boss. A kid-designer can tell you which one their game is doing, and which one the player is reacting to. A kid who’s only played games can’t.

The kid shows their friend the game and points at a specific moment. Not “look at my game.” “Watch this part. This is the part where the arrows feel scary.” That sentence is identity-as-maker. It says: this thing is mine, I made it, I know which parts are good. Kafai’s longitudinal work shows that sentence, said often enough, changes who the kid thinks they are.²

None of these require AI. They require a kid building something, watching it run, and deciding what’s next. AI just changes how far the kid can reach. The thing being learned is the thing Kafai and Burke have been writing about since the kid’s parents were in high school. Now there’s a partner inside the loop who can keep up.