An AI Tutor Is Now a Student Class Project. So What Should Kids Actually Learn?

What this is: a look at what happens to "learn to use an AI tutor" as a goal once building an AI tutor stops being hard — and an argument that the more durable thing to teach kids is the set of skills that don't commoditize: noticing what's true, belonging to people who care, and directing AI systems toward something they actually want.

The two-week tutor

Here is a data point from this year: in Stanford's Lean LaunchPad course, taught by Jennifer Carolan of Reach Capital and Steve Blank, student teams had working prototypes by week two, and one education team put its AI tutor on the leaderboard of Scale AI's TutorBench evaluation within a few weeks. (Carolan's account, as reported by Edtech Insiders in August 2026.) Being on a leaderboard is not the same as being a good tutor, and the same newsletter notes that new tutoring benchmarks keep finding gaps in every model's ability to teach over time. But the direction of travel is not in dispute. Building "an AI that can explain a concept, check understanding, and adapt to a student's level" has gone from a research problem to a class project. The primitives — a capable base model, a prompt scaffold, a retrieval layer over a curriculum, a feedback loop — are now off-the-shelf.

That's worth sitting with, because a huge amount of the "AI in education" conversation over the past three years has been organized around the premise that delivering instruction well is the hard, valuable thing to build. Khan Academy invested years in Khanmigo specifically to get tutoring right. If a two-week class project can now clear a competent tutoring bar, the premise needs updating: delivering instruction is becoming a commodity. Not free, not solved for every subject or every kid, but no longer scarce in the way it was.

So the actual question for parents, teachers, and product people isn't "how do we get kids a good AI tutor." It's: once every kid effectively has access to one, what should they spend their scarce human attention learning instead?

What doesn't commoditize

Three things, we'd argue, stay scarce even after tutoring gets automated.

Noticing. An AI tutor can tell a kid whether their answer is right. It's much worse at teaching a kid to notice when they don't actually understand something — the gap between "I can follow this explanation" and "I could produce this from scratch tomorrow." That metacognitive noticing is exactly what the OECD's 2026 Digital Education Outlook flags as the failure mode of ungoverned GenAI use in classrooms: students who outsource a task to AI without pedagogical scaffolding produce higher-quality work in the moment, but the advantage evaporates — and sometimes reverses — once the AI is taken away. In a field experiment across Türkiye high schools cited in the report, students using GenAI scored roughly 127% higher on practice problems, then scored about 17% worse than non-AI-assisted peers on the unassisted exam. The tutor delivered instruction fine. The student didn't notice they hadn't learned it.

Belonging. A tutor bot can be endlessly patient and always available, and that's genuinely useful. It cannot be a coach who's known a kid since second grade, a parent who notices they've gone quiet at dinner, or a friend who's proud of them for finishing something hard. Belonging is a human resource, and it's the thing that makes a kid willing to be bad at something in front of someone else — which is most of what learning actually requires. No tutoring product, however good, replaces that; the best ones (Khanmigo included) are explicit that they're a supplement to a teacher, not a substitute for one.

Directing. This is the one most directly displaced by cheap tutoring, and the most trainable. If an AI can explain any concept on demand, the valuable human skill shifts from "absorbing an explanation" to "knowing what to ask for, evaluating whether the answer is actually good, and iterating toward something better." That's a different muscle than being tutored. It's closer to being an editor, a director, or a product manager than being a student in the traditional sense — and it's the skill that transfers regardless of which AI model exists next year.

The builders learned more than the tutor would have taught

Here's the part of the two-week-tutor story that's easy to skip past: the students who built the tutor almost certainly learned more from building it than any student would have learned from being tutored by it.

Building even a simple adaptive tutor forces you to decide what "understanding a concept" means well enough to detect it in a transcript. It forces you to notice where your first version fails, argue about why, and revise. It forces you to evaluate whether the thing you made is actually good, not just whether it runs. None of that happens if you're on the receiving end of the finished product — you just get explanations, on demand, forever.

This is the create-versus-consume distinction in its starkest form. A finished AI tutor is a consumption product: excellent for filling a specific knowledge gap, low-leverage for building the underlying skill of directing AI well. The act of building one — even a rough, unpolished one — is a creation exercise that happens to teach noticing and directing as side effects of the work. The tutor is the output. The learning was in making it.

That's not an argument that every kid should build a tutor. Most won't, and most don't need to. It's an argument that the shape of the activity — describe intent, get a real result, evaluate it critically, iterate based on specific feedback — is the transferable thing, and it's available in far smaller, far more kid-appropriate forms than "build a benchmark-passing tutor." A 9-year-old directing an AI to build a working game about their favorite animal is running the same loop at a scale that fits their attention span and their stakes.

Districts are already reacting to the consumption side

It's worth noting that school systems are responding to exactly this risk right now, from the opposite direction. In September 2026, New York City's Department of Education announced a one-year moratorium on student-facing generative AI for roughly 600,000 students in grades K through 8, alongside screen-time guidance of 30 minutes a day for grades 3–5 and 45 minutes for grades 6–8. Days later, the Los Angeles Unified School District blocked generative AI on all district-owned student devices across roughly 378,000 students of all ages, with recommendations for a path forward due by the end of the 2026–27 school year.

Neither district has published a full rationale distinguishing consumption uses from creation uses of GenAI, and it would be unfair to read more into these moves than a broad, precautionary pause. But the underlying concern lines up with the OECD's finding: unstructured, teacher-unsupervised student access to general-purpose GenAI in a school building carries a real outsourcing risk, and districts serving hundreds of thousands of kids are choosing to slow down rather than sort that out live. That's a reasonable, defensible call for a district weighing tens of thousands of devices and a small tech team. It's a different calculation than the one a parent makes at home, deciding whether their own kid, with visibility into every prompt, should get to direct AI as a creative tool.

What this means practically

If you're a parent or teacher trying to make sense of the current AI-for-kids market, one useful filter is: does this tool ask the kid to reflect, iterate, and direct — or does it just deliver an answer?

A tutoring product that explains a concept and checks comprehension is valuable and worth using, but it's a consumption tool. Treat it the way you'd treat a good textbook with a tireless explainer inside it — helpful for closing a specific gap, not a substitute for a kid making something.
A coding platform like Scratch or Tynker builds real skill, but it's pre-generative — kids assemble blocks, they don't direct an AI. Different, complementary skill.
A classroom AI tool built for teacher workflows (lesson plans, grading, differentiation) is solving the teacher's problem, not the kid's. Worth having in a school, not a substitute for a kid-facing creative tool.
A generative tool where a kid describes what they want and an AI builds it, and the kid iterates — that's the create-not-consume category, and it's much rarer than it sounds, because most generative AI products either bar minors under 13 outright or require a school and teacher to sit in the loop for any kid under 13.

Where Xyplor sits in this, honestly

Xyplor is built around the create-not-consume bet: kids ages 6–17 describe what they want, get a real working game, quiz, podcast, or website in about 60 seconds, then iterate by giving specific feedback — "the dragons need to be faster," "add a final boss." Every creation ends with a short reflection on what the AI actually did and a concrete next step, because the magic of instant output doesn't teach anything on its own; the loop of seeing a result, understanding it, and trying again does.

We're not a tutor, and we don't try to be one — see our comparison with Khanmigo for where a dedicated tutoring product is genuinely the better pick. We also don't claim Xyplor teaches programming; it teaches AI direction, which is a narrower and, we think, more durable skill given how fast the underlying tools are changing. And we don't have outcome data yet showing that kids who use Xyplor get better at noticing or directing over time — that's a claim we'd need a longitudinal study to make honestly, and we don't have one.

What we do think is defensible: if delivering instruction is heading toward a commodity, the scarcer and more valuable thing to hand a kid isn't another explanation. It's the chance to build something, watch it not quite work, and figure out — with a person who cares about them nearby — what to try next.