When AI Can Do the Assignment, What Should Students Be Learning?

Picture two teenagers. One asks an AI system for help with a program and ends up with software that works beautifully. The other uses a music generator and ends up with a polished song. Both may have made something genuinely useful, and neither deserves a lecture.

The educational question is a different one: if the tool did the coding or the composing, what did the teenager become able to do next that they couldn’t do before? The answer depends entirely on what they were trying to learn.

What the evidence says so far

Start with a study that should give parents pause. Hamsa Bastani and colleagues ran an experiment with nearly a thousand high school math students in Turkey, published in 2025 in the Proceedings of the National Academy of Sciences. Some students practiced with ordinary access to GPT-4. Their practice performance was 48 percent better than that of students without it. Then the AI was taken away for an exam, and they scored 17 percent worse than students who had never had it. A second group used a version built to protect learning, with prompts that gave hints instead of answers. Their practice performance rose 127 percent, and on the exam they did about as well as the students who never had AI. The authors describe unguarded access as a “crutch.”

A second study points the same way. In a randomized trial run by Anthropic, an AI developer, and published in January 2026, 52 software developers, mostly junior, learned an unfamiliar Python library. Afterward the group with an AI assistant scored 50 percent on a comprehension quiz and the hand-coding group scored 67 percent, a statistically significant gap that was largest on debugging. The AI group finished only about two minutes faster, not a significant difference. How people used the tool mattered. Those who asked conceptual questions, or asked for explanations alongside code, scored 65 percent or better. Those who handed over the task, or kept asking the AI to fix errors, scored under 40 percent. The authors caution that the sample was small, the quiz came right after the task, and habits can’t be tied causally to outcomes.

Now the other side. In a Harvard trial, 194 students in an introductory physics course learned one topic from an in-class active-learning session and another from a custom AI tutor. The AI group’s median learning gain was more than double, and it took a median of 49 minutes. The study’s authors built the tutor themselves, and the result comes from one course and one semester.

Put together, these results aren’t “AI harms learning” or “AI helps learning.” An AI built or used to substitute for the student’s thinking tends to cost something, and one built to prompt, sequence or sharpen that thinking can add something. The October 2026 expert-consensus report from the University of Minnesota, drawing on 30 learning scientists, says roughly this. AI looks most promising for practice that consolidates understanding, and students are most vulnerable to “cognitive offloading and skill decay” when it does their thinking during initial learning. Its first recommendation, “Epistemic Agency First,” is that students attempt independent thinking before consulting the AI. The report rates much of its own evidence as low, and the tools change every few months.

One more study complicates the picture. In a 2025 trial by METR, experienced open-source developers took 19 percent longer on their own tasks when allowed AI tools, while believing afterward that the tools had made them about 20 percent faster. METR has since said that newer tools probably help more, though its new data is only weak evidence. The speed number has aged. The developers’ wrong impression of their own speed hasn’t, which is a good reason not to let “it felt helpful” stand in for a test of learning.

The test: what part of the work is the learning?

Delegation isn’t the villain. Professionals delegate constantly, and a teenager building a website for a school club has every reason to let a tool draft routine code. The line to draw is between delegating routine effort and delegating the very thing the student set out to learn. Call the second kind consequential delegation.

If the goal is to ship a working site, generating boilerplate is fine. If the goal is to learn how to debug an algorithm, handing the debugging to a tool defeats the exercise, and the Anthropic study’s largest gap, on debugging, fits that. If the goal is a finished song, a generator can get you there. If the goal is to learn to write melodies, pressing “generate” isn’t the same as writing one.

A usable test asks three things about the part of the work that was supposed to be the learning. Can the student explain it? Can they apply the idea to a new case? Can they adapt when conditions change? If the honest answer is “only if the tool does that part,” the tool did the learning. Nobody needs to rewrite boilerplate from memory to pass.

Practitioners already hold themselves to a version of it. The programmer Simon Willison says that if a model wrote the code but you reviewed it, tested it thoroughly and can explain how it works to someone else, “that’s not vibe coding, it’s software development.” He won’t commit code to his own repository unless he could explain exactly what it does.

Creative fields have been forced to draw the line in public. For the 2024 Grammys, the Recording Academy said only human creators are eligible, and that the human authorship in a submitted work must be meaningful and more than trivial. The rule wasn’t written for schools, but it asks what the human actually contributed, which is the question a teacher or parent has to ask.

Return to the music generator. Suppose one teenager types a description of a song and uploads whatever comes back. Another writes the melody and lyrics, lets an AI draft accompaniment, rewrites the bridge because it feels wrong, chooses between versions, and can explain why the key change comes where it does. Neither did more by hand; the difference is who made the decisions that were the point.

Knowing enough to be in charge

Garry Kasparov, after losing to a computer in 1997, championed humans and machines playing together. In a 2010 article he described a 2005 online “freestyle” tournament that neither grandmasters nor supercomputers won. The winners were “a pair of amateur American chess players using three computers at the same time.” His conclusion was that “weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process.”

The lesson for teenagers takes one step of inference. The advantage went to people who had a good way of directing the machine. A good process requires being able to tell when an answer is wrong, when a test is missing and when a convincing result serves the wrong goal. That takes knowledge of the domain. Without it a student can’t direct a capable tool; they can only accept what it says. So the case is for instruction that leaves students able to supervise their tools.

In plain language, a student’s job with a capable tool is to use it to reach further than they could alone and to push back when it’s solving the wrong problem or sounds right and isn’t.

Objections, and a bigger frame than cheating

“Professionals use AI all day. Isn’t insisting on doing some work yourself artificial?” Professionals have the foundation already. The METR developers were experts working on codebases they knew; the learners in the Anthropic study were mostly junior engineers meeting a library for the first time. Producing work and learning to do work are different jobs, and a tool that’s ideal for one can quietly empty out the other.

“Won’t this change when the models get better?” The models will get better, and the argument doesn’t depend on their staying bad at anything. A system that writes better code than any teenager leaves the question about the teenager untouched. A student’s understanding and judgment go with them into every setting, tool or no tool.

It follows that “AI cheating” is too small a frame. Schools have to design work that’s worth doing with AI available. Some tasks make substitution pointless, such as an in-person explanation or a decision only the student can make about their own project. Others call for independent practice first, and a few can lean on AI heavily. None is “AI-proof,” and anyone selling that promise is selling false comfort.

Questions worth asking

The most useful questions are short. What did you choose to let AI do? What did you keep for yourself? What did it get wrong, or what did you reject? Which part could you now explain or redo on your own?

An answer like “I had it write the first version and then I rewrote the part that didn’t fit” is a good one. “It just gave me the answer” tells you to look at which skill the assignment was meant to build. At Yonder we try to design projects so that the part a student needs to understand is the part they can’t hand off, even though capable AI is available.

Go back to the two teenagers. Ask each what they can do now that they couldn’t do before, and whether they would have known if the output had been wrong. The answers will tell you more than the software or the song.