Andrej Karpathy told a school board in November 2025 that the war on AI homework is over. His exact words: “You will never be able to detect the use of AI in homework. Full stop.” Detectors don’t work, can be defeated, and are in principle doomed. His advice was to stop trying. Flip the classroom. Practice with AI at home, assess in person.
He’s right about the detection problem. But the detection problem is not the actual problem. The actual problem is older, and AI is just the thing that finally made it visible.
The proxy was always thin
For most of the history of schooling, we have not really been measuring intelligence. We have been measuring artifact production. The essay, the worksheet, the lab report, the rendered diagram, the formatted output. Those are the things teachers can collect at the end of a period and grade in a stack. We trusted them as evidence of the underlying capability because we had no cheaper way to look at the underlying capability directly.
That was a budget decision, not an epistemic one. Evaluating someone’s reasoning takes a panel of experts and a multi-hour conversation. Evaluating their five-paragraph essay takes one person with a rubric and twenty minutes. So we built an entire credentialing apparatus on the worksheet, and we let the worksheet stand in for the thing we actually cared about.
The thing we actually cared about is harder to name and much harder to measure. Executive function, the ability to take a problem apart, taste, judgment, the capacity to envision something that doesn’t exist and figure out the steps to bring it into being. None of that produces a clean artifact you can score in a stack.
So we stopped trying to grade it directly, and let the artifact serve as the proxy.
The defense is the proof we already know how to do this
We do know how to measure the real thing. We just only do it at the very end of a long credentialing pipeline, after we’ve already filtered tens of thousands of people down to a few dozen who are worth the cost.
The dissertation defense is the proof. A candidate spends years assembling a body of work. They stand in front of a panel of people who actually know the field, and the panel picks the work apart in real time. The candidate has to defend choices, reconstruct reasoning under pressure, respond to objections that nobody scripted. We trust the outputs of that process. Nobody worries about whether the defense is gameable.
That process works. We just can’t afford to do it on every fourth-grader. The teacher of a fourth-grade class doesn’t have the time, doesn’t have the budget for an expert panel, and frankly, in many cases, doesn’t have the domain expertise to ask the kind of follow-up question that would surface real reasoning. The cheap test exists because the expensive test was unaffordable. Not because the cheap test was a good test.
What AI changes
Once a machine can produce the artifact, the artifact is no longer evidence of the underlying capability. The five-paragraph essay used to be a noisy but somewhat-useful proxy for whether a student could think. Now it’s noise without the signal. Everyone can produce the artifact. Some of those people had something to say. Most of them didn’t. And the worksheet can no longer tell us which is which.
The institutions that built their business model on sorting smart people from less-smart people are panicking, and the panic reads on the surface as a panic about cheating. It is not. It is a panic about the realization that the sorting mechanism was always doing less work than they advertised. AI didn’t break the sorter. AI revealed that the sorter was already broken.
The honest version of what’s happening: we are being forced, finally, to grade the expensive thing. The cheap proxy has collapsed and the only options are to find a new cheap proxy (which will collapse again as soon as the next capability lands) or to actually invest in measuring the underlying capability.
The kid with the bees
Here is the part of this that gets me up in the morning.
Imagine a fourth-grader with a real idea. A kid who has been watching the bees in her grandmother’s garden, has noticed they treat certain flowers differently depending on the time of day, and has a working theory about why. She doesn’t have the biology vocabulary. She doesn’t have the statistics. She doesn’t have the prose chops to write a coherent five-paragraph essay about it, and she doesn’t have the drafting skill to diagram her observations so an adult can look at them and nod.
Under the old regime, that kid was invisible. Not because she wasn’t smart. Because the only channel we’d built for her to demonstrate what she was capable of was a channel that required a whole separate set of skills (writing, drawing, formatting, citation) that we had no business expecting her to have yet. She failed the artifact test. We concluded she failed the intelligence test. Those are not the same test, but the school system treated them as the same test because it couldn’t afford to administer the other one.
Now that kid can sit down with an AI and produce the artifact. Not as a cheat. As a translation layer. The observations are hers. The decomposition is hers. The taste about which paths are interesting and which are dead ends is hers. The writing and the rendering and the formatting, which were never the point, are no longer the bottleneck.
The flip side is also real. The kid who used to coast on neat handwriting and a tidy paragraph but had nothing underneath is no longer indistinguishable from the kid with the real idea. The cheap signal is gone. The noise is exposed.
This is the part the panicking institutions don’t seem to want to say out loud. The democratization of artifact production isn’t only lowering the floor. It’s also raising the question of whether the floor was ever measuring the right thing in the first place.
What we should actually build
Karpathy’s flipped-classroom answer is one cut at the problem, and it’s a reasonable one. Move the practice to where AI is allowed and useful. Move the assessment to where you can look someone in the face and ask follow-up questions they didn’t script for.
There are other cuts. Apprenticeship models, where you watch someone work over weeks rather than scoring a single artifact. Project-defense models, where the student presents the work and a panel asks where the choices came from. AI-assisted oral assessment, where the proctor uses the student’s submitted work as a starting point and probes the reasoning behind it.
The unifying idea is that the human has to come back into the loop on the thing that matters, which is the reasoning, not the output. We had outsourced that to the artifact because the artifact was cheap and the conversation was expensive. The arithmetic has changed. The conversation is still expensive to run at human-only scale, but AI is now able to assist the conversation. It can ask follow-up questions, propose objections, summarize what it heard, surface inconsistencies for a human evaluator to probe further.
That is, ironically, the most interesting application of AI in education. Not as the thing students cheat with. As the thing that finally makes the expensive form of assessment affordable enough to use on more than a handful of dissertation candidates per institution per year.
The bill is due
The fact that AI is forcing this conversation isn’t a tragedy. It’s an audit. The proxies we’d been using to identify smart people were always thin. We knew they were thin. We tolerated the thinness because the alternative was unaffordable. Now the proxies are gone, and the alternative is, for the first time, within reach.
What scares the establishment isn’t that students are cheating. It’s that the cheap version of the test no longer separates the signal from the noise, and the institutions don’t yet have a story for how to grade the expensive version. That story is what the next decade of education work is actually about. The detection problem is a distraction. The measurement problem is the one to solve.
This is not only a schooling problem. Any organization that judges technical work by how finished the deliverable looks is running the same broken test. Once a model can produce a clean report or a well-formatted proposal on demand, the document stops being proof that the reasoning behind it is sound. For a science or engineering firm leaning on AI-assisted output, the useful question is no longer whether the artifact looks right. It is whether the person who signed it can defend the choices and trace where the numbers came from when someone asks the follow-up nobody scripted. That is the expensive test, and it is the one worth building a process around.