Cheating with AI has stopped being a shortcut. A New York University student built a bot that does his calculus homework at the pace of a student who is falling behind, so that the timestamps his professor checks look plausible. A sophomore at UMass Amherst spent four hours degrading a slide deck an OpenAI coding agent had produced, because it looked too professional to pass as his. New York Magazine interviewed several students doing this kind of work. The labor did not disappear when the models arrived. It moved from doing the assignment to concealing that a machine did it.
The NYU student, Jared, treated WebAssign as a challenge. The platform his introductory calculus course used is built specifically to make copying harder. He and a friend used an AI coding tool to write a bot that completes all their homework and does it slowly, at the rhythm of a struggling student, so a professor scanning submission times sees nothing unusual. Jared then handed the tool to friends, and says he no longer knows how many of his classmates are running his automated cheating system.
That last detail is the one worth sitting with. Cheating has historically scaled one student at a time, through a tutor, a group chat, a shared answer key. Software scales by copying. It is also a working demonstration of what AI agents are for: models that execute complex tasks on a person's behalf without human intervention, capable of writing entire programs and conducting research. Here the capability has been pointed at a submission scheduler.
The students frame this as a response to how their institutions reacted to mass AI cheating. Will, a recent graduate of Carleton College, argues that professors who fixate on the obvious tells of machine-written text — the long dash being the canonical example — are telling students precisely what to hide. Students will find ways to mask those stylistic markers, he says, and if the only anti-cheating measure is an announcement that a detector exists, it will not work.
Will's own professors banned AI-generated text. One of them boasted of having an entire "battery" of tools to enforce the ban. Will used AI anyway: he had Claude analyze his source material and write large sections of his essays, then rewrote the model's paragraphs by hand.
The most elaborate case in the piece is Theo's. A sophomore at UMass Amherst who describes himself as self-taught and has no interest in his general-education courses, he assembled an astronomy presentation using OpenAI's Codex coding agent together with a plugin that improved its web search. The finished deck looked too professional. So he spent four hours manually making the slides worse, until they resembled the work of a mediocre second-year. It took significant effort, he said.
Four hours of deliberate degradation is not obviously less work than building a mediocre deck from scratch. What these setups optimize is not effort but the ratio of effort to grade, and the certainty of the result. That trade only pencils out if the assignment is worth nothing to the person completing it, which is roughly what Theo says about his general-education requirements in the first place. Read that way, each of these contraptions is a price signal about the coursework rather than a verdict on the student.
There is a harder version of the same observation. What Jared, Will and Theo have actually practiced is agent orchestration, output laundering and the management of a system that has to behave believably under inspection. Those are skills with a labor market behind them, acquired inside courses designed to teach something else. The transcript will record calculus.
Danielle Carr, a historian and anthropologist at UCLA, argues that instructors who rely on observation and detection tools will simply be stuck in an endless game of cat and mouse. AI cheating, she says, is not a technical problem that can be fixed with a technical solution.
Nothing in the account describes a consequence. No student is caught, no bot is found, no professor changes what is assigned rather than how it is policed. The detection regime has produced undergraduates who write and distribute software to satisfy an institution that is still grading the artifact rather than the person who submitted it.