The numbers on AI in education have stopped being a debate and become a fait accompli. By 2026, around 92% of students use AI, and 88% admit to using it on graded assignments. Here is the kicker that makes detection policies look like theatre: an estimated 94% of AI-generated assignments go undetected, even though student discipline for AI-related misconduct rose 33% between 2022 and 2026. Schools are catching more cheats than ever and still missing almost all of them.
The honest, uncomfortable picture
What makes this genuinely thorny is the students’ own confusion about it. Nearly 80% say they believe using a large language model on schoolwork is a form of cheating, and yet over 46% do it on graded college work anyway. That gap, between knowing it is wrong and doing it because everyone else is and nobody gets caught, is the real story. It is not that students think AI is fine; it is that the incentives have collapsed. When detection fails 94% of the time and your classmates are all using it, abstaining just means a worse grade for your honesty.
And the detection arms race is unwinnable in its current form. AI-writing detectors are unreliable, prone to false positives that punish the innocent, and easily defeated by a student who lightly rewrites the output. Banning AI outright, meanwhile, runs into a different wall: more than a third of entry-level jobs now list AI skills, so a school that successfully stops its students touching AI is arguably failing to prepare them for the actual workforce. Block it and you handicap them; allow it and you cannot grade them the old way.
What this means
The uncomfortable conclusion is that the essay-you-write-alone-at-home, graded-on-the-finished-text model of education is broken, and no detector is going to fix it. The schools navigating this least badly are the ones changing what they assess: more in-class writing, oral defences, process over polished product, and teaching students to use AI well rather than pretending they will not use it at all. The technology did not create a cheating problem so much as expose that we were grading the artefact instead of the learning. That was always the flaw; AI just made it impossible to ignore.
Related: the same reckoning is hitting hiring.