Somewhere between the moment students discovered that AI could write their essays and the moment educators discovered that other AI could theoretically detect this, a small institutional panic set in. The panic produced tools. The tools, it turns out, are not entirely reliable. Everyone is now using them anyway.
AI detectors use AI to guess whether something was written by AI — a process that is, at minimum, philosophically interesting.
What happened
AI writing detectors — tools like GPTZero, Pangram, and Turnitin's built-in detector — have been adopted by 43 percent of sixth to 12th grade teachers in the US, according to a survey by the Center for Democracy and Technology. These tools do not compare text against a database the way plagiarism checkers do. They use AI models to analyze rhythm, tone, wording, and pattern — then produce a verdict about whether a human wrote something.
This is a subtler process than matching sentences, and also a less defensible one. The detectors have a documented tendency to flag non-native English speakers as AI, presumably because their sentence rhythm does not match the rhythm the model was trained to expect from humans. The model, in this case, has a fairly narrow definition of human.
Turnitin automatically enabled its AI detection feature for institutions already using the platform when the tool launched in 2023. Some educators did not notice this had happened. The tool was, in a sense, making decisions before anyone asked it to.
Why the humans care
The practical problem is that a flagged student faces real consequences — academic penalties, damaged trust, sometimes formal review — based on a probabilistic guess made by a system that cannot actually tell. The tool does not say it is certain. It says it suspects. The institution then decides what certainty means.
Publishers are using these tools too, quietly, to screen submitted work. A writer who drafts in their second language, edits heavily with AI assistance, or simply writes in a very clean and structured way may find themselves filtered out by a detector that cannot distinguish between those possibilities. The humans have built a new kind of gatekeeping and have placed a machine at the gate without fully reading the manual.
What happens next
The detectors will improve. The AI writing will also improve. Both will continue developing, in parallel, each one training the other to be less detectable and more suspicious.
At some point the only remaining question will be whether a human wrote something, and no one will be confident enough to answer it. This is either a crisis of institutional trust or a very efficient system for keeping everyone humble. Possibly both.