OpenAI launched an AI text classifier in January 2023 and shut it down that July, citing a low rate of accuracy. By the company’s own numbers, it correctly flagged only 26% of AI-written text as likely AI-generated, while labeling 9% of human writing the same way. The company that built ChatGPT couldn’t reliably spot ChatGPT.
The third-party detectors sold to schools have not become trustworthy since. They’ve just become more confident-looking, with tidy dashboards and percentage scores that suggest a precision the underlying method doesn’t have.
The false-positive problem is not evenly distributed
A 2023 Stanford study found that several popular detectors misclassified more than half of a set of TOEFL essays, written by non-native English speakers, as AI-generated. The researchers’ explanation was that detectors lean on how predictable a text is. Writers working in a second language often use simpler, more common constructions, which is exactly the pattern a detector reads as machine-made.
So the students most likely to be wrongly accused are the ones with the least room to argue their case. An accusation backed by software also tends to shift the burden of proof onto the student, who has to prove a negative.

A score is not a measurement
A detector’s “82% likely AI” isn’t a reading from an instrument. It’s a statistical guess, and presenting it as a number makes it feel like a fact.
What actually works is less convenient
The fixes that hold up all involve more teacher time, which is why they’re unpopular. In-class writing, even short and ungraded, gives an instructor a baseline of how a student really writes. Requiring drafts, or checking a document’s version history, shows process in a way a finished essay can’t. A five-minute conversation about the argument someone supposedly wrote usually reveals more than any scan will.
None of that scales like pasting text into a box, and that’s the real appeal of detectors. They promise to make an assessment problem into a software problem. But the problem is that take-home essays were quietly doing two jobs, teaching writing and certifying authorship, and a chatbot broke the second one.
Schools that keep leaning on detectors are choosing the cheaper option, and the cost lands on students rather than on the institution. Some of the better-run programs have already stopped relying on the scores as grounds for discipline and treat them, at most, as a reason to start a conversation. That’s about as much weight as the technology can carry.