Turnitin, GPTZero and the rest try to infer who wrote a piece of work from the words on the page. That's a guess, and a guess that gets a real student hauled in front of a misconduct panel is the worst kind. Here's why the whole approach is breaking down, in the words of the universities and researchers who studied it.
OpenAI built ChatGPT, then built a tool to detect it. It caught just 26% of AI text while wrongly flagging 9% of human writing, and OpenAI quietly shut it down, citing its low rate of accuracy. If the people who made the model can't reliably spot its output, third-party detectors are guessing.
OpenAI, via Ars Technica (2023) →After evaluating it, Vanderbilt disabled Turnitin's AI detection for all staff, citing false positives, a lack of transparency, and bias against non-native English speakers. Its verdict: false positives are not just a possibility but an inevitability, and the tool cannot be fairly used to judge whether work was AI-written.
Vanderbilt University (2023) →A Stanford study found leading detectors wrongly flagged 61% of TOEFL essays by non-native English speakers as AI-generated, while rating native-speaker essays as human. Detectors reward bursty, idiomatic prose, so the simpler, consistent writing of EAL and many neurodivergent students reads as machine-made.
Stanford, Liang et al. (2023) →Soheil Feizi's team at the University of Maryland concluded that no publicly available AI detectors are sufficiently reliable in practical scenarios, adding that they have a very high false-positive rate and can be pretty easily evaded. As models converge on human writing, the statistical signal detectors rely on disappears.
University of Maryland (2023) →Detectors are trained on yesterday's models. Today's models (GPT-4o, Claude, Gemini) write with far more human-like variety, and a few reworded sentences defeats most checkers. So the honest student gets flagged while the one who actually used AI, and edited a little, sails through.
arXiv survey of detection robustness (2024) →Detectors won't explain how they reached a score, so a student can't challenge it, you can't verify it, and your institution can't audit it. Even Turnitin tells schools not to treat its score as proof of misconduct. A number you can't explain is not evidence.
Turnitin guidance to institutions →“False positives are not just a possibility, but an inevitability.”
A student who writes their own essay, and a student who generates one with AI and edits it heavily, can produce text that is statistically identical. The finished page often doesn't contain enough information to reveal who (or what) wrote it. No amount of cleverness fixes that, because the missing evidence was never in the text to begin with.
That evidence lives in the process: the drafts, the pauses, the pasting, the rhythm of the keys. So that's what Learnaway reads, not the prose. Tellingly, the detector industry is quietly moving the same way, towards revision history and writing timelines. We just started there.
Set your next homework with Learnaway and see how the work was written, free for your whole school to start.
