Here is a claim that sounds reasonable and has launched a thousand classroom confrontations. If a machine can write an essay, another machine can reliably tell when it did. Paste the text into a detector, read off a percentage, and you have your answer. It is a tidy idea. The evidence says it is mostly wrong, and wrong in ways that fall hardest on the people least able to push back.
Start with the most cited result. A Stanford study ran seven popular detectors over essays written by non-native English speakers and found that the tools flagged about 61 percent of them as AI-generated. These were human-written essays. On writing by native English speakers the same tools were close to perfect. The detectors had, in effect, learned to read a smaller vocabulary and simpler sentence structure as a sign of a machine, which is also a fair description of plenty of human writers.
The disparities do not stop at language. Common Sense Media reported a false-positive rate of around 20 percent for Black students, against roughly 7 percent for white students. When a tool's mistakes line up that neatly with who the writer is, it is not acting as a neutral referee.
What the detectors actually measure
Part of the trouble is that these systems cannot see where text came from. They estimate how predictable a piece of writing is, on the theory that AI prose tends to be smoother and less surprising than human prose. That is a real statistical tendency, but a tendency is not a fingerprint. A careful human writer and an AI told to write loosely can both slip past the assumption. OpenAI's own early classifier caught only about a quarter of AI-written text while wrongly flagging some human writing, and the company quietly pulled it.
Where this is heading
The institutions that bought into detection are starting to back away. A New York court reversed a student's expulsion that had rested on a detector's verdict, and a growing number of universities have switched the feature off rather than defend it. The emerging view among researchers is modest and sensible. A detector score is at best one weak signal, never proof, and should never on its own decide that someone cheated.
None of this means AI writing is undetectable in every case, or that the worry behind it is silly. The flood of machine-generated text is real, and we have written about how much of it is now sloshing around online. The myth is the precision. The confident number a detector hands back carries all the authority of a measurement and almost none of the reliability, and treating it as hard evidence has already cost real people their reputations. For now, if you want to know whether a person wrote something, there is no shortcut that beats knowing the person and their work.
Commentarii · 0