The Myth That AI Detectors Can Reliably Catch AI Writing

A student turns in an essay. A detector flags it as "likely AI." The student insists they wrote every word, and they may well be telling the truth. This scene plays out daily now, and it rests on a comforting assumption that does not hold up: that software can reliably tell human writing from machine writing. The evidence in 2026 says it cannot, at least not well enough to justify the weight people are putting on it.

The claim, and why it is tempting

The promise is simple. Paste in a document, get back a percentage, and treat that number as a verdict. For a teacher facing a stack of essays or a manager screening applications, a single confidence score is appealing. The trouble is that the score is far softer than it looks, and the people most likely to be wrongly flagged are often the ones least able to push back.

What the testing actually shows

False positives, human writing wrongly tagged as AI, are the central problem. Measured rates range widely, from around 2 percent for the better tools up to roughly 15 percent for weaker ones, according to comparative testing this year. A 15 percent false-positive rate means that in a class of 100 honest students, fifteen could be accused on the strength of a bad reading.

Consistency is just as shaky. In one round of tests, the same 500-word post scored 23 percent AI one day and 67 percent two days later, with no change to the text. A tool that contradicts itself within a week is not something to hang an academic-integrity case on. Even the strongest performers, like Turnitin at around 92 percent overall accuracy in spring testing, still carry a false-positive rate that becomes a real number of wronged people once you scale it across a school.

Who gets hurt

The damage is not spread evenly. Writers working in a second language are flagged at two to three times the rate of native speakers, because the clean, careful prose of someone writing in a learned language can resemble the patterns detectors are trained to suspect. Technical documentation and some creative writing trip the alarms for similar reasons. The verdict that is supposed to protect fairness ends up landing hardest on people who did nothing wrong.

The honest position

None of this means detectors are useless. As one signal among several, alongside a conversation with the writer, a look at draft history, and plain editorial judgment, they have a place. The myth is not that they detect anything. The myth is that they detect reliably enough to stand alone as proof. The expert consensus is blunt on this point: no current detector is accurate enough to serve as sole evidence of who, or what, wrote a piece of text. Treat the percentage as a prompt to ask questions, not as an answer.