Can an AI Detector Really Tell Who Wrote It?

A student is hauled in front of an academic panel because a piece of software flagged their essay as machine-written. It is a scene playing out often enough that it is worth examining the belief underneath it: that an AI detector can reliably tell whether a person or a model wrote a given passage. The honest answer, in 2026, is that it cannot do so with anything close to certainty.

Start with the strongest case for the tools. The best detectors are genuinely good in controlled conditions. Pangram, widely regarded as the leader, reports around 99 percent accuracy, and an independent University of Chicago Booth working paper found it had the lowest false-positive rate of the detectors tested, close to zero across different passage lengths. That is a real achievement, and it is why the claim persists. On clean, unedited text, a good detector is right far more often than it is wrong.

Where the claim breaks down

The trouble starts the moment the text is edited. A cottage industry of humanizer tools now exists to rewrite AI output until detectors wave it through, and independent testing shows they largely succeed. A 2026 head-to-head test of several detectors against a humanizer found that no tool stayed reliable once the text had been rewritten a few times. This week the pattern hardened: a company called StealthGPT announced a system it markets as producing text that current detectors, Pangram included, no longer catch. Whether it lives up to the boast, the direction of travel is clear. Detection and evasion are locked in an arms race, and evasion keeps closing the gap.

Then there is the quieter problem, the one that hurts real people. Even a very low false-positive rate becomes a lot of wrongly accused writers once you run millions of essays through it. Pangram's own guidance acknowledges that false positives happen and warns against treating a result as proof. Non-native English writers and people with plain, formulaic styles are flagged more often, which turns a statistical tool into an unfair one when it is used to decide someone's grade.

So the sensible reading is neither that detectors are useless nor that they are proof. They are probabilistic. A flag is a reason to look closer, ask questions, maybe have a conversation. It is not, on its own, evidence of anything. The myth is not that these tools work. It is that a percentage on a screen can substitute for judgement. On that point the research is consistent, and the case is worth keeping in mind whenever a machine claims to know something for sure.