Can AI Content Detectors Actually Tell
Schools and employers increasingly run text through AI detectors. We look at how they work, how often they are wrong, and why a confident score is the most dangerous thing about them.
Eddie Ochieng
July 16, 2026
AI content detectors promise something people badly want, a machine that can tell whether a machine wrote something. Universities buy them, employers run CVs through them, and editors use them to screen submissions. The trouble is that the promise is far more confident than the technology behind it.
This matters because the cost of being wrong is not symmetrical. A missed AI essay is an annoyance. A student wrongly accused of cheating is a catastrophe for that student, and it happens more often than the marketing admits.
How we compared
We looked at how these tools describe their own accuracy, at the published research on false positives, and at the pattern of user reports from students, teachers and writers on the receiving end. The vendors and the independent findings do not always agree, and where they diverge we say so.
How they actually work
Detectors do not find a watermark or a signature, because in most cases there is not one. They measure statistical properties of text. The two that matter are perplexity, roughly how surprising each word is given the ones before it, and burstiness, how much sentence length and structure vary.
Human writing tends to be lumpy. We write a long meandering sentence, then a short one. We choose an odd word. Language models, optimising for the most probable next word, tend to produce text that is smoother and more even. Detectors look for that smoothness and call it machine.
The flaw is in the method
Notice what this means. A detector is not detecting AI. It is detecting text that is unusually predictable and even. Plenty of humans write like that, and it is not a character flaw, it is a writing style. That is the root of every false positive story you have read.
Who gets wrongly accused
The false positives are not random, and that is the most uncomfortable part. Research and a steady stream of reports point the same way. Non native English speakers are flagged far more often, because writing in a second language tends to produce simpler, more regular sentence construction, which is exactly what these tools score as machine like.
The same applies to anyone writing in a plain, formal register. Technical documentation, legal writing, and the careful prose of someone who has been taught to write clearly and simply all trip the same wire. Ironically, being taught to write well can make you look like a robot.
The tools
GPTZero is the best known and the most widely used in education. Originality.ai aims at publishers and content teams and is more aggressive, which means it catches more and also flags more innocents. Copyleaks bundles AI detection with plagiarism checking, and Turnitin has the widest institutional reach simply because universities already use it for plagiarism.
They all publish impressive accuracy figures. Read those figures carefully, because accuracy on a clean benchmark of obviously human and obviously machine text tells you very little about accuracy on a real essay written by a real, nervous, second language student at two in the morning.
| Tool | Price | Best for |
|---|---|---|
| GPTZero | Free tier, paid plans available | The most common tool in education |
| Originality.ai | Paid, pay as you go credits | Publishers and content teams |
| Copyleaks | Free trial, paid plans | AI detection bundled with plagiarism |
| Turnitin | Institutional licence | Universities that already use it |
If you are on the receiving end
If you are a student or writer accused on the strength of a detector score, your best defence is process, not protest. Version history in Google Docs or Word, earlier drafts, notes, and a browser history that shows the research all demonstrate how the work came to exist. Keep them as a matter of habit.
FAQ
Can AI detectors be trusted?+
Not on their own. They produce false positives at rates that matter, particularly for non native English speakers and for plain, formal writing. Treat a score as a prompt to look closer, never as evidence.
Why do they flag human writing as AI?+
Because they measure how predictable and evenly structured text is, not whether a machine wrote it. Clear, simple, regular prose scores as machine like, whoever produced it.
Can you beat a detector by editing AI text?+
Usually, yes, which is the other half of the problem. Light editing to vary sentence length and word choice defeats most detectors, so the people they catch are often the least sophisticated, not the least honest.
Should schools use them?+
Only as one input among several, with a human conversation attached and a right of reply. Using a detector score as automatic proof of misconduct is indefensible given the known error rates.
If you use AI in your writing and want to do it honestly, see our guides to the best free AI writing tools for students and how to write better AI prompts.
