What these tools actually measure
Almost every detector works on two statistics. Perplexity is how surprising the next word is, given the ones before it. Burstiness is how much that surprise varies from sentence to sentence. Language models are trained to pick likely words, so their output tends to have low perplexity and steady burstiness. Detectors look for that signature.
The problem is that plenty of human writing has the same signature: technical documentation, legal boilerplate, anything written to a template, and — importantly — anything written by somebody writing in a second language, who reaches for the common word because it is the one they are sure of.
The main tools
| Tool | Aimed at | Worth knowing |
|---|---|---|
| Turnitin AI detection | Universities and schools | Built into existing plagiarism workflows. Reports a percentage, not a verdict. |
| GPTZero | Teachers, general use | Reports per-sentence scores, which is more useful than one number for the whole document. |
| Originality.ai | Publishers and agencies | Paid, aimed at content teams checking freelance work. |
| Copyleaks | Enterprise | Combines AI detection with plagiarism checking. |
| Winston AI | Education and publishing | Includes OCR, so it can check scanned or photographed work. |
How reliable are they?
Vendors publish accuracy figures above 95%. Independent testing consistently finds lower, and — the part that matters — false positives are not evenly distributed. Published research has repeatedly found that text written by non-native English speakers is flagged far more often than text by native speakers, because simpler, more predictable vocabulary is exactly what the detectors are looking for.
OpenAI withdrew its own AI text classifier in 2023, citing a low rate of accuracy. That is the clearest available statement about how hard the problem is: the organisation with the most access to the models could not make a detector it was willing to stand behind.
How to use a score responsibly
- Treat it as a prompt to look, not a finding. A high score is a reason to read the work and talk to the person, nothing more.
- Never use it as sole evidence. No current detector is reliable enough to support an academic misconduct case on its own, and several institutions have said so in writing.
- Look at process instead. Draft history, version history and a short conversation about the argument tell you far more than a percentage does.
- Be aware of who gets flagged. If you apply a detector across a cohort, you will disproportionately flag your second-language students. That is a known property of the method, not a coincidence.
If your own writing gets flagged
It happens, and it is not evidence of anything. Keep your drafts — version history in Google Docs or Word is the single most useful thing you can show. Being able to walk somebody through how the argument developed settles the question in a way no counter-detector can.