Since ChatGPT arrived, AI detection has been switched on across schools, universities and some journals. Every week, people who wrote their own work are told a tool thinks otherwise. How often does that happen, and how much should a score count for?
We looked at the independent research and the vendors' own numbers. Then we worked through the arithmetic most discussions leave out: what a flag actually means once you account for how many people are using AI in the first place.
- 61%Average rate at which seven detectors labelled human-written TOEFL essays as AI (Liang et al., 2023)
- 26%Share of AI text OpenAI's own classifier caught before it was withdrawn in 2023
- 14 toolsTested in one study; none was found to be accurate or reliable (Weber-Wulff et al., 2023)
- ~750Papers a year Vanderbilt estimated could be wrongly flagged at a 1% false positive rate
How AI detectors work
Most detectors don't look for a hidden signature. They estimate how predictable the text is. Language models write by choosing likely next words, so their output tends to have low perplexity (each word is unsurprising) and low burstiness (sentences are similar in length and structure). Detectors score text on features like these and report a probability.
The weakness is built in. Predictable, uniform text isn't unique to machines. People produce it too, especially when they:
- write in a second language, with a smaller working vocabulary and safer constructions;
- follow the conventions of a genre closely, as in a methods section or a lab report;
- write short texts, where there is too little to measure;
- have their text polished heavily by grammar and style tools.
Academic writing hits several of these at once.
What the research found
| Source | What was tested | Finding |
|---|---|---|
| Liang et al., Patterns, 2023 | Seven detectors on 91 TOEFL essays and 88 US eighth-grade essays, all written by people | Near-perfect on the native speakers' essays. On the TOEFL essays, an average false positive rate of 61%. All seven detectors flagged 18 of the 91, and at least one flagged 89. |
| Weber-Wulff et al., Int. J. for Educational Integrity, 2023 | 14 tools, including Turnitin and PlagiarismCheck, on human, AI, translated and edited texts | The tools were "neither accurate nor reliable", leaned towards labelling text as human, and did worse when AI text had been paraphrased or edited. |
| OpenAI AI Classifier, 2023 | OpenAI's own detector, on its own evaluation set | Caught 26% of AI-written text and wrongly flagged 9% of human text. Withdrawn in July 2023 for low accuracy. |
| Turnitin, 2023 | The vendor's own testing | Under 1% document-level false positives for documents with more than 20% AI writing. Scores below 20% now show an asterisk instead, because false positives were more common there. Minimum length raised to 300 words. |
Two things are worth noticing. The vendor's figures are much better than the independent ones, which is normal: vendors test on data that suits their product, and the conditions in real classrooms are messier. And the studies agree on direction. Detectors do best on long, unedited AI text and worst on exactly the cases where a decision matters: short texts, edited texts, and writers whose English is formal or non-native.
The base-rate problem
Even a detector with a low false positive rate can produce a lot of false accusations. What a flag means depends on how common AI use is in the group being checked, and that's rarely discussed.
Take a detector that catches 90% of AI-written work and wrongly flags 1% of human work, roughly Turnitin's claimed rate. Run it over 1,000 submissions and count the flags:
| If this many submissions used AI | Flags on AI work | Flags on human work | Share of flags that are wrong |
|---|---|---|---|
| 20% (200 of 1,000) | 180 | 8 | 4% |
| 5% (50 of 1,000) | 45 | 9.5 | 17% |
| 2% (20 of 1,000) | 18 | 9.8 | 35% |
| 5%, with a 5% false positive rate | 45 | 47.5 | 51% |
In a group where AI use is rare, a third of flags can land on innocent people even with a 1% error rate. With a higher false positive rate, plausible for second-language writers given Liang's results, a flag is about as likely to be wrong as right. The score shown to the instructor looks the same in every row.
This is why Vanderbilt University turned off Turnitin's AI detector in August 2023. It calculated that a 1% false positive rate, applied to the 75,000 papers it submitted in 2022, would have meant around 750 students wrongly flagged. Other universities have since made the same decision, or told staff not to treat a score as evidence on its own. Turnitin also says its indicator shouldn't be the sole basis for action against a student.
Who gets flagged most
- Non-native English writers. The clearest finding in the research. A narrower vocabulary and more conventional phrasing look like low perplexity.
- Formulaic sections. Methods, abstracts and literature summaries follow conventions that make them predictable by design.
- Heavily polished text. Extensive grammar and style correction can make prose more uniform.
- Short pieces. Turnitin raised its minimum to 300 words because accuracy falls on shorter text.
What to do if your work is flagged
- Don't panic, and don't admit to something you didn't do just to end the conversation.
- Ask for specifics: which policy, what evidence, and what the score actually was. A score alone is a probability, not a finding.
- Show your process. Dated drafts, document version history, research notes, annotated PDFs, browser history of the sources you read. This is far stronger evidence than any score.
- Offer to explain the work. Someone who wrote an argument can explain why it's structured as it is and where each claim comes from.
- Bring the research. The Liang and Weber-Wulff studies, and Turnitin's own guidance, are reasonable things to put in front of a panel.
- Ask for human review, and contact your student union, ombudsperson or academic advisor if the process doesn't allow one.
Protect yourself before it happens
The best defence is a visible process. Write in a tool that keeps version history, or save dated drafts. Keep your notes and the sources you read. If your course or journal allows AI help, use it within the rules and say so. Our guide to what journals allow when you use AI covers disclosure for published work.
Be wary of AI "humanizer" tools that promise to get text past detectors. If AI use wasn't allowed, disguising it doesn't make it allowed, and deliberate concealment is usually treated more seriously than the use itself. If it was allowed, there's nothing to hide.
Where AI drafting is permitted, the honest approach is to make the work genuinely yours: your argument, your evidence, your judgment about every sentence. Our use case on editing AI drafts into your own voice walks through that process.
For instructors and editors
- Treat a detector score as a reason to talk to the writer, never as proof.
- Be especially cautious with second-language writers and short or formulaic texts.
- Ask for process evidence and a conversation before any formal step.
- State clearly what AI use is and isn't allowed for each assignment, so students aren't guessing.
- Assessments that include drafts, oral discussion or work done in class make the question much easier to answer.
The bottom line
AI detectors can tell you that a piece of text is statistically similar to machine output. They can't tell you who wrote it. The research shows they're least reliable for the writers most likely to be flagged, and the arithmetic shows that even a small error rate produces many false accusations where AI use is rare. Use them, if at all, to start a conversation, and decide on the evidence.
