Are machine-text detectors accurate? What the evidence shows

Are machine-text detectors accurate? Not accurate enough to prove that someone used a chatbot. OpenAI withdrew its own detector in 2023 "due to its low rate of accuracy". A peer-reviewed study found seven popular detectors flagged, on average, 61% of essays by non-native English speakers as machine-written, and a test of 14 tools concluded they were "neither accurate nor reliable". Turnitin, one of the two commercial tools in that test and widely used in academia, says its false positive rate is below 1% only for documents it judges to be more than 20% machine-written text, and about 4% at the level of individual sentences. This guide explains what the evidence shows, why these tools make mistakes, and what to do if you are a teacher, a student or a publisher. We did not run a detector test ourselves: the figures below come from the studies and vendors cited.

What the evidence says, in one table

Table: what studies and vendors report about machine-text detectors

Why machine-text detectors make mistakes

A language model writes by choosing likely next words, as our guide to what an LLM is explains. Many detectors judge how predictable a text is (its perplexity, in the jargon), and that is the weakness: plenty of humans write predictably too.

Liang's study shows the effect clearly. The non-native writers' essays were flagged because of "constrained linguistic expressions": a smaller vocabulary and simpler phrasing. When the researchers asked ChatGPT only to "enhance the word choices to sound more like that of a native speaker", the average false positive rate on the same essays fell from 61.22% to 11.77%. Of the 91 essays, 89 were flagged by at least one detector, and 18 by all seven. The authors caution against using detectors "in evaluative or educational settings".

OpenAI listed the same limits for its own tool. It was "very unreliable on short texts (below 1,000 characters)", recommended only for English, unreliable on code, and could not judge text that is predictable by nature: that it was impossible to predict whether a list of the first 1,000 prime numbers was written by a model or by a person. It also warned that "machine-written text can be edited to evade the classifier".

That last point cuts both ways. Weber-Wulff's team found that "content obfuscation techniques significantly worsen the performance of tools", so a determined cheater can often slip through while an honest writer with a plain style gets flagged.

Turnitin's numbers, read carefully

Turnitin publishes more detail than most vendors, and its own statements show where the risk lies. In March 2023 it said it aimed for "a less than 1% false positive rate" but acknowledged "a small risk of false positives". In May 2023, after testing 800,000 documents written before ChatGPT existed, it reported:

• Documents above 20% machine-written text: a document-level false positive rate "less than 1%".
• Documents below 20%: "a higher incidence of false positives", so Turnitin now shows an asterisk instead of a reliable score.
• Individual sentences: a false positive rate of about 4%.
• Short texts: the minimum length rose from 150 to 300 words.

A rate below 1% sounds small, but it applies to every honest paper submitted. Turnitin itself tells instructors: "given that our false positive rate is not zero, you as the instructor will need to apply your professional judgment".

What to do with a detector score

• If you teach: treat a score as a reason to talk, never as proof. OpenAI said its classifier "should not be used as a primary decision-making tool". Compare the work with the student's earlier writing, ask about their process, and know your institution's policy before you act.
• If you are accused: ask which tool was used and what score it gave, point to the documented error rates above, and show your process: drafts, notes, sources and your document's version history.
• If you publish: detection is not the way to prove where content came from. The EU's answer is provenance: Article 50 of Regulation (EU) 2024/1689 requires generative models providers to mark output "in a machine-readable format and detectable as artificially generated or manipulated". Our guide to Article 50 explains who must label what and when.

The bottom line

Detectors can sometimes spot unedited machine-written text, but the evidence shows they mislabel human writing, especially from non-native speakers, and miss machine-written text that has been lightly edited. If a decision affects someone's grade or job, a detector score should never be the only evidence. For the related problem of model output that sounds right but is wrong, see our guide to model hallucinations, and for how models choose words at all, what a token is.