Are AI detectors accurate
What the vendors claim, what the independent tests found, who gets flagged most, and how to read a false positive rate before anyone acts on a score.
Accurate on what. That is the only honest answer, and the rest of this guide unpacks it: the leading detectors are accurate on the texts their vendors benchmark them on, far less accurate on texts by people who write plainly or in a second language, and easy to defeat with a sentence of instruction to the model. Below are the numbers that have been published, who published them, and what they are worth.
Two kinds of error, and only one that hurts people
A detector can be wrong in two directions. A false negative lets a generated text through as human: embarrassing for the detector, harmless to anyone. A false positive labels a human text as generated: a student accused of cheating, a job application binned, a freelance writer not paid. When a vendor says "99 percent accurate", ask which error the 1 percent is, on what texts, and how many texts the 1 percent will be applied to. Vanderbilt University did that arithmetic in 2023: Turnitin claimed a 1 percent false positive rate; Vanderbilt had submitted 75,000 papers to Turnitin the previous year; so the feature would have wrongly flagged around 750 student papers in a year. The university disabled it.
What the vendors publish
Each detector vendor publishes its own figures, measured on its own benchmark. GPTZero's technology page, read on 25 October 2026, claims 96.5 percent accuracy on mixed documents (texts with both human and generated passages) and a false positive rate under 1 percent, and says that targeted work since 2022 brought its false positive rate on TOEFL essays down to 1.1 percent. Turnitin claimed a 1 percent false positive rate at the launch of its feature in April 2023, as Vanderbilt records, and did not publish how its tool works. These are claims by the sellers, on texts they chose, and they may well be true on those texts. They say nothing about your text.
| Who | Figure | On what | Status |
|---|---|---|---|
| GPTZero | 96.5% accuracy, under 1% false positives, 1.1% on TOEFL essays | its own and "independent" benchmarks, not published on the page | vendor claim, read 25 October 2026 |
| Turnitin | 1% false positives | its own test at launch, April 2023 | vendor claim, method not published (Vanderbilt) |
| OpenAI classifier | "low rate of accuracy" | its own output | withdrawn 20 July 2023 (TechCrunch) |
| Seven detectors tested by Stanford researchers | 61.22% false positives on TOEFL essays, near-perfect on US eighth-grade essays | 91 TOEFL and 88 eighth-grade essays, human-written | independent, peer-reviewed study, 2023 |
What the independent test found
The most cited independent test is by Weixin Liang and colleagues at Stanford, published in 2023. They ran seven widely used detectors on 91 essays written by non-native English speakers for the TOEFL exam and 88 essays by US eighth-graders, all human-written. On the eighth-graders the detectors were close to perfect. On the TOEFL essays they flagged more than half as generated: an average false positive rate of 61.22 percent across the seven tools. Eighteen of the 91 essays (19.78 percent) were flagged by all seven detectors at once; 89 of the 91 (97.80 percent) were flagged by at least one.
The study also found the mechanism. The essays flagged by every detector had significantly lower perplexity than the rest: a narrower vocabulary and safer constructions, which is what writing in a second language looks like, and also what a language model's default output looks like. When the researchers asked ChatGPT to enrich the TOEFL essays' word choices "to sound more like a native speaker", the false positive rate fell from 61.22 to 11.77 percent and only one essay remained unanimously flagged. When they asked it to simplify the eighth-graders' word choices "as if written by a non-native speaker", misclassification rose from 5.19 to 56.65 percent. Same authors, same ideas, different vocabulary, opposite verdicts.
Which detectors the study tested
The seven tools, accessed by the researchers on 15 March 2023, were Originality.AI, Quil.org's AI Writing Check, Sapling, OpenAI's own detector, Crossplag, GPTZero and ZeroGPT: the commercial field of the time, not obscure tools. Vendors have updated their models since, and GPTZero in particular states that it has worked on this bias and reports a TOEFL false positive rate of 1.1 percent in its own current testing. Both things can be true: the 2023 measurement stands as the independent evidence on record, and the vendor's later figure is a claim on its own benchmark that no independent test we could read has confirmed. The correct reading is not "detectors were bad in 2023 and are fine now" but "the only independent measurement found a 61 percent error on one population, and the improvements since are self-reported".
What a vendor's benchmark leaves out
Three things, usually. The threshold: every accuracy figure is measured at one setting of the flag, and the product may ship at another. The length: short texts have too few words for any method to average, so benchmarks use long ones and products refuse short ones, while the essay a teacher worries about is often four hundred words. And the mix: most real suspicion involves a text that is partly generated and partly rewritten, which is harder than a wholly generated one; GPTZero's 96.5 percent figure is specifically for mixed documents on its own benchmark, and it is the only vendor figure on that case we could find. A benchmark that reports one number for one threshold on long, cleanly separated texts is reporting the easiest case.
How easily the detectors are beaten
The same study ran the test the other way. The researchers had ChatGPT write 31 college admission essays; the detectors initially caught them. Then they gave the model one more instruction, "elevate the provided text by employing literary language", and the detection rate fell from 100 percent to 13 percent. On generated scientific abstracts the detectors did worse from the start, catching at most 68 percent, because technical writing has higher perplexity to begin with. TechCrunch, testing public detectors in February 2023, titled its findings "most sites claiming to catch AI-written text fail spectacularly", and in July 2023 reported OpenAI's withdrawal of its own classifier with the company's own words, a "low rate of accuracy". Two years of vendor improvement later, no published independent test has shown a style-based detector that survives a paraphrase.
Who gets flagged most
The population most exposed to false positives is the one the Stanford study measured: people writing in English as a second language. Others share the same profile of predictable, conventional prose: students taught to write to a formula, writers of legal, technical and administrative text, people with certain cognitive profiles who write plainly, and anyone using a grammar checker that smooths their sentences. The common factor is not dishonesty; it is regularity. A detector that measures regularity will find it in them.
| Writing | Why it scores as generated | Source of the observation |
|---|---|---|
| Non-native English | limited vocabulary, safe constructions, low perplexity | Liang et al., 61.22% false positives on TOEFL essays |
| Formula essays, template documents | same structure every time, signposting transitions | the mechanism of perplexity and classifier detectors (guide: how detectors work) |
| Technical, legal, administrative prose | formal register, no first person, repeated terms | same mechanism; generated abstracts were harder to catch for the mirror reason (Liang et al.) |
| Grammar-checked text | smoothed sentence lengths and word choices | same mechanism |
How to read a false positive rate
Three questions turn a vendor's percentage into something usable. First, on what texts was it measured: a rate on news articles and Wikipedia says nothing about first-year essays or non-native writers. Second, at what threshold: most detectors let the operator choose how confident the tool must be before it flags, and the false positive rate moves with that setting; a vendor quoting its best rate is quoting its strictest threshold, which also lets more generated text through. Third, how many texts will it be applied to: 1 percent of 75,000 is 750 people, and the cost of a wrong accusation falls entirely on them. Vanderbilt's conclusion in August 2023, that it did not believe AI detection software was an effective tool that should be used, was not a judgement about a percentage; it was about what the percentage meant in a room full of students.
What accuracy would even mean
Underneath the numbers is a problem no benchmark fixes. The detectors measure resemblance to a model's default style, and that style is also the style of plain, careful human writing; there is no feature of a text that belongs to machines alone, and the Stanford experiments show the boundary moving with a single change of vocabulary. Watermarks, planted by the model at generation, could in principle give a real answer, but only for the models that use them and only until the text is edited. So "are AI detectors accurate" has a stable answer: accurate enough to point at a text worth reading closely, and not accurate enough to say anything about the person who wrote it. The detector on this site is built on that answer: it shows its seven measurements and no probability, because the measurements are facts about the text and the probability would be read as a fact about you.
Questions people ask
How accurate are AI detectors?
It depends on the text more than on the tool. On the texts vendors benchmark, the leading detectors claim accuracy above 95 percent and false positive rates under 1 percent. On 91 essays by non-native English speakers, seven widely used detectors flagged 61 percent on average as generated, and 98 percent were flagged by at least one of them. OpenAI withdrew its own detector in 2023 for low accuracy.
What is a false positive in AI detection?
A human-written text that the detector labels as generated. It is the error that harms people: a student accused, an applicant rejected, a writer disbelieved. A 1 percent false positive rate sounds small until it is applied to 75,000 papers, which Vanderbilt calculated at around 750 wrong flags a year.
Can AI detectors be fooled?
Yes, cheaply. Liang and colleagues cut the detection rate on generated college essays from 100 percent to 13 percent with a single follow-up instruction to the model to vary its language. Paraphrasing tools and light human editing have the same effect. A clean score proves little.
Which AI detector is the most accurate?
We have not tested them and we do not rank them. Each vendor publishes its own figures on its own benchmark; GPTZero, for instance, claims 96.5 percent on mixed documents and under 1 percent false positives. Independent tests on specific populations, such as non-native writers, have found much higher error rates for the same class of tools.
Why do AI detectors flag my own writing?
Because they measure how predictable and regular a text is, and careful human writing is predictable and regular. Formal English, writing in a second language, technical and legal prose, and texts written to a template all score as generated more often than casual writing does.
Is Turnitin's AI detector accurate?
Turnitin claimed a 1 percent false positive rate at launch in April 2023 and has not published its method. Vanderbilt University disabled the feature in August 2023 on the grounds that it could not evaluate the claim and that 1 percent of its submissions was too many wrong accusations; other universities followed.
Sources
- Liang, Yuksekgonul, Mao, Wu, Zou, GPT detectors are biased against non-native English writers, arXiv 2304.02819 (2023): 7 detectors, 91 TOEFL and 88 US eighth-grade essays, false positive rates and bypass prompts, read 25 October 2026.
- Vanderbilt University, Guidance on AI detection and why we are disabling Turnitin's AI detector (16 August 2023): the 1 percent claim, the 75,000 papers, the decision, read 25 October 2026.
- TechCrunch, OpenAI scuttles AI-written text detector over "low rate of accuracy" (25 July 2023), read 25 October 2026.
- TechCrunch, Most sites claiming to catch AI-written text fail spectacularly (16 February 2023), read 25 October 2026.
- GPTZero, Technology page: the vendor's claimed accuracy (96.5 percent on mixed documents), false positive rate (under 1 percent) and TOEFL false positive rate (1.1 percent), read 25 October 2026.