Blog

Are AI Detectors Accurate?

What independent studies found, where detectors go wrong, and how to read a score from different tools.

By the AI Checker Humanizer team. Published . Updated .

The short answer

It depends on the detector and on the text. In recent independent tests, some AI detectors flagged almost no human writing and missed little unedited AI text, so on long, unedited text the better tools are now quite accurate. Performance on short or rewritten text varies by detector: several studies found that rewriting reduced detection, and some tools have wrongly flagged large shares of writing by non-native English speakers.

A score is never proof on its own. Even a detector with a very low error rate will wrongly flag some people when thousands of essays go through it, which is why many schools tell staff not to act on a score alone.

What the studies found

These are independent studies with published methods and numbers. Vendor claims are covered separately further down.

Independent studies of AI detector accuracy
StudyTexts testedAI detectorsWhat they found
Liang et al., Patterns (2023)91 TOEFL essays by non-native English writers, 88 US eighth-grade essays7 detectors, including GPTZero and OriginalityNon-native essays wrongly flagged 61% of the time on average; the US essays were judged almost entirely correctly
Weber-Wulff et al., International Journal for Educational Integrity (2023)Human, AI and edited AI texts14 tools, including TurnitinDescribed the tools as “neither accurate nor reliable”; edits and paraphrasing made results worse
Jabarian and Imas, NBER working paper (2025)1,992 human texts across six genres, plus matching AI texts from four modelsPangram, Originality, GPTZero, an open-source modelFalse positive rates of 0.1% or less for Pangram, 0.1% to 0.3% for Originality, about 0.7% for GPTZero; the open-source model was far worse
Epoch AI (2026)495 human texts of about 500 words; AI texts from three current modelsPangram, GPTZero, OriginalityNo false positives for Pangram or GPTZero, 3.8% for Originality; when told to imitate a human style, AI texts were missed 10% to 18% of the time
Karr et al., preprint (2026)642 abstracts of published papers, 2013–15 and 2023–25Pangram, GPTZeroNo pre-2016 abstracts flagged; abstracts lightly refined with AI flagged 38% to 80% of the time; text run through a humanizer was missed over 96% of the time
Sources checked November 11 and 12, 2026.

Two things stand out. The 2025 and 2026 studies reported low false positive rates for some detectors on unedited text; their samples and methods differ from the 2023 studies, so the figures aren't directly comparable. And several studies found that rewriting AI text reduced detection, though results varied by tool.

False positives: when human writing gets flagged

A false positive is human writing called AI. For the strongest AI detectors in the recent studies, the rate on ordinary human text was well under 1%. That still adds up: at 0.5%, a school checking 10,000 essays would wrongly flag about 50.

The studies above also show where rates go up. Liang's study found seven 2023-era detectors wrongly flagged most TOEFL essays by non-native writers, though that doesn't tell you today's rate for every tool. And in the Karr study, recent abstracts were flagged more often than older ones, which the authors note may reflect real AI editing rather than detector mistakes: human writing polished with AI is partly AI writing by then.

If you've been flagged for work you wrote, our guide to AI detector false positives goes through what to do.

Non-native English writers

The 2023 Stanford study by Liang and colleagues is the one most people cite. Seven AI detectors of that time wrongly flagged 61% of TOEFL essays on average, and all seven agreed on about a fifth of them. In that study, the TOEFL essays flagged by all seven detectors had more predictable wording than the other TOEFL essays.

Newer tools say they have addressed this; GPTZero, for example, reports a 1.1% false positive rate on TOEFL essays (vendor-reported). Our own detector flagged none of the 600 essays by non-native English writers in our test set:

Our detector on English texts it hadn't seen before
TextsHow manyResult95% bound
Human-written texts1,0961 flagged as likely AI (0.09%)At most 0.52%
Essays by non-native English writers6000 flagged as likely AIAt most 0.64%
AI-generated texts from 6 model families278270 caught as likely AI (97.1%)At least 94.4%
English, November 2026, detection engine version w0p5-38c6d591.

Even so, be careful with any score on second-language writing, especially short pieces.

Humanized and paraphrased text

Results for rewritten and humanized text vary a lot by detector. Research on paraphrasing attacks, such as the DIPPER paraphraser in 2023, showed that rewriting AI text with another model could take one detector's catch rate from 70% to under 5%. More recent results vary by tool: in the NBER study, Pangram stayed robust against a commercial humanizer while GPTZero missed about half of the humanized texts, and in the Karr study text run through a humanizer was missed over 96% of the time.

Detector makers have responded. Turnitin added detection of text changed by “AI bypasser” tools, also called humanizers, for English in August 2025.

Why AI detectors give different results

Paste the same essay into three AI detectors and you'll often get three different numbers. There are a few reasons.

  • Different models. Each tool trains its own classifier on its own collection of human and AI writing.
  • Different thresholds. Each decides where to draw the line between AI and human, trading false positives against missed AI text.
  • Different units. Some score sentence by sentence, some by segment, some the whole document at once.
  • Different meanings. Turnitin, Grammarly and QuillBot report roughly what share of the text looks AI-generated. GPTZero reports how confident its model is in one label for the whole document. Our score is the chance that most of the text is AI-written.

So a 40% from one tool and a 40% from another can mean different things. Before comparing, check what each number measures.

Is a 20% or 40% AI score bad?

It depends on the tool. On Turnitin, 20% means about a fifth of the qualifying text was labeled AI; scores from 1% to 19% are shown only as an asterisk because they're less reliable. On GPTZero, a percentage is the model's confidence in its label for the whole text, not a share of the text.

On our detector, a score is a probability, and this is how to read it:

How to read the AI score in English
ScoreWhat it usually meansWhat to do next
91 to 100Likely AI-writtenLook at the red sentences and the patterns in them. If you wrote them, keep your drafts as a record.
21 to 90Mixed: no confident verdict on the text as a wholeCheck the red and yellow sentences one by one. A mixed result doesn't confirm a mix of human and AI writing.
0 to 20Likely written by a personUsually nothing. A few red sentences can still be worth a look.
Two extra rules: if red sentences make up 25% or more of the words, the result is never "likely written by a person", whatever the score. This table is for text in one language; when a text mixes languages, read the score for each language on the result.

The score is the chance that most of the scored text is AI-written, calibrated on equal amounts of human and AI writing.

You can check a text below and see a color for every sentence, so you know where the score comes from. Our AI detection page also lists how often we caught unedited text from six model families. For the general picture, read how AI detectors work.

0 / 2,000 words

How accurate are popular detectors?

Independent results and vendor claims are labeled below. Check the test date and detector version, since updates can change performance. For free limits and which detector suits which job, see our roundup of the best AI detectors.

Is GPTZero accurate?

In independent tests from 2025 and 2026, GPTZero had low false positive rates on ordinary human text (about 0.7% in the NBER study, none of 495 texts in the Epoch AI test). It did much worse on humanized text. GPTZero says its error rate is under 2% when it reports high confidence (vendor-reported).

Is Turnitin accurate?

Turnitin reports a document-level false positive rate under 1% for documents with more than 20% labeled AI (vendor-reported). In a Temple University test, it scored 28 of 30 human papers as fully human and 23 of 30 AI papers as fully AI, but gave a partial score to only 13 of 30 papers that mixed the two. Our article on Turnitin AI detection has the details.

Is QuillBot's AI detector accurate?

QuillBot's own blog calls its detector highly accurate without explaining how that was measured (vendor-reported). A small peer-reviewed study by Paullet and colleagues (Issues in Information Systems, 2025) found it judged 35 of 36 human texts and all 34 AI texts correctly; in a 2025 preprint by Alshammari and Rao, its average AI score on humanized AI text fell to 58%.

Is Grammarly's AI detector accurate?

Grammarly advertises 99% detection accuracy (vendor-reported) but its own help page says the detector “is not 100% accurate and should not be used as a definitive assessment.” We found no peer-reviewed study of it; a small Authors Guild test in May 2026, using ten human articles published in 2022 or earlier, found low scores on all of them. (Sources checked November 12, 2026.)

Is Pangram accurate?

In the 2025 and 2026 independent studies above, Pangram had the lowest false positive rates and held up best against humanizers, though a 2026 study still found it missed most text run through one humanizer. Pangram itself reports a false positive rate of about 0.01% on its own tests (vendor-reported).

Frequently asked questions

Can an AI detector be wrong?

Yes, in both directions. It can flag human writing and miss AI text, especially when the text is short, formulaic or has been rewritten.

Are AI detectors accurate for academic writing?

Results depend on the detector and on how AI was used. In one 2026 study, detectors flagged 38% to 80% of abstracts lightly polished with AI, without distinguishing that editing from full AI drafts. Treat a score on academic work with extra care.

Which AI detector is the most accurate?

In the recent independent tests we cite, Pangram had the lowest false positive rates, with GPTZero close behind on unedited text. That tells you how often human writing was wrongly flagged; to compare overall accuracy, check missed AI text too, and the date of the test.

Why do AI detectors give different results?

They use different models and thresholds, split text differently, and report different things: some give the share of text that looks like AI, others a confidence for the whole document.

Should a teacher rely on an AI score?

Not on its own. Many universities say in their policies that a detector score can't be the only evidence, and Turnitin says the same about its own tool.

How accurate is your detector?

On English texts it hadn't seen before, it flagged 1 of 1,096 human texts and none of 600 essays by non-native English writers, and caught 270 of 278 AI texts.