PROGRAMS

AI Detector Accuracy Tracker

8 popular AI detectors, the accuracy and false positive rate each company claims, and what several independent studies found when they tested them. Includes the latest available data.

Tracker status

Last verified

Oct 8, 2026

Detectors tracked

8

Independent studies

8

Changes this update

First version

See the update log →

What the companies claim

6

of 8 popular AI detectors claim 99% accuracy or higher

ZeroGPT claims 98.4%. Turnitin publishes no accuracy figure.

What independent tests found

13

of 26 independent test results put detection below 80%

All 13 are from studies published from 2023 to 2025.

6

publish a false positive rate

Each is 1% or lower. Originality.ai and Phrasly publish none.

2

give no details of the test behind their claim

Phrasly and ZeroGPT

8

independent studies are tracked here

They cover 7 of the 8 detectors. Phrasly has no result.

What Each AI Detector Claims

DetectorClaimed accuracyClaimed false positive rateWhat the claim is based onSource
CopyleaksOver 99%0.03%Its own tests, the latest on 500,000 textsCopyleaks AI Detector, Testing methodology
GPTZero99%1%Its own testing, Penn State benchmarks and the RAID benchmarkGPTZero
Originality.ai99%+Not publishedIts own benchmark of 456,872 samplesOriginality.ai accuracy study
PangramOver 99%0.01% (1 in 10,000)Public datasets of tens of millions of documentsPangram
Phrasly99.8%Not publishedNot statedPhrasly AI Detector
TurnitinNot publishedLess than 1%Its own test on more than 700,000 papers written before ChatGPTTurnitin AI detector, Turnitin FAQ
Winston AI99.87%0.12% (7 of 6,000)Its own test of 12,000 documentsWinston AI, Winston 5.0 test
ZeroGPT98.4%Less than 1%Not statedZeroGPT

A false positive is human writing that a detector labels as AI. Turnitin’s figure applies to documents it scores as more than 20% AI-written.

Other Figures the Companies Publish

These are other accuracy and false positive figures the same companies publish on their own websites.

DetectorMeasureHeadline claimOther published figureSource
CopyleaksFalse positive rate0.03%0.30% at the default sensitivity setting in its V11 test. 0.16% at Extra Safe and 0.66% at Extra Sensitive.Copyleaks testing methodology
GPTZeroAccuracy99%99.76% in its February 2026 benchmarkGPTZero benchmark
GPTZeroFalse positive rate1%0.08% in its February 2026 benchmarkGPTZero benchmark
Originality.aiAccuracy99%+99.4% at the default 15% AI Allowed setting. 98.0% at 5%, 95.8% at 25% and 88.7% at 40%.Originality.ai accuracy study
PangramFalse positive rate0.01% (1 in 10,000)0.0041% (41 of 1,000,000) on its Pangram 4 model cardPangram 4 model card

What Independent Tests Found

DetectorDetection resultHuman text wrongly flaggedTest setYearSource
PangramMissed 0 of 297 basic AI texts and 30 of 297 written to imitate a real author0 of 495495 human passages and 297 AI passages per test2026Epoch AI
GPTZeroMissed 2 of 297 basic AI texts and 32 of 297 written to imitate a real author0 of 495495 human passages and 297 AI passages per test2026Epoch AI
Originality.aiMissed 1 of 297 basic AI texts and 53 of 297 written to imitate a real author19 of 495495 human passages and 297 AI passages per test2026Epoch AI
GPTZeroMissed 0% to 7% of AI text0% to 2.38%1,992 human and 7,968 AI passages2025Jabarian and Imas
Originality.aiMissed 0% to 4.48% of AI text0% to 2.18%1,992 human and 7,968 AI passages2025Jabarian and Imas
PangramMissed 0% to 2% of AI text0% to 0.75%1,992 human and 7,968 AI passages2025Jabarian and Imas
Pangram98.0% of AI articles detected2%300 articles2025Russell et al.
GPTZero85.3% of AI articles detected0.7%300 articles2025Russell et al.
GPTZero97.22% accuracy0%72 human abstracts plus AI-generated abstracts2025Pratama
ZeroGPT64.35% accuracy16.67%72 human abstracts plus AI-generated abstracts2025Pratama
GPTZero66.5% of AI text detectedHeld at 5% by the testMore than 6 million AI texts2024RAID benchmark
Originality.ai85.0% of AI text detectedHeld at 5% by the testMore than 6 million AI texts2024RAID benchmark
Winston AI71.0% of AI text detectedHeld at 5% by the testMore than 6 million AI texts2024RAID benchmark
ZeroGPT65.5% of AI text detected5% target not reachedMore than 6 million AI texts2024RAID benchmark
Originality.ai100% of AI articles detectedNot reported150 medical articles2024Liu et al.
Turnitin94% of AI articles detected0%150 medical articles2024Liu et al.
GPTZero70% of AI articles detected22%150 medical articles2024Liu et al.
ZeroGPT96% of AI articles detectedNot reported150 medical articles2024Liu et al.
Turnitin50% accuracy. 7.9% after the text was edited to evade detection.–114 samples2024Perkins et al.
GPTZero26.4% accuracy. 16.7% after the text was edited to evade detection.–114 samples2024Perkins et al.
ZeroGPT31.3% accuracy. 17.3% after the text was edited to evade detection.–114 samples2024Perkins et al.
Copyleaks73.9% accuracy. 58.7% after the text was edited to evade detection.–114 samples2024Perkins et al.
Turnitin76% accuracy0 of 1854 documents2023Weber-Wulff et al.
GPTZero54% accuracy9 of 1854 documents2023Weber-Wulff et al.
Winston AI67% accuracy0 of 1854 documents2023Weber-Wulff et al.
ZeroGPT59% accuracy0 of 1854 documents2023Weber-Wulff et al.

Each study tested the version of the detector available at the time. The RAID benchmark sets every detector to a 5% false positive rate before measuring detection.

How We Maintain This Tracker

We check each company’s website for new accuracy and false positive data, and we add independent studies as they’re published.

Each figure is confirmed on the company’s own website or via the study itself.

Update Log

Oct 8, 2026

Published the first version with 8 detectors and 8 independent studies.

How We Track AI Detector Accuracy

Detectors

We track 8 widely used AI-text detectors: Copyleaks, GPTZero, Originality.ai, Pangram, Phrasly, Turnitin, Winston AI and ZeroGPT.

Accuracy Figures

Accuracy and false positive rate are based on the figures each company publishes on its own website for its current English-language detector. “Not published” means we found no figure from the tool/company.

Independent Tests

An independent test is a peer-reviewed paper, working paper or benchmark run by researchers who do not work for a detector company, and that reports a result for a detector that we track.

Detection Results

The detection result is each study’s own primary measure. Some studies report overall accuracy, some report the share of AI text detected and some report the percentage of AI content the tool missed. A result counts as below 80% when the study’s accuracy or detection figure is under 80%.

Comparing the Numbers

The studies use different sample text, AI models and scoring rules, and each tested the version of the detector available at the time. Therefore it’s worth noting that company claims and the independent results are not measured the same way.