Is Copyleaks Accurate? What the Evidence Shows
Copyleaks gives four different false positive rates on its own pages, and the few independent tests disagree. Here are its claims, its sensitivity settings and what three independent studies found.
Short answer: Copyleaks claims over 99% accuracy and a very low false positive rate, but its own pages give four different false positive figures. Independent tests are few and small, and they disagree: one found a single false positive in 114 human texts, another flagged half of its human controls.
What Copyleaks says about its accuracy
- "Over 99% accuracy" on the product page, and 99.12% in the help center (AI Detector, help center).
- Its false positive rate appears as 0.03% (product page), 0.2% (help center), 0.30% at the default setting (methodology page) and under 1% for non-native English writers (2024 blog).
- Its September 2026 methodology reports internal tests on hundreds of thousands of texts, with false positive rates by sensitivity setting (methodology):
| Setting | False positive rate | Missed AI text |
|---|---|---|
| Extra Safe | 0.16% | 0.88% |
| Balanced (default) | 0.30% | 0.64% |
| Extra Sensitive (for humanized text) | 0.66% | 0.49% |
These are Copyleaks' own tests. Its terms say the AI features must not be the sole basis for adverse decisions such as academic misconduct findings (terms).
Copyleaks also cites a 2023 study "by Cornell Tech researchers" that ranked it most accurate. That study is Orenstrakh et al.; its listed authors are from the University of Toronto Mississauga, Maranatha Christian University and ESPOL, with no Cornell affiliation.
What independent studies found
Copyleaks was not tested in Liang et al. 2023, Weber-Wulff et al. 2023 or the RAID paper, and it is not on the RAID leaderboard.
| Study | Tested | What it found for Copyleaks |
|---|---|---|
| Orenstrakh et al. 2023 (preprint) | 2023 | 1 false positive in 114 computer science submissions written before ChatGPT; 95% of ChatGPT texts caught. On 10 QuillBot-paraphrased texts, accuracy fell from 100% to 50%. On Spanish human submissions it was wrong every time |
| Perkins et al. 2024 (preprint) | Autumn 2023 | The best of seven tools at catching AI text (65%), but it also flagged 5 of 10 human controls, the most of any tool |
| Alshammari & Rao 2025 (preprint) | March 2025 | No false positives on 49 human answers; scored humanized AI text at 71% on average |
Sources: Orenstrakh et al., Perkins et al., Alshammari & Rao.
What the evidence adds up to
- Plain English AI text: caught well in every study.
- Human writing: results range from almost no false positives to half of a small control set.
- Paraphrased or humanized text: detection dropped in both studies that tested it, which fits Copyleaks offering a separate setting for it.
- Other languages: the one independent test in Spanish went badly, while Copyleaks publishes high figures for six languages.
FAQ
Is Copyleaks accurate? On plain English AI text, yes. Its false positive rate is uncertain: Copyleaks gives four figures, and the small independent studies disagree.
Is Copyleaks free? You can scan up to 25,000 characters without logging in, and a free account adds limited monthly credits (Copyleaks).
Which sensitivity setting should I use? Copyleaks' own figures show the trade-off: the setting built to catch humanized text also flags the most human writing.
Can Copyleaks detect humanized text? Copyleaks says its Extra Sensitive setting is built for it. Independent tests found detection dropped on paraphrased and humanized text.