The short answer
The evidence is mixed. Copyleaks claims over 99% from internal tests on English text. In a 2026 peer-reviewed study of 160 English papers, it flagged none of 40 human papers. It also gave none of the 40 fully AI papers a full AI result, and it was right on only 30% of mixed papers. So a low score from Copyleaks did not mean a text was human in that test.
What Copyleaks claims
- Accuracy. Its home page says over 99%. A footnote says that rating is based on internal testing of English datasets.
- False positives. Its pages give different figures: 0.03% in one place and 0.2% in others. We could not tell which is current.
- Method. Its testing methodology page says its data science and QA teams tested the detector on data that differed from the training data.
- Non-native writers. A Copyleaks post reports 99.84% accuracy, with 12 of 7,482 texts wrong. That is the company’s own test.
- Third parties. It also says third-party studies rank it most accurate. Read any such study yourself. Results depend on the texts, as the next section shows.
- Languages. Its API documentation says AI detection works in 30 languages.
What published tests found
| Study | Texts | Copyleaks result |
|---|---|---|
| Van Vlasselaer et al., 2026 (peer reviewed) | 160 English papers in the social sciences, 40 of each type | Human: none flagged. Fully AI: 75% missed, 25% partly flagged. Mixed: 30% correct. AI text rewritten to sound like a student: 22.5% correct. |
| Alshammari and Rao, 2025 (preprint) | 49 human and 49 DeepSeek samples, plus 196 rewritten versions | Near-perfect on unedited and paraphrased text. 71% accuracy on text put through a rewriting tool. |
| Elkhatat et al., 2023 (peer reviewed) | 30 paragraphs from ChatGPT and 5 human paragraphs | Highest sensitivity of five tools on GPT-4 text, at 93%. Weaker on GPT-3.5 text. |
The 2026 study used ChatGPT Deep Research for its fully AI papers, and long academic text. It may be a harder test than short paragraphs, which could explain the gap with the older studies. The human papers came from non-native English-speaking graduate students, which makes the zero false positives more interesting, though 40 papers cannot prove a rate. The paper does not say whether the free or paid Copyleaks was used. In the same test, GPTZero, Turnitin and Pangram had different results, and the authors say detectors should not be the only evidence in high-stakes decisions.
Why a Copyleaks score is not proof
The studies above show the risk runs both ways. A high score can be wrong, and a low one can be too. For a direct look at Copyleaks and Turnitin, see Copyleaks vs Turnitin.
Three things hold for any AI detector, this one included:
- It reads the finished text. It does not see how the text was written, so it can only estimate.
- Its figures depend on the test. An accuracy number holds for the texts it was measured on. Your text may be shorter, edited or in another language.
- A small error rate is still real people. Liang and colleagues (2023) found seven detectors, which the paper does not name, flagged 61.3% of TOEFL essays by non-native speakers on average. Detectors have changed since, but plain, predictable writing remains the risk.
What to do with a result
- If you are a student. Keep your drafts and notes, and read your course policy. See how to prove you didn’t use AI.
- If you are a teacher. Treat a score as a reason to talk, not a finding. See how teachers check for AI.
Our AI detector gives a separate estimate for every sentence. It can disagree with Copyleaks. We publish no accuracy figure for it.
Sources
- Copyleaks (page text seen in search results, 1 November 2026): home page, testing methodology, non-native writers, third-party studies post and API documentation.
- Studies: Van Vlasselaer et al., Alshammari and Rao and Elkhatat et al.