Is Copyleaks accurate?
Copyleaks topped two independent detector tests on unedited and lightly edited AI text, but it wrongly flagged 5 of 10 human samples in a 2024 study, and its headline accuracy figure comes from its own testing.
Short answer
Copyleaks is one of the stronger AI detectors in published tests, but not one to trust on its own. It classified all 126 documents correctly in a 2023 study and kept 58.7% accuracy after evasion edits in 2024, the best of seven tools. The same 2024 study found it wrongly flagged 5 of 10 human samples.
Is Copyleaks accurate? Copyleaks sells AI and plagiarism detection to schools and companies, and its free checker is one of the most used. Independent researchers have tested it twice, and both times it finished at or near the top on catching AI text. The open question is human writing: one small test found it accused human work far more often than its rivals. Every number below is dated and linked to its source, and vendor figures are labelled. For how we weigh sources, see how we score AI accuracy.
Copyleaks AI detector test results
- Independent study
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| Walters 2023, 126 documents | Documents classified correctly (GPT-3.5, GPT-4 and student essays) | 100%, tied 1st of 16 | Source 1Independent study | |
| Perkins et al. 2024 | Accuracy on unedited AI text plus human controls | 64.8%, 1st of 7 | Source 2Independent study | |
| Perkins et al. 2024 | Accuracy on AI text after evasion edits | 58.7%, 1st of 7 | Source 2Independent study | |
| Perkins et al. 2024 | Human control samples wrongly marked as AI (5 of 10) | 50% | Source 2Independent study | |
| Copyleaks homepageVendor's own figure, not independently checked | Detection accuracy, internal testing of English datasets | 99% | Source 4Vendor claim | |
| Copyleaks V11 methodologyVendor's own figure, not independently checked | True positive rate / true negative rate, 500,000 English texts | 0.989 / 0.999 | Source 5Vendor claim |
What the Copyleaks accuracy numbers mean
The independent record on AI text is good. Walters ran 16 detectors on 42 essays written by GPT-3.5, 42 by GPT-4 and 42 by first-year students, and Copyleaks and Turnitin were the only two that got every document right (Walters, 2023-10-06). Most other detectors could tell GPT-3.5 essays from student work but struggled with GPT-4.
Perkins and colleagues tested seven detectors in early 2024 and then reworded the AI text with simple evasion techniques. Copyleaks led on unedited text at 64.8%, just ahead of Turnitin at 61%. After the edits it still scored 58.7%, down from 73.9%, while the seven tools averaged 22.2%. That makes it the most robust detector to light editing in that test.
The weak point is human writing. In the same Perkins study, Copyleaks marked 5 of its 10 human control samples as AI, the highest false accusation share of the seven. Turnitin and ZeroGPT marked none. The human sample was small, so treat the figure as a warning, not a rate.
Copyleaks' own numbers are much higher. Its homepage claims 99% AI detection accuracy, with a note that the figure comes from internal testing of English datasets. Its methodology page reports a true negative rate of 0.999 for its V11 model on 300,000 human texts (Copyleaks, 2026-09-28). Both are vendor claims on the vendor's own data, and the independent tests above used older versions.
Where Copyleaks goes wrong
Human writing flagged as AI
It wrongly flagged 5 of 10 human samples in Perkins et al. (2024), the worst of seven detectors.
Edited AI text
It held up best, but still fell from 73.9% to 58.7% once the text was reworded, so about four in ten edited AI samples got through.
Old independent data
The academic tests date from 2023 and early 2024. Copyleaks' current V11 model has only been tested by Copyleaks.
No fairness test
Copyleaks was not among the seven detectors in the Liang et al. study of non-native English writers, so there is no independent figure for how it treats them.
How Copyleaks compares in the same tests
Only compare detectors inside one study.
| Test | Copyleaks | Others in the same test |
|---|---|---|
| Walters 2023, documents classified correctly | 100% | Turnitin 100%, Originality.ai 98%, ZeroGPT 87%, GPTZero 81% |
| Perkins et al. 2024, unedited AI text | 64.8% (1st of 7) | Turnitin 61%, ZeroGPT 46.1%, GPTZero 26.3% |
| Perkins et al. 2024, after evasion edits | 58.7% (1st of 7) | ZeroGPT 17.3%, GPTZero 16.7%, Turnitin 7.9% |
| Perkins et al. 2024, human samples wrongly flagged | 50% (5 of 10) | Turnitin 0%, ZeroGPT 0%, GPTZero 10% |
Copyleaks beat Turnitin, ZeroGPT and an older GPTZero on AI text in both studies, and fell behind all three on human samples in the second. Originality.ai was only in the first study. For the full picture, see are AI detectors accurate.
What a Copyleaks score can and cannot prove
No detector score should be the only evidence in a misconduct case.
It can show
- A Copyleaks score can tell you that a text reads like model output to its detector. In the published tests it caught AI text more often than most rivals, so a high score is a reason to look closer.
It cannot prove
- A Copyleaks score cannot tell you who wrote the work or how AI was used. One independent test found it flagged half of the human samples it saw, so a flag on human writing is a real possibility.
What to do if Copyleaks flags your work
Ask for the full Copyleaks report, including the score and the highlighted passages.
Collect your drafts, outline, notes and version history from Google Docs or Word.
Point to the 2024 study in which Copyleaks flagged 5 of 10 human samples.
If English is not your first language, cite the Liang et al. findings on detectors and non-native writers.
Offer to talk through your argument and sources, or to write a short sample under supervision.
Ask for the appeal process in writing, and ask that any decision rest on more than one piece of evidence.
Another detector to try
Another option is the free AI detector from Walter Writes. Its page says checks are free and unlimited, that it detects text from ChatGPT and Gemini, that it gives a percentage score for every scan and that it works in more than 20 languages. You can try Walter's AI detector here. A second detector's score is one more signal, not proof either way.
Sources
- Independent study
- Independent study
- Independent study
- Vendor claim
- Vendor claim
Questions
How accurate is Copyleaks?
On AI text, among the best in published tests. It classified all 126 documents correctly in a 2023 study of 16 detectors and led seven detectors in 2024 at 64.8% on unedited text and 58.7% after evasion edits. Its own 99% figure comes from internal testing.
Does Copyleaks give false positives?
Yes. In Perkins et al. (2024), Copyleaks marked 5 of 10 human samples as AI, the highest share of seven detectors, though the sample was small. Copyleaks reports a true negative rate of 0.999 on its own data. Keep your drafts so you can show your process.
Is Copyleaks more accurate than Turnitin?
In the two studies that tested both, they were close on AI text. Both got all 126 documents right in Walters (2023). In Perkins et al. (2024), Copyleaks scored 64.8% to Turnitin's 61% on unedited text and 58.7% to 7.9% after edits, but flagged 5 of 10 human samples where Turnitin flagged none.
Can Copyleaks detect paraphrased AI text?
Better than most detectors, not reliably. After simple evasion edits, Copyleaks kept 58.7% accuracy, the best of seven tools in a 2024 study, while the average fell to 22.2%. About four in ten edited AI samples still got past it.
Should a school rely on a Copyleaks score?
Not on its own. Copyleaks caught AI text well in independent tests, but one study found it flagged half of the human samples it checked. Use a score to start a conversation with the writer, and look at drafts and version history before deciding anything.
