Is Copyleaks Accurate? What Its AI Detector Claims, What Studies Found, and How to Read a Result

Is Copyleaks accurate? On its own tests, very: the company reports accuracy above 99% and a false positive rate well under 1%. Independent evaluations are fewer, older and smaller, and they show the same tool doing well on English essays, worse on paraphrased text, and badly on a small set of Spanish papers. This page sets the vendor's figures beside the two published studies that name Copyleaks, explains what each number counts, and shows how to read a Copyleaks result about your own draft or a student's without treating it as proof.

Is Copyleaks accurate? What the company says

A 99% headline tells you about the vendor's test set, not about your essay.

The Copyleaks AI detector page, read on November 2, 2026, advertises accuracy above 99% and "an industry-low .03% false positive rate." It lists per-language figures for its latest model, given as accuracy on human text and on AI text: English 99.97% and 99.20%, French 99.88% and 96.18%, Spanish 99.85% and 98.02%, Portuguese 99.95% and 93.08%. The same page says the detector supports over 30 languages, needs at least 255 characters on the web platform and 350 in the browser extension, and allows a free scan of up to 25,000 characters.

The more useful document is Copyleaks' own testing methodology page, which describes model V11, tested on September 10, 2026 and published on September 28, 2026. Two internal teams ran English-only tests on texts longer than 350 characters. The data science team used 300,000 human texts and 200,000 AI texts and reports a true positive rate of 0.989 and a true negative rate of 0.999. The QA team used 326,120 human texts, of which 923 were wrongly flagged as AI (0.9970 accuracy on human text), and 128,225 AI texts from four model families, of which 826 were missed (0.9936).

Notice the gap. The product page says .03%; the method page gives 0.30% for the default setting and 0.16% for the most cautious one. The product page does not say which test produced its figure, so the method page, with its counts, is the one to quote. Its human texts came from datasets the company assembled; the student-essay set (74,212 texts, 6 flagged) is not your class.

What independent studies found

Two published evaluations name Copyleaks and report its results separately. Both date from 2023 and tested older models than V11.

Orenstrakh, Karnalim, Suárez and Liut (2023). In Detecting LLM-Generated Text in Computing Education, posted to arXiv in July 2023, researchers from the University of Toronto Mississauga and two other universities ran eight detectors on 164 computing-course submissions: 124 written by students between 2016 and 2018, before ChatGPT existed, 30 generated by ChatGPT, and 10 generated and then paraphrased with QuillBot. Copyleaks ranked first. Using a 50% threshold, it classified 99.12% of the human data and 95.00% of the ChatGPT data correctly, 97.06% overall, and produced 1 false positive across the human submissions. On the 10 paraphrased papers its accuracy fell from 100% to 50%. On 20 Spanish submissions (10 human, 10 AI) it labeled every paper as AI, so its accuracy on the Spanish human papers was 0.00%. It appears to be the July 2023 arXiv study that Copyleaks' product page attributes to Cornell; arXiv is a preprint service, and the authors' universities are listed on the paper.

Walters (2023). The Effectiveness of Software Designed to Detect AI-Generated Writing, published in Open Information Science, compared 16 detectors on 126 documents: 42 essays generated by ChatGPT-3.5, 42 by ChatGPT-4, and 42 written without AI by students in a first-year composition course. The abstract reports that three detectors, Copyleaks, Turnitin and Originality.ai, had high accuracy with all three sets, while most of the others struggled to tell GPT-4 papers from student papers.

What a Copyleaks false positive means for you

A false positive is human writing labeled AI. At the default setting, Copyleaks' own figure means about 3 in every 1,000 human texts in its test sets were flagged. That sounds small until you multiply it: across a few thousand essays a term, even the vendor's figure predicts several wrong flags, and the small Spanish sample above suggests some writing fares worse.

Rates also depend on settings you may not see: the methodology table shows false positives rising about fourfold from Extra Safe to Extra Sensitive. What a percentage on an AI detection report means covers why a score is a probability about the text, not a measure of how much of it a machine wrote.

A worked example: reading a flagged paragraph

Here is an example paragraph written for this page, of the kind a first-year student might submit.

Suppose a detector marks this paragraph as likely AI. Ask four questions first. Is it long enough? At 43 words and about 300 characters it barely clears the 255-character web minimum, and short texts are where every detector is least sure. Is it generic? Short, list-like sentences with "moreover" and "therefore" are exactly what careful students are taught to write, and also what models produce. Is there evidence beside the score? Drafts, notes and version history matter more than the percentage. Does a second tool agree? Different detectors often disagree about the same paragraph.

Copyleaks' report helps here. Its product page says the AI Logic feature shows AI Phrases, with how often a phrase appears in AI rather than human text, and AI Source Match, which finds passages published elsewhere. A flag driven by stock phrases is weaker evidence than a passage found online.

How to read a Copyleaks result fairly

1. Check the length and language

   Note the character count and the language. Short texts and languages other than English are where the published evidence is weakest.

2. Ask which setting was used

   Extra Safe, Balanced or Extra Sensitive change the false positive rate by a factor of about four in the vendor's own table.

3. Read the highlighted passages yourself

   Look at what was marked and why. Stock phrases are weak evidence; matched published text is stronger.

4. Get a second opinion

   Run the same text through one more detector and compare. Agreement raises confidence a little; disagreement is a reason to stop.

5. Look at the writing record

   Drafts, notes and the document's version history say more about authorship than any score.

For the second opinion, some readers use Walter Writes AI's detector as a second check. Walter Writes AI is an AI humanizer with a built-in AI detector that supports over 80 languages, and every plan includes voice matching and the detector. Its result is one more estimate, not a verdict: no detector, Walter's included, can guarantee 100% accuracy, and two detectors disagreeing about a paragraph is common. Whatever any tool says, the text still needs your own read.

Pitfalls

• Quoting the headline instead of the method. Over 99% and .03% come from the product page; the counted results are on the methodology page.
• Assuming the studies tested today's model. Both independent evaluations are from 2023. Copyleaks now reports model V11.
• Treating a paraphrased draft as proof either way. The 2023 computing study saw accuracy halve after QuillBot. A low score on paraphrased text does not show human authorship.

Expert tips

• Quote figures with their base. "923 of 326,120 human texts" tells a reader what was counted; a bare percentage does not.
• Compare like with like. When choosing a tool for a class, the comparison of AI detectors for teachers lists what each vendor documents, and what any AI text detector can and cannot tell you covers the limits all of them share.

The QuillBot AI detector page covers a tool that reports its figures on a public benchmark, which makes a useful comparison.

Is Copyleaks accurate? By its own counts and in two independent studies, it is one of the stronger detectors available, and it publishes more about its testing than most. It is not infallible: its default setting flags a few human texts per thousand, and paraphrasing halved its accuracy in one study. Use it as one reading, check the setting and the length, and let the drafts decide. More on checking and editing AI writing is on the AI writing hub.