Copyleaks AI Detector: Its Own False-Positive Rate, .03% to .2%

What Is Copyleaks?

Copyleaks is an online integrity platform, and its two flagship products are an AI-content detector and a plagiarism checker. You paste or upload a document and it hands back a reading: an estimate of how much of the text looks machine-generated, a similarity score against other sources, or both at once. The audience is exactly who you would guess, and it is broader than most detectors reach: students giving a paper a last check before submitting, freelance and agency writers who get screened by clients or publishers, and the institutions, publishers, and HR teams that license the tool to screen work in volume.

Worth saying at the top, in Copyleaks’ favor: it does publish accuracy claims about itself, which plenty of consumer detectors never bother to do. Its product page states “over 99% accuracy and an industry-low .03% false positive rate” (EV-copyleaks-ai-detector-01), and its blog gestures at outside studies it treats as validation (EV-copyleaks-ai-detector-03). That openness is real, and it is more than several rivals offer. It is also the reason this page can exist at all: the moment a vendor commits to specific numbers, those numbers become checkable against one another and against the wider world, which is precisely what the sections below do.

Copyleaks Ltd: Who Builds It

The tool comes from Copyleaks Ltd, a company founded around 2015 that carries a public encyclopedia entry and a company page (both linked in this page’s source record). Unlike the anonymous consumer detectors I have reviewed elsewhere, this is a named, established B2B vendor with a documented corporate footprint, LMS integrations, and enterprise contracts behind it. That shapes how you should read its claims. A company with a real reputation and real customers to answer to is unlikely to be running a deliberate trick, which is the fair reading. It also means the bar such a vendor is held to is higher, so a headline error rate that says two different things on two of its own pages is a reasonable thing to press on.

AI Detector or Plagiarism Checker? Copyleaks Runs Both

Before a single accuracy number, the most useful thing to grasp about Copyleaks is that it is not one detector. It is two separate scoring systems sharing a brand, and mixing them up is the most common reason a flagged user cannot work out what actually happened.

System one is the AI detector. It estimates how much of a document was likely produced by generative AI and returns an AI percentage. System two is the plagiarism checker, the older and more familiar product, which compares your text against other online sources and a shared database and returns a Similarity Score. A false positive, the way this page uses the term, is human writing that one of those systems scores as machine-made when it was not. The two run independently, and Copyleaks says so plainly in its own FAQ document: “The AI percentage does not influence the Similarity Score calculation, nor does the Similarity Score change the AI percentage. They are independent metrics produced by separate analyses” (EV-copyleaks-ai-detector-11).

Which Score Actually Flagged You?

Here is where the real confusion lives. When an instructor or a client tells you “Copyleaks flagged your work,” that sentence by itself does not reveal which of the two systems fired. A high Similarity Score means your text matched existing sources, which is a citation-and-originality problem. A high AI percentage means the AI detector estimated machine involvement, a completely different problem with completely different evidence behind it. The two readings can even point opposite ways on the same document: work you wrote yourself can score low on similarity yet draw a high AI reading, and a heavily quoted piece can do the reverse.

So the question to settle first, before any argument about accuracy, is which number actually flagged you. If all you were told is that “Copyleaks flagged it,” ask to see the exact report and the exact percentage, because defending a plagiarism match and defending an AI reading are not the same job. Everything below about accuracy concerns the AI detector, not the Similarity Score.

Is Copyleaks Accurate? The Vendor Claim vs Independent Tests

Begin with Copyleaks’ account of itself. The product page claims “over 99% accuracy” (EV-copyleaks-ai-detector-01), and the company blog frames that as backed by outside researchers, leading with a study that scored Copyleaks 99.12% on human-written data and 95.00% on ChatGPT-generated data (EV-copyleaks-ai-detector-03). Taken alone, that is a strong self-report.

The outside record is messier, because independent tests of Copyleaks do not converge. They scatter across a wide band, and the scatter is itself the honest finding. Here is what I was able to open and verify this run, sorted by who ran each test and how much they disclosed.

Read that band as what it is: 53.4% to 99.12%, collected by different people on different dates against different models under different methods, and never a live side-by-side. A few rows carry caveats I would rather show you than smooth over. GPTZero’s 90.7% is a competitor’s test that naturally flatters its own tool, so weigh it as a rival’s number rather than a neutral one (EV-copyleaks-ai-detector-05). Supwriter sells an AI-writing-adjacent product, which makes its 77.5% a self-interested party’s result, and I flag it that way (EV-copyleaks-ai-detector-08). Global 100’s 94.7% comes off a page that never states its per-tool sample size, so it earns a directional reading, not a rigorous one (EV-copyleaks-ai-detector-06). The two rows I would trust furthest, the academic paper and the aidetector.ac benchmark, both disclose their methods and sample sizes, and they land at opposite ends of the band. That is the whole point: no two independent tests of Copyleaks that I could verify agree on how accurate it is.

One figure gets misattributed over and over, so it is worth settling here. GPTZero’s comparison lists a 0.24% false-positive rate, but that is GPTZero’s own claimed rate for its own tool, not Copyleaks’ (EV-copyleaks-ai-detector-05). Anyone handing you “0.24% for Copyleaks” is repeating a mix-up between two numbers printed on the same page. Copyleaks’ own false-positive figures are a separate matter, and they have a problem of their own, which the next section is entirely about.

V9 and AI Logic: Which Version Is Being Tested?

Part of why the band is so wide is that these tests are not all measuring the same Copyleaks. The webspero run tested version 7.1 in March 2025 (EV-copyleaks-ai-detector-09). The aidetector.ac benchmark ran in March 2026 (EV-copyleaks-ai-detector-07). The academic paper measured a 2023 build (EV-copyleaks-ai-detector-04). In between, Copyleaks’ own release notes show it launched AI Logic on May 23, 2025 and AI Detector V9 on June 24, 2025 for the web platform, with the LMS and API rollout landing on July 8, 2025 (EV-copyleaks-ai-detector-14). A stale detail worth correcting, because several reviews recite it: there was no “February 2026 AI Insights” release. “AI Insights” was rolled into AI Logic as a renamed sub-feature called “AI Phrases,” not a launch of its own (EV-copyleaks-ai-detector-14). So whenever you meet an accuracy figure for Copyleaks, ask which version and which date produced it, because a 2025 result and a 2023 result are not describing the same detector.

Copyleaks’ Own False-Positive Rate Doesn’t Agree With Itself

This is the finding that carries the page, and the one nobody else in this niche seems to have written down. Put the outside tests aside for a moment and look only at what Copyleaks says about its own false-positive rate, meaning how often it labels genuine human writing as AI. It gives two different answers on two of its own pages.

Both are Copyleaks’ own words. The product page leads with .03%, an industry-low figure it uses to sell the tool (EV-copyleaks-ai-detector-01). Its own official FAQ, a dated document hosted on copyleaks.com, says “less than 0.2%” (EV-copyleaks-ai-detector-02). That is roughly a sevenfold spread between two of the company’s own numbers for the single most important reliability figure a detector has. I am not claiming either one is the true value, and I am not saying the tool is broken. I am reporting that a reader trying to answer “how often does Copyleaks wrongly flag human writing” cannot get a straight answer from Copyleaks itself.

The same gap turns up in smaller ways. The two documents also give slightly different per-language accuracy pairs for English, with the product page and the FAQ printing figures that do not match (EV-copyleaks-ai-detector-01, EV-copyleaks-ai-detector-02). It is a minor version of the same pattern: the numbers on Copyleaks’ own pages have drifted apart over time and never been reconciled. If you are going to rest a decision on a vendor’s stated error rate, expecting that rate to read the same across the vendor’s own material is not much to ask, and here it does not.

Why the July 2023 Validation Study Is Getting Old

The other half of Copyleaks’ self-report is the “independent validation” it points back to, and that anchor is aging badly. The study Copyleaks headlines on its blog is a July 2023 arXiv paper that tested eight detectors on pre-ChatGPT student work and found Copyleaks the most accurate of the group, at 99.12% on human data (EV-copyleaks-ai-detector-03, EV-copyleaks-ai-detector-04). The paper is real, and I pulled it directly. Two things keep it from meaning what the marketing implies. First, that 99.12% is one table’s threshold-based figure; the same paper reports other, lower and inconsistent Copyleaks numbers on different sub-tests, so it is not a single clean result (EV-copyleaks-ai-detector-04). Second, and more to the point, a July 2023 study necessarily predates GPT-4o, Claude 3.5, and Gemini 2, all of which arrived after mid-2024. It has nothing to say about how Copyleaks handles the models students and writers actually use today. Copyleaks moved forward, shipping AI Logic and V9 in mid-2025 (EV-copyleaks-ai-detector-14), but the validation it keeps citing stayed behind. A present-day headline propped up by an old benchmark is not proof of a present-day error rate.

Does Copyleaks Flag ESL Writing or Short Essays More Often?

If you came to English later in life and keep watching sentences you wrote yourself get flagged, this is the part of the page written for you, and the part I weigh most carefully. The evidence here is solid, but it sits at the level of the whole detector category rather than at Copyleaks by name. In a peer-reviewed study led from Stanford, researchers ran seven commercial AI detectors against real TOEFL essays by non-native English writers; on average the tools wrongly tagged 61.3% of that genuine human writing as AI, while passing native-speaker essays nearly every time ( Liang et al., Patterns / Cell Press; EV-best-ai-humanizer-01, July 10, 2023). That 61.3% describes the set of detectors the study measured, and Copyleaks was not singled out within it. I will not repackage it as a Copyleaks figure, because it is not one, and I could not find a peer-reviewed benchmark this run that measures Copyleaks by itself on second-language writing. The careful version is narrow: Copyleaks belongs to a category with a documented bias against non-native writing, and nothing I found shows it either escaping that bias or being unusually bad at it.

The short-essay worry runs along the same lines. Detectors as a group grow less reliable on very short passages, roughly under 200 to 300 words, because a probability model simply has less text to read, and that is a general limitation reported across the field rather than a Copyleaks-only fault. If your flagged piece was short, carry that context into any appeal, alongside the documented cases below.

Two Documented False-Positive Cases

Beyond the category-level research, there are two specific, named, dated cases of Copyleaks scoring genuine human writing as AI, and specific beats generic every time.

The first comes from webspero.com, where reviewer Gursharan Singh ran a three-phase test of Copyleaks v7.1 in March 2025. In the third phase he fed the tool a passage that was 100% human-written, on the subject of large language models, and Copyleaks returned 100% AI. His reaction, verbatim: “CopyLeaks, what the heck? To all of our surprise, this 100% human-generated content was deemed AI” (EV-copyleaks-ai-detector-09). That same three-phase run is where the 53.4% overall figure in the table above comes from, so it is one study, not three.

The second carries more weight because of who ran it. ZDNet’s David Gewirtz, a named and credentialed technology journalist, ran his own confirmed-human writing through several detectors. Copyleaks scored it 100% AI, while Pangram, QuillBot, and ZeroGPT each read the very same text correctly as human across all five of his test blocks (EV-copyleaks-ai-detector-10, October 29, 2025). One journalist’s writing is no more a false-positive rate than webspero’s passage is. But two independent, documented instances of a fully-AI verdict on writing a person actually produced is exactly the failure a flagged student or freelancer dreads, and it is on the record.

Does Copyleaks Catch Humanized or Paraphrased AI Text?

Fair question, and it has two honest halves. The independent tests above suggest Copyleaks is uneven against paraphrased or edited AI text: the webspero run caught 60% of AI-paraphrased content in its first assessment but only 33.3% of human-edited AI content in its second (EV-copyleaks-ai-detector-09), and the aidetector.ac benchmark logged a 22% false-negative rate overall (EV-copyleaks-ai-detector-07). Those are two disclosed-method tests pointing at the same soft spot most detectors share, which is that rewritten machine text is harder for them than raw machine text. I would not inflate that into a flat “Copyleaks misses humanized AI,” since the samples are modest and the tests are not live, but the direction is consistent.

Does MeteGPT’s Own Humanized Text Pass Copyleaks? (FC-METEGPT-001)

Here I do have a first-party figure, and I am going to bury it in more caveats than the number deserves, precisely because it is my own tool and my interest in it is obvious. In a small, controlled in-house test from mid-May 2026, spanning about 30 academic-style passages, the machine-rewritten text our own humanizer produces scored 6% AI when I put it through Copyleaks (FC-METEGPT-001). Take that line literally and no further. It describes how our own engine’s output happened to read under Copyleaks on that particular set, on that particular day. It is not a ruling on how accurate Copyleaks is in general, and I refuse to stretch it into one. The caveats are the substance, not the footnote: the test is small and controlled, about 30 passages; unusual input like code or dense technical prose can push far higher on strict detectors; the data is ours, verifiable by the owner but not audited by an outside party; detector behavior and versions shift over time; and one reading on one sample on one date guarantees nothing about your document. I include it to show a tendency, not to advertise a pass rate, and it stays inside this scoped subsection rather than seeping into the general accuracy answer above.

Copyleaks vs Turnitin, GPTZero, ZeroGPT, and Originality.ai

People search these matchups wanting a clean ranking, and the honest answer is smaller than a ranking but more useful, because no independent same-samples, same-day test in the record I gathered pits these tools against each other head-to-head. What I can do is stand them beside one another on what is documented and send you to each tool’s own dated page.

Copyleaks vs Turnitin

These two serve different buyers, and that gap matters more than any single percentage. Copyleaks is self-serve: anyone can sign up and paste text in, and it carries a broad B2B and classroom footprint. Turnitin, by contrast, lives inside a school’s license and hands its AI indicator to the instructor rather than the writer. Turnitin’s reporting also has a quirk worth flagging: on a low reading it prints no percentage and highlights no text, leaving only an asterisk once the score falls below a 20% threshold, and that is why I treat it as an “under 20% AI” ceiling instead of a precise number (EV-turnitin-10). So if the system that will actually grade your work is your institution’s, a Copyleaks score is only a check you ran yourself, not the figure your school ends up seeing. That is the reason I keep Turnitin’s AI checker on a separate dated page, held to the same evidence rules for precisely that question.

Copyleaks vs GPTZero

The most-cited comparison here is GPTZero’s own. In its 3,000-sample test, GPTZero scored Copyleaks at 90.7% overall against its own claimed 99.3% (EV-copyleaks-ai-detector-05). Read that as a competitor measuring a rival, which naturally tilts toward the tool doing the measuring, not as a neutral benchmark. GPTZero makes its own accuracy claims that deserve the same scrutiny, and I put GPTZero’s own accuracy claims through the same check on a separate page.

Copyleaks vs ZeroGPT

ZeroGPT is a different tool from GPTZero despite the name, and the one place they overlap with Copyleaks in my record is the Gewirtz test: on his own human writing, Copyleaks returned 100% AI while ZeroGPT read the same passage correctly (EV-copyleaks-ai-detector-10). That is a single journalist’s five-block trial, not a rate, and ZeroGPT has gaps of its own that I document elsewhere, including no published false-positive number for itself. I cover ZeroGPT, and the places its own numbers disagree with each other on its own page.

Copyleaks vs Originality.ai

Originality.ai is the tool most freelance writers get judged by before they get paid, and it tends to run tuned to catch more AI at the cost of more false alarms. I could not verify a clean same-samples test pitting Copyleaks against Originality.ai this run, so I will not invent a head-to-head number between them. What holds is the posture: both publish confident accuracy claims, and both deserve to be read against the outside record rather than their own marketing. I put Originality.ai through the same self-consistency audit on its own dated page.

Who Actually Uses Copyleaks? Canvas, LMS, and Enterprise

One genuine differentiator for Copyleaks is reach beyond a single classroom system. It integrates with eight named learning management platforms, including Brightspace, Canvas, Google Classroom, Moodle, Blackboard, Edsby, Sakai, and Schoology (EV-copyleaks-ai-detector-12). Its original Canvas partnership with Instructure dates to October 2020, well before Copyleaks built an AI detector, and it began on the plagiarism-detection side of the product (EV-copyleaks-ai-detector-12). Keep that straight: the 2020 deal was a plagiarism-era integration, not a recent AI-detector announcement.

Beyond the Classroom: Publishers, Enterprise, and HR

Where Turnitin is largely an academic system, Copyleaks sells to a wider set of buyers: enterprises screening internal or vendor content, publishers vetting submissions, and HR teams checking applications. For a reader deciding which tool matters to them, the practical question is which system will actually judge their work. A student inside a Canvas course may be read by Copyleaks; a freelancer may be screened under a publisher’s Copyleaks contract; an applicant’s writing sample may pass through an HR deployment. The same detector, and the same accuracy questions this page raises, sits behind all of them, which is why the version-and-date scoping above matters wherever you run into it.

Is Copyleaks Free? Pricing for the AI Detector

Copyleaks’ live pricing page, captured on July 20, 2026, shows bundled plans that scan for both AI and plagiarism rather than a cheap AI-only option. The Personal plan is $16.99 a month, or $13.99 a month billed annually, and includes 100 credits, which Copyleaks describes as roughly 25,000 words or 100 images. The Pro plan is $99.99 a month, or $74.99 a month billed annually, with 1,000 credits (EV-copyleaks-ai-detector-13). Enterprise and Education pricing is custom, contact-sales only.

One correction, because it is everywhere. A cluster of third-party reviews still quote a $7.99 AI-only Copyleaks plan. That figure did not appear on Copyleaks’ live pricing page this run; it reads like stale copy recycled without a recheck (EV-copyleaks-ai-detector-13). If you saw $7.99 somewhere, confirm it against copyleaks.com before you plan around it, because the current structure is the $13.99 and $74.99 annual bundles above. The pricing page I captured did not list a standalone free plan either, so if you want a no-cost AI check to hold up against a Copyleaks result, the free detector we run at MeteGPT, described next, is one option, and there are others.

To keep my own stake in view instead of selling around it, let me name our ceiling rather than hide it. Our own detector at MeteGPT costs nothing and needs no sign-in, and an anonymous run is capped at 125 words. It is a deliberately small box, meant for a quick second opinion on a short passage rather than a whole-manuscript pass, and I am not going to pretend it is sharper than Copyleaks, since I never ran the two side by side. You can check a passage in our own free detector for an independent second reading.

Copyleaks Flagged My Work: What Should I Do Next?

If Copyleaks flagged writing you actually produced, the useful response is to build a record, not to spiral and not to hunt for a trick. One reading carries far less weight than the moment makes it feel like it carries, especially next to the two cases above where Copyleaks scored genuine human writing as fully AI (EV-copyleaks-ai-detector-09, EV-copyleaks-ai-detector-10). Here is the procedure that stands up.

First, find out which system flagged you. Ask to see the actual report and confirm whether it was the AI percentage or the Similarity Score, since they are separate readings with separate defenses (EV-copyleaks-ai-detector-11). Second, request the raw AI report itself, the full percentage and any highlighted passages, rather than accepting a secondhand summary of it. Third, hold onto your drafts, outlines, notes, and edit history; a trail showing how the piece came together over time is something no probability score can generate or refute, and if you write in a cloud document its version history is already keeping that trail for you, so do not clear it. Fourth, if you are assembling an appeal, the two documented false-positive cases on this page are citable precedent that this specific tool has scored real human writing as AI, and the peer-reviewed Stanford finding on non-native writing (EV-best-ai-humanizer-01) belongs in any appeal from a second-language writer. Fifth, get a second reading from a differently-built engine before anything turns formal, so a single tool’s verdict is never the whole basis for a decision; the same passage can go through a free detector such as ours and you can set the two independent readings side by side.

If what you actually want is to rewrite AI text so it stops getting flagged, that is a different task from checking one, and it is not what this page is for. None of the steps above are about slipping past a detector. They are about being able to defend real work when a number gets it wrong.

Is Copyleaks Worth Trusting? The Verdict

Should a single Copyleaks AI score be trusted on its own? No, and the reason deserves precision, because it is not that the tool has no value. The trouble is narrower: Copyleaks’ own figures never settle into anything steady enough to lean a decision on. Its published false-positive rate reads .03% on one of its pages and “less than 0.2%” in its own FAQ (EV-copyleaks-ai-detector-01, EV-copyleaks-ai-detector-02). Its “independent validation” rests on a July 2023 study that predates the models people use now (EV-copyleaks-ai-detector-03, EV-copyleaks-ai-detector-04, EV-copyleaks-ai-detector-14). Independent tests of it range from 53.4% to 99.12% with no agreement (EV-copyleaks-ai-detector-04 through EV-copyleaks-ai-detector-09). And two documented cases show it scoring genuine human writing as AI (EV-copyleaks-ai-detector-09, EV-copyleaks-ai-detector-10). That is not a case for avoiding Copyleaks, and it is not a claim that its detector fails. It is a case for treating any lone Copyleaks AI reading as a single unaudited data point, the kind that should have a person, a second tool’s opinion, and a documented draft trail standing behind it before it settles anything.

Credit where it is due, and this belongs on the record too: Copyleaks publishes its claims in the open, it does not sell an evasion product, and it says as much directly, arguing on its own blog that “fooling advanced AI detectors is an increasingly difficult and ineffective strategy” and that “responsible AI detection is defined by transparency, context, and ongoing innovation” (EV-copyleaks-ai-detector-15). That is a more honest posture than several tools in this niche take, and it is worth weighing against the inconsistencies above, not in place of them.

Common Questions

Is Copyleaks accurate? Copyleaks claims over 99% accuracy (EV-copyleaks-ai-detector-01), yet the independent tests I could verify run from 53.4% to 99.12% with no agreement (EV-copyleaks-ai-detector-04 through EV-copyleaks-ai-detector-09), and its own false-positive rate reads two different ways across its own pages, .03% and “less than 0.2%” (EV-copyleaks-ai-detector-01, EV-copyleaks-ai-detector-02). Read any one score as a single unaudited signal, not a verdict.

Is Copyleaks free? The live pricing page this run showed paid bundled plans, a Personal plan at $13.99 a month billed annually and a Pro plan at $74.99 a month billed annually, with no standalone free plan and no separate AI-only tier; the $7.99 figure some reviews still quote did not appear (EV-copyleaks-ai-detector-13).

Which score flagged me, AI or plagiarism? Copyleaks runs two separate systems, and a flag by itself does not tell you which one. The AI percentage and the Similarity Score are independent metrics from separate analyses (EV-copyleaks-ai-detector-11), so ask to see the specific report before you decide how to respond.

Is Copyleaks better than Turnitin? They serve different buyers, and no independent same-samples test in my record pits them head-to-head. Turnitin suppresses its numeric score below a 20% threshold, showing only an asterisk (EV-turnitin-10), and it runs inside your institution’s license, so if a school will judge your work, a Copyleaks result is a self-run signal, not the reading it will see.

Copyleaks flagged my essay, is that proof I used AI? No. One flag settles nothing on its own, particularly from a tool that has twice been documented scoring real human writing as AI (EV-copyleaks-ai-detector-09, EV-copyleaks-ai-detector-10). Find out which of its two systems flagged you, preserve your drafts and edit history, and run the passage through a second, differently-built detector before you accept one number as final.

Does Copyleaks detect humanized or paraphrased AI text? Not uniformly. Disclosed-method tests show it catching a majority of paraphrased AI in one assessment but only a third of human-edited AI in another (EV-copyleaks-ai-detector-09), and one benchmark logged a 22% false-negative rate overall (EV-copyleaks-ai-detector-07). Rewritten machine text is harder for detectors as a class, and Copyleaks is not exempt.