Guide

Copyleaks AI Detector in 2026: How It Works, How Accurate It Is, and What It Misses

A plain-English 2026 guide to the Copyleaks AI Detector: what the 99% accuracy and 0.03% false-positive claims are measured on, why the published methodology documents a model Copyleaks has since replaced, what the largest peer-reviewed test found on long academic papers, how the three sensitivity levels change the score your marker sees, what AI Logic and AI Source Match actually do, what the terms of use disclaim, and why this detector fails by missing rather than by accusing.

September 28, 2026 · Updated September 29, 2026 · The AI Rankings

Quick answer: The Copyleaks AI Detector is a commercial AI-text detector sold to schools, universities and businesses, and usable free for up to 25,000 characters without an account (Copyleaks). Its published claim is over 99% accuracy with a 0.03% false positive rate, and Copyleaks publishes a dated testing-methodology page behind that number (Copyleaks testing methodology). Two facts qualify that claim. The methodology documents model V10, tested on 16 October 2025 — and Copyleaks shipped AI Text Detection model V11.0 on 22 September 2026, with no published evaluation of the new model (Copyleaks release notes). And the largest peer-reviewed test of Copyleaks on long-form academic work found the opposite failure to the one most people fear: across 40 fully AI-generated master’s-level papers of at least 4,000 words each, Copyleaks recorded 0% strict accuracy and underestimated AI content by a mean of 85.97%, while clearing 100% of the human-written papers with no false positives at all (Van Vlasselaer, Van Droogenbroeck and Spruyt, International Journal for Educational Integrity, 29 June 2026). That is the shape of this tool. Copyleaks is the detector least likely to accuse an innocent writer and among the most likely to clear a guilty one — a profile that makes it defensible for institutions and close to useless as proof of anything.

This guide covers the Copyleaks AI Detector specifically: the product, the score, the evidence and the company. For the underlying detection mechanism across all tools, see how AI detectors work. For ranked alternatives with independent accuracy scores, see best AI detectors. For the institutional detector most students are actually assessed by, see Turnitin AI detection. Copyleaks also sells plagiarism detection, which is a different product with a different evidence base — that side is covered in best plagiarism checkers.

What the Copyleaks AI detector actually is

Copyleaks began as a plagiarism-detection company in 2015 and added AI-text detection later. The AI detector is now the front door to the business, and four facts about the product shape everything below.

It is an institutional product with a free consumer entrance. Copyleaks sells to universities, schools, publishers and enterprises, and integrates with learning management systems. But anybody can paste up to 25,000 characters into the web detector without logging in and get a percentage back (Copyleaks). The number a student sees on the public page and the number an instructor sees inside Canvas are produced by the same engine at potentially different settings — which matters, and is covered below.

It reports a score plus an explanation, not a score alone. Copyleaks brands this layer AI Logic, and it has two components: AI Phrases, which highlights the specific sentences whose linguistic patterns drove the score, and AI Source Match, which flags where flagged text matches content already published on AI-generated websites (Copyleaks). Source matching sits alongside the classifier rather than replacing it.

Detection sensitivity is a setting, not a constant. Copyleaks exposes three detection levels: level 1 catches text copied straight from a model with no edits, level 2 catches text with minor changes such as tense adjustments or added words, and level 3 catches heavily modified AI text (St Petersburg College faculty documentation). The same document notes that at that institution the Canvas default is level 2, and that an instructor can only change the setting before students submit.

It is one product in a suite. Copyleaks sells AI text detection, AI image detection, AI video detection (announced 2 June 2026), a plagiarism checker, and Codeleaks for AI-generated and plagiarised source code (Copyleaks). The accuracy claims on this page concern the text detector only and should not be read across to the others.

The detection model changed on 22 September 2026, and the published evidence did not

This is the most important current fact about Copyleaks.

Copyleaks’ published testing methodology documents model V10. The page states the test was run on 16 October 2025 and published on 12 November 2025. It reports results from two separate teams: a data-science evaluation on 300,000 human-written and 200,000 AI-generated texts, and a QA evaluation on 229,843 human-written and 18,712 AI-generated texts, all above the 350-character product minimum (Copyleaks testing methodology).

Copyleaks replaced that model on 22 September 2026. The release notes record “AI Text Detection Model V11.0” shipping on that date, stating that the release “improves detection accuracy and extends coverage to recently released large language models,” and that the new version now appears in API responses as modelVersion: v11.0 (Copyleaks release notes).

No evaluation of V11.0 has been published by Copyleaks or by anyone else. As of 28 September 2026 the testing-methodology page still documents V10, and we could find no independent test of V11.0. Every accuracy figure on this page, vendor or independent, describes a model that is no longer the one running.

Two things follow, and they pull in opposite directions. Publishing a version number at all puts Copyleaks ahead of most of this category: ZeroGPT publishes no version history, so a ZeroGPT score cannot be tied to the system that produced it (ZeroGPT AI detection). But a headline “99% accuracy” advertised on the product page while the methodology behind it describes a superseded model is a claim about the past presented in the present tense. If you are relying on a Copyleaks score today, the honest position is that its current accuracy is data not available.

How Copyleaks’ detector works

Copyleaks publishes more about its method than most consumer detectors and less than Turnitin.

It classifies per section and aggregates. The API returns “per-section classification (human or AI), an overall human-versus-AI summary, and AI Logic explanations” (Copyleaks API documentation). The web product presents the same thing as a document-level percentage with AI-attributed phrases highlighted.

AI Phrases is a frequency argument, not a semantic one. Copyleaks describes it as identifying the specific sentences whose linguistic patterns are statistically characteristic of model output. That is the same broad family of signal every classifier-based detector uses, and the mechanism is set out properly in how AI detectors work.

AI Source Match is genuinely different, and it is not detection. It checks flagged passages against content already published on AI-generated websites — in effect running the plagiarism engine against the AI-content web. When it fires, it gives something a classifier cannot: a named external source. When it does not fire, it says nothing either way. Treating a Source Match hit and a classifier score as the same kind of evidence is a mistake, because only one of them is checkable.

It claims to catch modification, at a cost. Copyleaks says the detector identifies “mixed text where human-created text has been amended with AI-generated text” and “paraphrased content, including text that has been processed through spinners or modified with deliberate typos” (Copyleaks AI Detector FAQs), and the API documentation adds detection of character swaps and hidden characters. The independent evidence below shows this claim holds unevenly and fails hard against one specific paraphrasing tool.

What Copyleaks does not publish is the architecture. There is no model card, no released evaluation set and no training-corpus description. The methodology page describes how the test was run, not what was tested. That is a meaningful step up from a vendor that publishes nothing, and short of what would let anyone reproduce a result.

What Copyleaks will and will not scan

The 255-character floor is very low, and that is a risk rather than a feature. Copyleaks will return a percentage on roughly 40 words. Every published accuracy figure, vendor and independent, was measured on text far longer than that — the vendor’s own test set used 350 characters as a minimum and the peer-reviewed 2026 study used papers of at least 4,000 words. A Copyleaks score on a paragraph is not a weaker version of a real result; it is outside the range anyone has measured.

The language gap is between detection and explanation. Copyleaks claims detection across 30-plus languages but offers AI Logic phrase-level explanations in six. In practice that means a marker looking at a flagged Japanese or Arabic submission gets a number without the sentence-level evidence that makes the number arguable. Every independent study cited on this page tested English.

How to read a Copyleaks AI score

A percentage is not a proportion of your document. It is an aggregate of section-level classifications. This is the single most common misreading of every detector of this type.

The same text can score differently at different sensitivity levels. Because the level is an institutional or instructor setting, two markers at two universities can run the identical essay through Copyleaks and see different results without either of them doing anything wrong. If a score is being quoted at you, the sensitivity level it was produced at is a fair question, and one most people asked will not know the answer to.

The AI score and the similarity score are separate, unrelated numbers. Copyleaks states this plainly: the similarity score shows matching against online sources, the AI percentage estimates generative-AI involvement (Copyleaks AI Detector FAQs). A 0% similarity score and an 80% AI score are not contradictory, and neither is the reverse.

Read the AI Phrases, not the headline. The highlighted sentences are the only part of a Copyleaks result you can actually interrogate.

Which AI models Copyleaks says it can detect

Copyleaks names ChatGPT, GPT-5, Gemini, Claude, DeepSeek and Llama on the detector page, and the V10 QA evaluation covered text from “OpenAI family models,” GPT-5, Gemini family models, Claude family models and Grok family models (Copyleaks testing methodology).

Three qualifications belong with that list.

Naming a family is not publishing a per-model rate. Copyleaks reports aggregate true-positive rates across its whole AI set, not a figure per model. Per-model detection rates are data not available for every model named.

The V11.0 release explicitly extends coverage to newer models, which concedes that V10 did not have it. “Extends coverage to recently released large language models” is the vendor stating that the model behind the published 99% figure did not cover the current generation.

One independent result on a specific model is unusually strong. Alshammari and Rao, in March 2025, reported Copyleaks at 100% accuracy on human text and 99.7% on DeepSeek output — a figure Copyleaks cites itself (Copyleaks). It is a narrow test on one model and it is a genuine data point.

What Copyleaks claims about accuracy

There are three different vendor false-positive figures on three different Copyleaks surfaces, and the spread between them is an order of magnitude.

The measured V10 numbers are more informative than the headline, and Copyleaks deserves credit for publishing them. The data-science evaluation reported a true-positive rate of 0.988 and a true-negative rate of 0.999; the QA evaluation reported 0.9997 accuracy on human text and 0.992 on AI text (Copyleaks testing methodology). These are internally consistent with the 99% headline, and they were produced on the vendor’s own test set at the default balanced sensitivity.

Copyleaks also publishes per-language accuracy, which almost nobody in this category does. English is given as 99.97% on human text and 99.20% on AI text; French 99.88% and 96.18%; German 99.94% and 95.63%; Italian 99.88% and 97.00%; Portuguese 99.95% and 93.08%; Spanish 99.85% and 98.02% (Copyleaks). Note the pattern: human-text accuracy barely moves across languages while AI-text accuracy drops nearly seven points from English to Portuguese. The detector’s conservatism travels; its sensitivity does not.

What is missing is the composition of the test set. An accuracy figure on a detector is a function of which models generated the AI half, how long the passages were, and how much of the text was paraphrased or edited. Copyleaks publishes the counts and the minimum length; it does not publish the length distribution, the prompt set, or the proportion of hybrid and humanised text. That omission is exactly where the independent evidence diverges from the claim, as the next two sections show.

What Copyleaks’ list of “third-party studies” actually contains

Copyleaks backs its headline with “99% accuracy backed by independent third-party studies” and maintains a page listing them. It is worth reading the list rather than the claim about the list.

The page cites 17 items (Copyleaks). By Copyleaks’ own descriptions of them, they break down as follows.

Three observations, stated neutrally.

“Third-party” is doing a lot of work. A bachelor’s thesis and a press-wire editorial team are third parties. They are not independent evaluations in the sense a reader of “backed by independent third-party studies” would reasonably assume. Several of the strongest-sounding results come from the weakest-standing sources — the “only tool to correctly identify AI across all models” line is attributed to a press-wire editorial team, and one “100% accuracy” result is described as being on four test documents.

The sample sizes are mostly tiny. Where Copyleaks’ summary names a sample, the numbers are in the single or double digits of documents. The one recent study with a large, purpose-built corpus — 160 papers of at least 4,000 words each — is not on the list, and it is the one that disagrees.

The page is maintained without a visible revision date. It carries a publication date of 8 July 2025 while citing studies published in January 2026, so it is evidently updated in place. That is normal practice and it means the “as of” date on the claim is unknowable.

We have not read all 17 of these papers, and we are not disputing any individual figure in them. The point is narrower and it is about what the sentence on the product page implies: a list assembled by the vendor, weighted towards small favourable tests and including non-academic sources, is vendor-selected evidence. It is not the same thing as a literature.

What independent testing actually found

Four measurements of Copyleaks are worth knowing, and they do not tell one story.

The 2026 study is the one that matters most, and it is unambiguous about the direction of the failure. Across 40 fully AI-generated master’s-level papers, Copyleaks classified every one as a false negative under strict scoring, underestimating the AI content by a mean of 85.97 percentage points. Across 40 human-written papers it produced no false positives at all. On hybrid papers it reached 30% and on humanised AI 22.5% (Van Vlasselaer et al.). The same study’s best performer, Pangram, reached 65% to 97.5% on AI-generated text. The authors’ conclusion applies to the whole category: detection tools “should not be used as sole evidence in high-stakes decision-making but should be implemented in a broader evaluation strategy.”

The QuillBot result is the sharpest single number on this page. Kar and colleagues found Copyleaks scored an unedited AI article at 100% AI, a Grammarly-paraphrased version at 100%, a ChatGPT-paraphrased version at 100% — and a QuillBot-paraphrased version of the same text at 0% (Kar et al.). That is a total collapse triggered by one specific tool while two others left the score untouched. It was a single sample of about 500 words, so treat it as a signal rather than a rate. The direction matches every study of machine-paraphrased text across the category, covered in best AI humanizers and best AI paraphrasing tools.

Copyleaks has plainly changed since the 2023 tests, and in the direction you would want. Elkhatat and colleagues found it produced false positives and uncertain calls on human control text in 2023; the 2026 study found zero false positives across 40 human papers. That is real improvement on the metric that hurts people most. It has come with, or alongside, a large loss of sensitivity on long AI text.

One clarification, because the name similarity invites an error. The widely cited 2023 study by Weber-Wulff and colleagues, which tested 14 detectors and is the source of many circulating false-positive figures, did not test Copyleaks (Weber-Wulff et al.). Copyleaks appears in that paper’s related-work table only. Any page attributing a Weber-Wulff false-positive rate to Copyleaks is wrong, and we are stating it plainly so nobody makes that error from here.

The real Copyleaks problem is false negatives, not false positives

Almost all public argument about AI detectors is about wrongful accusation. Copyleaks is the tool where that framing fits worst, and understanding why is the most useful thing on this page.

On the best available evidence, Copyleaks does not over-flag human writing. Zero false positives across 40 human-written master’s-level papers in the 2026 study; a measured 0.026% false-positive rate on 229,843 human texts in the vendor’s own V10 QA run; per-language human-text accuracy between 99.85% and 99.97%; no false positives in Scribbr’s comparative test. Four separate measurements, two independent, pointing the same way.

The cost of that conservatism is that it misses. A detector tuned to almost never call a human “AI” will inevitably call a good deal of AI “human.” The 2026 study quantified it: 0% strict accuracy on fully AI-generated long papers, and a mean underestimation of 85.97%. A 4,000-word paper written entirely by a model came back looking substantially human.

This inverts the practical advice for both sides. For an institution, a low false-positive rate is the property that makes a tool defensible in a disciplinary process, and Copyleaks has strong evidence for it on both vendor and independent measurement. For anyone hoping a Copyleaks pass means a document is human-written, that inference does not hold at all — a clean Copyleaks score on a long document is weak evidence of anything.

Copyleaks concedes the one fairness point that most vendors deny. Its own FAQ acknowledges that non-native English speakers face a higher risk of misclassification (Copyleaks AI Detector FAQs). That is unusual candour, and it matters: the systematic disadvantage detectors of this class impose on second-language writers is the best-documented harm in the field, and is covered in how AI detectors work.

There is a documented case of Copyleaks flagging, then clearing, the same paper. In April 2024 Marley Stevens, a student at the University of North Georgia, was accused of cheating after using Grammarly’s grammar-checking features. Turnitin flagged the paper and Copyleaks initially flagged it as bot-written; when Stevens later ran the same paper through Copyleaks herself, it came back human-written. Copyleaks did not respond to requests for comment. She lost a scholarship, was placed on academic probation and paid $105 for a required cheating seminar (EdSurge, 4 April 2024). The case is the clearest illustration available that two runs of the same tool on the same text are not the same evidence, and that grammar-checking software sits inside the blast radius — see best AI grammar checkers for what those tools actually do to your text.

What Copyleaks’ own terms of use say about accuracy

The contract is the document that would matter in a dispute, and Copyleaks’ is unusually direct about the AI detector specifically.

It disclaims accuracy of the AI features by name. The terms, last updated 24 August 2026, state that “Copyleaks does not provide, and expressly disclaims, any warranty regarding the accuracy of the result provided by the AI Features” (Copyleaks terms of use).

It states that the detector does not determine whether content was AI-generated. The same document: “The AI Features are an assessment and diagnostic tools which does not guarantee whether any given content was AI Generated.” That is the vendor describing its own product as diagnostic rather than probative, in the binding document, while the product page advertises over 99% accuracy. Both statements are standard practice in their respective genres. Anyone proposing to use a Copyleaks score in a formal process should know which of the two is enforceable.

Liability is capped at the fees you paid in the last twelve months. For a free user that is zero. The terms are governed by New York law with exclusive jurisdiction in New York courts, and name the operating entity as Copyleaks, Inc., at 115 E 23rd St, 7th Floor, New York, NY 10010.

Copyleaks in the LMS: who sees the score

This is where Copyleaks differs most from a consumer detector, and where the practical consequences sit.

It runs inside the major learning management systems. Copyleaks launched AI Logic across Canvas, D2L Brightspace, Moodle, Blackboard, Schoology, Edsby and Sakai on 29 July 2025 (Copyleaks). Where an institution has it enabled, a submission can be scanned automatically on upload and the result appears in the instructor’s view.

The sensitivity level is chosen upstream of you. An administrator or instructor sets the level, and at least one institution’s documentation says it can only be changed before submission. The level is not part of the headline percentage, so it is worth asking which one was used.

Copyleaks’ own guidance is that the score starts a conversation. Its FAQ says “data provided by AI detectors should be used to inform the situation and offer the option for a learning opportunity,” and that detection works best “when paired with human judgment and used as part of a broader content evaluation process,” explicitly discouraging reliance on the tool alone for high-stakes decisions. That is the vendor agreeing with the peer-reviewed literature. Institutions that treat a Copyleaks percentage as a finding are going further than Copyleaks does.

Copyleaks says it serves millions of users across enterprise and education and appears on the Inc. 5000 list. A precise count of institutional customers is data not available; Copyleaks publishes no figure and we found none in a primary source.

Plagiarism checking and Codeleaks are separate products

Copyleaks is frequently described as a plagiarism checker that added AI detection, and the two are worth keeping apart because their evidence bases are not the same.

The plagiarism engine is older, more mature and differently evaluated. It matches submitted text against published sources across more than 100 languages, including where text has been edited or paraphrased. Nothing on this page speaks to its performance. If you are choosing a tool for source matching rather than AI detection, best plagiarism checkers is the page for that comparison, and the similarity score it produces is a different number from the AI percentage discussed here.

Codeleaks is the source-code product. It detects AI-generated and plagiarised code and identifies software licences. No independent evaluation of it appears in the literature above — the one study Copyleaks cites on code detection is a student thesis reporting that it exceeded an 80% accuracy requirement. For reviewing AI-written code on its merits rather than its provenance, best AI code review is the relevant comparison.

The image and video detectors are newer than any published test. AI image detection gained confidence heatmaps on 23 September 2026 (Copyleaks release notes); AI video detection was announced on 2 June 2026. For the state of image detection generally, including how badly the category performs on authentic photographs, see best AI image detectors.

What it costs

Prices below were read from Copyleaks’ pricing page on 28 September 2026 and are in USD.

(Copyleaks pricing)

The credit system is the thing to understand before you buy. Credits are shared across text and image scanning, so an image-heavy month eats the word allowance. The annual plans are not simply discounted monthly plans — annual Personal carries twelve times the monthly credit allowance for roughly ten months’ worth of money, which makes the monthly plan poor value for anything beyond an occasional check.

Institutional pricing is not published and we are not estimating it. Education and Enterprise quotes depend on seat counts and integration scope. Any USD figure you see for Copyleaks institutional licensing on a third-party directory is an unsourced guess.

How Copyleaks compares with the tools it gets compared with

Read that as a warning about the category, not a leaderboard. The June 2026 comparison found that every tool it tested handled human writing well and that most of them failed badly on fully AI-generated academic papers (Van Vlasselaer et al.). Full rankings with current scores and pricing are on best AI detectors. The longer-term alternative to detection — provenance signals attached at generation rather than inferred afterwards — is covered in what is AI watermarking.

What to do, by situation

If you are a student whose work has been flagged by Copyleaks. Two facts are on your side and worth stating calmly. Copyleaks’ own terms disclaim any warranty as to the accuracy of its AI results and describe the feature as diagnostic rather than determinative. And its own guidance says the score should inform a conversation, not settle one. The strongest material you have is process evidence: version history from Google Docs or Word, timestamped drafts, outlines, notes and reading. Ask which sensitivity level the scan ran at and whether AI Source Match returned a named source or only a classifier score — those are different quality of evidence, and only the first is checkable.

If you are a teacher or administrator choosing a detector. Copyleaks has a well-evidenced false-positive record, which is the property that matters if your process can ruin somebody’s year. Go in knowing what you are buying: on the only large peer-reviewed test of long-form work it missed every fully AI-generated paper and underestimated AI content by 86 percentage points on average. It is a screening tool that will rarely accuse the innocent and will frequently clear the guilty. Best AI for teachers covers the assessment-design approaches that survive the detector problem, which is where the durable answer lies.

If you are a writer or freelancer checking your own work. The free 25,000-character scan is enough, and Copyleaks’ conservatism works in your favour here: a flag on your own prose is more likely to be meaningful than a flag from a tool with a higher measured false-positive rate. Read the AI Phrases rather than the percentage and treat a highlight as a note that a passage reads formulaically.

If you are a publisher or content team screening submitted work. AI Source Match is the feature to test, because a named source is worth more than a probability. Check it against material you know the provenance of before you build a workflow on it, and assume the classifier will pass long, edited AI text.

If you are deciding whether to pay. Test the free tier on documents whose origin you already know, including one long AI-written piece, before spending anything. The current model, V11.0, shipped on 22 September 2026 without a published evaluation, so your own test on your own content type is better evidence than any figure on this page.

This page describes what the tool does and how reliable the evidence for it is. It does not explain how to get a particular result from it.

Who runs Copyleaks

Copyleaks, Inc. was founded in 2015 by Alon Yamin and Yehonathan Bitton, and is headquartered at 115 E 23rd St, 7th Floor, New York, NY 10010 (Copyleaks terms of use). Yamin is chief executive and Bitton chief technology officer.

Disclosed funding is small for the category. Copyleaks raised a $6 million Series A in April 2022 led by JAL Ventures with participation from Connecticut Innovations, bringing total disclosed funding to $7.8 million including a $1.8 million seed round. At that point the company had around 20 full-time staff, roughly half in Israel (Calcalist, 19 April 2022). Any funding since then is data not available.

The company is more findable than most of this cluster. It publishes a press-release archive, names its executives, maintains public API documentation with dated release notes, and its chief executive speaks on the record in trade coverage. There is no disclosed corporate parent. By contrast, Superhuman, the company formerly named Grammarly, announced on 23 June 2026 that it would buy GPTZero.

Where this leaves Copyleaks

Copyleaks publishes more about its detector than most vendors in consumer and institutional AI detection, and its headline number is also far from what independent long-form testing measures. Both of those are true at once, and neither cancels the other.

What it does well is real. It publishes a dated testing methodology, per-language figures, a version number in its API responses and release notes you can read. It concedes the non-native-speaker risk its competitors elide. Its false-positive record is strong on both vendor and independent measurement, and the peer-reviewed 2026 study found it cleared every human-written paper it was given. If your worry is a student wrongly accused, Copyleaks is the least likely of these tools to do it.

What it does badly is also real, and it is the mirror image. On the only large peer-reviewed test of long academic work, Copyleaks caught none of the fully AI-generated papers under strict scoring and underestimated AI content by a mean of 85.97%. One paraphrasing tool took a 100% score to 0% in a single pass. The 99% accuracy advertised on the product page is measured on the vendor’s own corpus, and it describes model V10 — which Copyleaks replaced on 22 September 2026 with a model nobody has yet evaluated. And the terms of use expressly disclaim any warranty of accuracy for exactly the feature the marketing quantifies.

The honest summary is that Copyleaks is a well-documented conservative screening instrument and not a source of proof about a person. Used as a prompt to look harder at a piece of work, it is among the better tools available. Used as a finding, it is indefensible — and Copyleaks, in its own contract and its own guidance, says so.

Frequently asked questions

Is Copyleaks accurate?

It depends entirely on what you are asking it to do. Copyleaks advertises over 99% accuracy and a 0.03% false-positive rate, and its published testing methodology supports those figures on its own corpus of roughly 500,000 texts. On human-written text the independent evidence agrees: the peer-reviewed study published in the International Journal for Educational Integrity on 29 June 2026 found Copyleaks produced zero false positives across 40 human-written master’s-level papers. On AI-written text the same study found the opposite, recording 0% strict accuracy on 40 fully AI-generated papers and a mean underestimation of AI content of 85.97%. Copyleaks is accurate at not accusing innocent writers and inaccurate at catching long AI-generated work.

Does Copyleaks detect ChatGPT?

It says so, and the evidence is mixed by text length. Copyleaks names ChatGPT, GPT-5, Gemini, Claude, DeepSeek and Llama on its detector page, and its own V10 test set included output from OpenAI, Gemini, Claude and Grok model families. A 2024 study in the Indian Journal of Psychological Medicine scored an unedited ChatGPT-written article at 100% AI. But the 2026 peer-reviewed study of 4,000-word academic papers found Copyleaks missed every fully AI-generated one. No per-model detection rate is published, and the model running today, V11.0, shipped on 22 September 2026 with no published evaluation.

Is Copyleaks free?

Partly. You can scan up to 25,000 characters without creating an account, which returns a percentage and highlighted AI phrases, and a free account adds monthly credits and sentence-by-sentence insights. Full AI Logic explanations and image detection require a paid plan. As of 28 September 2026 the Personal plan is $16.99 per month billed monthly, or $13.99 per month billed annually at $167.88 for 1,200 credits; Pro is $99.99 monthly or $74.99 per month billed annually. Education and Enterprise plans are quoted on request and the prices are not published.

Why does Copyleaks say my writing is AI when I wrote it myself?

Less often than most detectors, but it does happen. Copyleaks reports a measured false-positive rate of 0.026% on its own test corpus, and the 2026 peer-reviewed study found no false positives across 40 human papers, so a Copyleaks flag on genuinely human writing is comparatively rare. When it happens, the cause is the same as elsewhere: the detector responds to how predictable and regular your prose is, not to who typed it, which puts formulaic, technical and carefully edited writing at higher risk. Copyleaks’ own FAQ acknowledges that non-native English speakers face a higher risk of misclassification. Note also that the sensitivity level your institution chose changes the result.

Is Copyleaks better than Turnitin?

On the most recent peer-reviewed comparison, the two failed in the same direction and to a similar degree. The June 2026 study found both Copyleaks and Turnitin recorded 0% strict accuracy on fully AI-generated master’s-level papers, and both produced no false positives on human papers. The practical differences are about access and process rather than accuracy: Turnitin is institution-only and its AI indicator is visible to staff, while Copyleaks offers a free public scan anyone can run, publishes a dated testing methodology and a version number, and exposes three selectable sensitivity levels. Neither should be treated as proof.

Does Copyleaks detect paraphrased or humanised text?

Inconsistently, and the paraphrasing tool used matters more than anything else. In the 2024 study in the Indian Journal of Psychological Medicine, Copyleaks scored an AI-written article at 100% AI, a Grammarly-paraphrased version at 100% and a ChatGPT-paraphrased version at 100% — but a QuillBot-paraphrased version of the same text at 0%. That was a single sample of about 500 words, so treat it as a signal rather than a rate. The 2026 study of full-length papers found Copyleaks reached 22.5% strict accuracy on humanised AI papers and 30% on hybrid human-plus-AI papers, so the general pattern holds: modification degrades this detector substantially.

What does the Copyleaks AI percentage actually mean?

It is an aggregate of section-level classifications expressing how confident the model is that the text was machine-generated, not the proportion of your document that came from a chatbot. It is also a separate, unrelated number from the Copyleaks similarity score, which measures matching against published sources — a document can score 0% similarity and 80% AI, or the reverse. The percentage also depends on which of the three detection sensitivity levels the scan ran at, a setting chosen by the institution or instructor before submission.

What are AI Phrases and AI Source Match?

They are the two components of Copyleaks’ AI Logic layer, and they are different kinds of evidence. AI Phrases highlights the specific sentences whose linguistic patterns drove the score, so you can see what the classifier reacted to. AI Source Match checks flagged text against content already published on AI-generated websites and can return a named external source. A Source Match hit is checkable in a way a classifier percentage is not; a Source Match miss tells you nothing either way. AI Logic ships across Canvas, D2L Brightspace, Moodle, Blackboard, Schoology, Edsby and Sakai, and phrase-level explanations are available in six languages against detection in 30 or more.

How many words does Copyleaks need?

Copyleaks sets a minimum of 255 characters on the web platform and 350 characters in the browser extension — roughly 40 to 55 words. That floor is far below the length of any text in any published accuracy test: the vendor’s own V10 evaluation used 350 characters as its minimum and the 2026 peer-reviewed study used papers of at least 4,000 words. A Copyleaks score on a short passage is outside the range anyone has measured, so short-text results should be discarded rather than discounted. The free scan without an account accepts up to 25,000 characters.

Can Copyleaks be wrong?

Yes, and its own terms of use say so explicitly. The terms, last updated 24 August 2026, state that Copyleaks “does not provide, and expressly disclaims, any warranty regarding the accuracy of the result provided by the AI Features,” and that “the AI Features are an assessment and diagnostic tools which does not guarantee whether any given content was AI Generated.” Total liability is capped at the fees paid in the preceding twelve months, which for a free user is nothing. The documented case of Marley Stevens at the University of North Georgia in April 2024 is instructive: Copyleaks initially flagged her Grammarly-edited paper as bot-written and later returned a human verdict on the same text.

Do universities use Copyleaks?

Yes. Copyleaks integrates with Canvas, D2L Brightspace, Moodle, Blackboard, Schoology, Edsby and Sakai, and where an institution enables it a submission can be scanned automatically with the result shown to the instructor. Copyleaks says it serves millions of users across enterprise and education; a precise count of institutional customers is not published. Its own guidance is that results “should be used to inform the situation and offer the option for a learning opportunity” and work best “paired with human judgment,” which is narrower than how some institutions use the score.

Who owns Copyleaks?

Copyleaks, Inc. is an independent company founded in 2015 by Alon Yamin, who is chief executive, and Yehonathan Bitton, who is chief technology officer. It is headquartered at 115 E 23rd St, 7th Floor, New York, NY 10010. Disclosed funding totals $7.8 million, comprising a $1.8 million seed round and a $6 million Series A announced in April 2022 led by JAL Ventures with Connecticut Innovations participating. There is no disclosed corporate parent.

Is Copyleaks the same as its plagiarism checker?

No. Copyleaks sells two distinct products that produce two distinct numbers. The plagiarism checker matches submitted text against published sources across more than 100 languages and produces a similarity score; the AI detector estimates generative-AI involvement and produces an AI percentage. The plagiarism engine is the older product and nothing on this page speaks to its performance. Copyleaks also sells Codeleaks for source code, an AI image detector, and an AI video detector announced in June 2026, none of which share the text detector’s published accuracy figures.

← All guides