How the Copyleaks AI Detector works, what its results mean, and what to do if it flags you

What Copyleaks is, and where you meet it

Copyleaks was founded in 2015 by Alon Yamin and Yehonatan Bitton, two former programmers from Israel, and is now based in New York. It began as a plagiarism checker; AI detection came later. Today it sells an AI detector for text, images and video, a plagiarism checker, a grammar checker, and tools for companies that want to govern how staff use AI.

Most students meet it without choosing it. Copyleaks plugs into Canvas, Brightspace (D2L), Google Classroom, Moodle, Blackboard, Schoology, Sakai and other learning platforms, and in Canvas its AI alert appears under the similarity score, with suspected AI text underlined in red. Writers meet it through clients and publishers, and through its Chrome and Edge extension, Google Docs add-on and API.

How Copyleaks says it works

Copyleaks describes a model trained "to recognize human writing patterns, flagging text as potential generative AI when it detects deviations from known human patterns and specific AI signals." Among the signals it names are frequency ratios, parts of speech, syllable dispersion and hyphen usage. It says it trained on human writing from universities and companies, and that it does not train its detector on specific language models. The current model is version 11.

On top of the score sits what Copyleaks calls AI Logic, in two parts. AI Phrases "detects phrases that appear with higher frequency in AI-written text than in human writing", and AI Source Match checks whether any part of the text matches AI-generated content already published online. Phrase highlighting is a useful hint, but phrases that are common in AI text are often common in careful human writing too, because AI models learned them from people.

Whoever runs the scan can also choose a sensitivity: 1 "Extra Safe", 2 "Balanced" (the default) or 3 "Extra Sensitive". Copyleaks says the most sensitive level is "designed to flag AI text that was put through a 'humanizer' or text spinner", and it flags more human writing too. How AI detectors work covers the general ideas behind all of this.

What the result means

Copyleaks shows a percentage, highlighted passages and, with a free account, sentence-by-sentence detail. What the percentage means is less settled than it looks. Copyleaks' FAQ says it "estimates the total content in a document that generative AI may have created", while its developer documentation describes the overall figure as "the overall probability that the text is AI-generated". Sheridan College, which uses Copyleaks, tells its staff the percentage "is not a confidence score" and that Copyleaks judges larger chunks of text, so human sentences inside a flagged chunk can be marked as AI.

So if you are shown a Copyleaks figure, ask three things: which sensitivity level was used, which passages were highlighted, and whether those passages are quotations, references or standard phrasing that you would expect in any essay on the subject. What a detector score means explains why a single number hides so much.

• Length: at least 255 characters on the website and 350 in the extension; Copyleaks recommends at least 350 words for a reliable result.
• Free check: up to 25,000 characters without logging in.
• Languages: 30, from English, Spanish and French to Arabic, Hindi and Vietnamese. Copyleaks says English gives the highest accuracy.

Grammar tools and humanizers

Copyleaks tested grammar tools itself in May 2024. When it applied ordinary grammar fixes to 1,000 human essays, all of them still scored as human; when it rewrote 500 essays with Grammarly's generative "Improve" feature, 31.6% scored as AI. The line is between correcting your writing and having AI rewrite it, and Copyleaks warns that tools like Grammarly "can be flagged as potential AI content". If you used one, say so.

AntiGPT is a humanizer, so we will be plain about the rest. Copyleaks says its sensitive setting is built to catch humanized text. We do not run Copyleaks, we publish no figures for it, and we cannot tell you how it will score any text, including text rewritten with AntiGPT. If your course does not allow AI, rewriting AI text with any tool is still using AI.

False positives: what Copyleaks says, and what others found

Copyleaks gives several false-positive figures: ".03%" on its product page, "0.2%" on its extension page, and 0.30% at the default setting in the test report for its current model, published in September 2026. That report is the one with stated conditions: an English-only test by Copyleaks' own teams, in which 923 of 326,120 human texts were flagged, including 6 of 74,212 student essays. At the most sensitive setting the rate rose to 0.66%. Copyleaks also acknowledges that the styles of "non-native English speakers, neurodiverse writers" can "occasionally mimic patterns misconstrued as AI-generated".

Independent tests vary with the sample. A preregistered study published in June 2026 in the International Journal for Educational Integrity found no false positives on 40 long academic papers by non-native English-speaking graduate students. In a 2024 study of 30 detectors, Copyleaks was one of two that classified every human essay in its English first- and second-language sets correctly. On the other side, Bloomberg Businessweek ran 500 essays written in 2022, before ChatGPT, through Copyleaks and GPTZero in 2024 and found 1 to 2% wrongly flagged, some with near-total certainty; a 2023 University of Toronto study found Copyleaks misjudging all ten human-written Spanish texts it tested; and a small 2024 study saw 5 of its 10 human control samples labelled as AI.

Low rates still add up. At 0.3%, a university scanning 100,000 submissions a year would wrongly flag around 300. Our own benchmark uses a different detector, and its human passages show the kind of text any detector can get wrong: famous, polished writing that a model finds very familiar.

If Copyleaks flags your work

Copyleaks' own position is the place to start: "Copyleaks does not decide whether misconduct occurred. It provides evidence educators can evaluate alongside the assignment, the student's prior work, and the institution's policy." Its chief operating officer has written that "detectors should be used as a prompt for further review, not as a final verdict", and its chief executive called a flag "kind of like a yellow flag". Then:

• Ask for the report, the sensitivity level and the highlighted passages. If your institution has enabled it, you may be able to see how your work will be reviewed before or after you submit.
• Check what was highlighted. Quotations, references, set phrases from your subject and text inside a mostly AI-flagged chunk are all worth pointing out.
• Show your process: version history in Google Docs or Word, outlines, notes and drafts. Copyleaks names "the student's prior work" as something to weigh, so offer earlier assignments in your own style.
• Mention the tools you used, including grammar checkers and translation. If English is not your first language, say so; our guide for non-native English writers explains why that matters.
• Copyleaks has no appeal process for students; your institution decides. If you are accused of using AI sets out the steps.

Where AntiGPT fits, honestly

AntiGPT's detector is Hello-SimpleAI's open RoBERTa classifier, not Copyleaks, so its score will not match Copyleaks' and does not predict it. We show it before and after a rewrite as an estimate of how predictable your text reads, which is useful feedback while you edit and nothing more.

Where your course or client allows AI help, AntiGPT rewrites a draft so it reads in your own voice and keeps your terminology; our guide to essays and coursework covers citations and disclosure. Check every fact and quotation after any rewrite. The free readability checker is a quick way to see whether a passage reads like you or like a template.