Does GPTZero Detect Grammarly? When Editing Gets Flagged as AI
By Ashikur Rahman11 min readQuick answer: GPTZero can detect Grammarly's AI rewrites, but it publishes no test of Grammarly's ordinary grammar fixes. Its "AI Polished" tag, filed under "Mixed", marks human text later improved by AI, and GPTZero names "the use of Grammarly" as an example of that enhancement. On its own benchmark of AI-edited documents, 65% came back Mixed, 24% AI and 11% human. Spelling corrections are a different case: GPTZero's method set aside texts changed too little to count as edited. Since June 2026, Grammarly's parent has also agreed to buy GPTZero.
What does GPTZero say about Grammarly?
GPTZero mentions Grammarly once, in the post that introduced its "AI Polished" label, and the sentence is short. In that June 2025 post it says: "We consider this "AI Polished" classification to indicate more advanced form of writing enhancement (like the use of Grammarly). Still, it indicates the work is original, unlike texts generated entirely with AI."
Two things follow. GPTZero separates polishing from generation, and a polished result is not meant to read as an accusation. But "like the use of Grammarly" is loose wording. It does not say which Grammarly feature, and Grammarly's features differ a great deal, from a red underline to a full paragraph rewrite.
The word "detect" also hides a technical point. GPTZero does not look for Grammarly as a brand. It has no way to tell from text alone which product made an edit. It estimates whether sentences read as if a language model wrote or changed them. So the real question is what your edits did to your sentences.
Which Grammarly edits are likely to trigger GPTZero?
The more a feature rewrites, the more likely a language-model signal appears. No published test maps this directly for GPTZero, so the table combines what GPTZero, Grammarly and Turnitin each state.
| Grammarly action | What changes | Likely reading in GPTZero | Basis |
|---|---|---|---|
| Spelling, punctuation, basic grammar | Small corrections | Probably human, not tested | GPTZero rejected low-edit samples as too close to human |
| Clarity and word-choice suggestions | Some words and phrases | Uncertain, more edits raise the risk | No published test |
| Rewrite full sentences | Sentence structure and wording | Could tag as AI Polished | Close to GPTZero's polish examples |
| Paraphraser or Humanizer agent | Whole paragraphs | Likely Mixed or AI | Grammarly: these rewrites "come from our LLM" |
| Prompt-generated draft | All of the text | Likely AI | This is AI generation |
The first row rests on GPTZero's method. To build its polished training set, it measured edit distance from the original and "defined thresholds to either accept or reject AI-edited documents." Texts were rejected when "they had not been changed enough to be considered edited." Its published rejected example swaps "led" for "contributed" and "said" for "noted", with few other changes. If that counted as too close to human during training, it suggests grammar-level correction is not what the tag targets. That is an inference from how the data was built, not a promise about any one essay.
Grammarly's AI Detector guide draws the same line on its own product: "traditional, non-generative corrections in the form of Grammarly red and blue underlines should typically not impact the percentage score." Meaningful rewrites "will raise the percentage score." That statement is about Grammarly's detector, not GPTZero, but the reason carries over. The full version, with Turnitin's answer, is in whether Grammarly shows up as AI.
What does "AI Polished" mean on a GPTZero result?
It means GPTZero thinks a human wrote the text first and AI then changed much of it. The result sits in the Mixed class, with edited sentences shown in light green. Ordinary AI passages in a mixed document are highlighted in yellow. GPTZero says "In most cases, editing a document with AI changes the majority of sentences."
To build the category, GPTZero prompted language models with instructions such as "Please polish this text to improve its overall quality". It then benchmarked 10,000 samples and published this table:
| Detector, on AI-edited documents | Mixed result | AI result | Human result |
|---|---|---|---|
| GPTZero, AI Polished | 65% | 24% | 11% |
| Copyleaks | 7% | 79% | 14% |
| Originality | N/A | 93% | 7% |
Read that as GPTZero's own benchmark, published by a vendor comparing itself with rivals. The useful figure is the 24%. Even on GPTZero's preferred reading, about one in four documents edited by AI still came back fully AI, a harder result for an honest writer to explain. GPTZero expected the tag on 4.4% of user submissions, and said the update "should reduce the number of false positive scans".
Why does it matter that Grammarly's parent is buying GPTZero?
Because the company selling the rewrite tools will also own a detector that flags rewrites. SiliconANGLE reported in June 2026 that Superhuman, the company formerly known as Grammarly, agreed to acquire GPTZero, with the price undisclosed. GPTZero is to be built into Superhuman Go. GPTZero's own site footer now lists Superhuman and Grammarly under Company.
| Product in the group after the deal | Role |
|---|---|
| Grammarly suggestions and agents | Corrects and rewrites text |
| Grammarly Humanizer | Tunes AI writing to sound "more natural, sincere, and distinctly human" |
| Grammarly AI Detector | Scores text for AI, using its own model |
| Grammarly Authorship | Records how a document was written |
| GPTZero | A second AI detector, plus Writing Replay |
Nothing read says the two detectors treat Grammarly's edits specially, or that they will. CriticNest has no evidence of that, and readers should not assume it either way. What changes is the need for caution about independence. When a seller of writing tools also owns a detection brand, its accuracy claims need outside checks. This article could not confirm whether the deal has closed. It also found no statement of how the products will share data, so check the current privacy terms if that matters to you.
Transparency note: CriticNest has no affiliate or commercial relationship with GPTZero, Grammarly, Superhuman, Turnitin, Pangram or QuillBot. Nothing here is a paid placement. No link in this article earns a commission. Every quotation was read from the linked pages on 6 November 2026, and no detector was run for this article. If that changes, the disclosure will move up.
Is GPTZero accurate on edited text?
GPTZero's own numbers are excellent, outside tests are mixed, and edited or mixed text is where the gap shows. Here is what each source reports, with the vendor claims kept apart.
| Source | Test | Result for GPTZero |
|---|---|---|
| GPTZero homepage | Vendor headline | "99.9% Accuracy" |
| GPTZero technology page | Vendor claim | False positives "no more than 1%"; 96.5% on mixed documents |
| GPTZero 4o post, 18 Oct 2026 | Vendor run on PERSUADE student essays | 1 false positive in 10,402 essays |
| Epoch AI, July 2026 | Independent; 495 human, 297 AI passages per condition | 0 of 495 false positives; missed 2 of 297 basic AI and 32 of 297 style imitations |
| IJEI study, June 2026 | Peer reviewed; 160 academic papers | All human papers read as human; 0.0% on hybrid and 2.5% on humanised papers |
| Karr et al., 2026 | 642 published abstracts refined by Gemini 3 Flash | 37.6% and 48.5% of refined abstracts flagged; 8.9% of recent originals flagged |
These do not contradict each other as much as it looks. Each measures a different thing, on a different model date. The Epoch test used version 2026-05-11-base, the IJEI paper's AI text dates from May 2025, and neither covers 4o. A pattern still shows: plainly human text is read well across these tests, plain AI text usually is too, with the Deep Research papers as a large exception, and sorting partly edited text is much weaker.
There is a small contradiction on GPTZero's own pages. The technology page says "Even if AI content has been altered to look more human-like, GPTZero can detect it." The FAQ says "Our classifier is not trained to identify AI-generated text after it has been heavily modified after generation". The FAQ may simply be older text, but on 6 November 2026 both were live.
Does GPTZero give false positives on non-native English writing?
The concern is real, and the company says it has worked on it. The widely cited 2023 study by Liang and colleagues ran seven detectors over 91 TOEFL essays and found an average false-positive rate of 61.3%. It used models from early 2023, and the figure is an average across detectors. GPTZero now says its work has "reduced AI detection's false positive rate on TOEFL texts to 1.1%", a vendor figure.
Later outside work is more reassuring. In the IJEI paper, the 40 human papers were by non-native English-speaking graduate students, and all four tools placed them in the 0% to 20% band. That sample is small, and long academic papers are not the same as short essays. GPTZero's FAQ also warns that accuracy rises with length and that the classifier "can sometimes flag other machine-generated or highly procedural text". Lab reports and templates deserve extra care.
Does a GPTZero score predict your Turnitin score?
No. They are separate models, and published tests show them behaving differently. In the IJEI study, Turnitin reached 60.0% on hybrid papers where GPTZero scored 0.0%, while on human text the paper describes GPTZero as the one with a small false-positive rate. Grammarly's guide warns its own scores "may differ from those of other solutions like Turnitin, GPTZero, Copyleaks". For Turnitin's mechanics, see what Turnitin's AI score means, and for its record, how accurate its detector is.
QuillBot has a parallel tag for AI-refined text, explained in whether QuillBot shows up as AI. Every vendor's label is its own scale, so a "Mixed" from one tool is not a finding in another.
What should you do if GPTZero marks your edited work as Mixed?
Treat it as a question to answer, and answer it with your writing process. GPTZero itself says "these results should not be used to punish students" and describes the tool as a way to start a conversation.
- Look at the classification and highlights. Light green suggests AI polishing; yellow suggests AI-generated chunks. They raise different questions.
- Be precise about what you used. "Grammar suggestions only" and "I accepted three sentence rewrites" are both clear, checkable answers. Vague answers sound worse than either.
- Bring your record. GPTZero tells educators to ask for "drafts, revision histories, or brainstorming notes", and says a document "typed out with several edits over a reasonable period of time" is likely authentic.
- Use process tools before you write next time. Google Docs keeps version history. Grammarly Authorship records typed, pasted and AI text. GPTZero's Writing Replay lets a reader "Watch documents being written, view copy and pastes".
- Check the policy. AI polishing may be allowed in one course and banned in the next. If rewrites are banned, a Mixed tag may be correct, and the honest move is to say so.
Turning off Grammarly's generative features is simple if you want correction only. Grammarly says any user can do it in "the Feature customization tab in their account settings", and says they are "not turned on by default".
How this was researched
GPTZero's homepage, technology page, FAQ and two posts were read on 6 November 2026, alongside Grammarly's AI Detector guide through its support API, SiliconANGLE's report on the deal, the Epoch AI data insight, the Karr preprint, the 2026 IJEI paper and the Liang paper.
Not verified: how GPTZero scores Grammarly grammar-only edits, since no test appears on the pages read; whether the Superhuman deal has closed; GPTZero's current prices, because its pricing page did not render figures. No detector was run. The rows in the first table marked as likely are this article's reading of those sources, not measurements. The author's role is careful reader of the fine print, not tester.
The verdict: GPTZero reads the edit, not the brand
GPTZero does not detect Grammarly as such. It looks for language-model patterns in sentences. Grammarly's rewrites and agents are language-model output, so GPTZero can reasonably tag them as AI Polished or worse. Small corrections look very different from that, though neither company has published a test that settles it.
If you only accept spelling and grammar fixes, your risk looks low. If you accept rewrites, assume they can be seen. Either way, keep a record of how you wrote, because a detector being bought by the group behind your editor is one more reason to rely on evidence you control.
Frequently asked questions
Does GPTZero flag Grammarly as AI?
It can flag Grammarly's AI rewrites. GPTZero's AI Polished tag marks human text improved by AI and names Grammarly as an example. Basic spelling and grammar corrections have no published GPTZero test, though its method set aside lightly changed texts as too similar to human writing.
Is GPTZero owned by Grammarly?
SiliconANGLE reported in June 2026 that Superhuman, the company formerly known as Grammarly, agreed to acquire GPTZero, with the price undisclosed. GPTZero's site footer now links to Superhuman and Grammarly. This article could not confirm the closing date, or how the two products will work together.
Does GPTZero work?
For plainly human and plainly AI text, independent results are strong. Epoch AI recorded no false positives on 495 human passages and missed 2 of 297 basic AI passages. For hybrid academic papers, a 2026 peer-reviewed study scored it 0.0%. Results depend on the text and model version.
Does GPTZero give false positives?
Yes, sometimes. GPTZero claims "no more than 1%" and cites 1 in 10,402 on one essay set. Its FAQ admits "human is classified as AI" in edge cases and warns about highly procedural text. In a 2026 study, 8.9% of recently published abstracts were flagged, though those authors may have used AI.
Is GPTZero as accurate as Turnitin?
Neither wins every test. In a 2026 peer-reviewed study of 160 papers, Turnitin did better on hybrid text, and both missed fully AI papers made with ChatGPT Deep Research in May 2025. Turnitin was not in Epoch AI's July 2026 test. Their models also keep changing.
Does GPTZero save your work?
GPTZero says it does not "store or collect the documents passed into any calls to our API", but it does "store inputs from calls made from our dashboard", used in aggregate to improve the service. Check its privacy policy before pasting confidential or unpublished writing.
Does GPTZero use AI?
Yes. GPTZero says it uses "an end-to-end deep learning approach", with a sentence-by-sentence classification model trained on human text and AI text from a range of language models. It estimates probability and does not know who actually wrote something.
Does GPTZero work for Turnitin?
Not as a preview. GPTZero and Turnitin are separate models, and a result from one does not predict the other. A 2026 study saw them score the same hybrid papers very differently. Only your institution's Turnitin report shows what your instructor sees.
Why did Grammarly's detector and GPTZero give different scores?
They were trained on different data and report different things. Grammarly gives a percentage, while GPTZero classifies text as human, AI or mixed with highlights. Grammarly warns that its scores "may differ" from GPTZero's. Disagreement is normal and shows why one score should not decide anything.
How do I prove I only used Grammarly for grammar?
Show your process. Grammarly Authorship separates typed text from pasted and AI text if you switch it on before writing, and Google Docs keeps version history. Keep drafts and notes as well. GPTZero's FAQ tells educators to look for exactly these records when a result is in doubt.
