Short answer: Grammarly's AI checker scores very well on the RAID benchmark it submitted to, but hands-on tests have been mixed, and Grammarly itself says no AI detector is 100% accurate and that a score should never be used alone (Grammarly). If a result matters, read the highlighted text, compare with a second tool such as our free AI detector, and, if you are the writer, consider Grammarly Authorship, which records how a document was written instead of guessing from the final text.
This guide covers what Grammarly's AI detector checks, how it differs from Grammarly Authorship, what the published tests say, and how to use any score fairly.
What Grammarly's AI detector does
Grammarly is best known as a grammar and writing assistant. Its free AI detector sits on its website next to the grammar checker, plagiarism checker and paraphraser. You paste text or upload a document, click "Scan for AI", and see what percentage of the text appears to be AI-generated (Grammarly AI Detector).
Grammarly says the detector splits a document into segments and checks each one for language patterns often linked to AI writing, then estimates how much of the content may be AI-generated. Its page says the model was trained on tens of thousands of human-written and AI-generated texts (Grammarly AI Detector). With Grammarly Pro you also get an "AI Detector agent" that shows which parts may sound AI-generated and suggests rewrites and citations.
Grammarly is open about limits. Its own FAQ says AI detectors "do not track authorship, which means even human-written text can be flagged as AI-generated", that lightly edited AI text may not be detected, and that detection "should never be used as a standalone verification method" (Grammarly AI Detector).
One more piece of context: Superhuman, the company formerly known as Grammarly, announced in June 2026 that it was acquiring the AI detector GPTZero (BetaKit). We cover that tool separately in Is GPTZero accurate?.
Grammarly Authorship: a different kind of check
Grammarly Authorship is not a detector. It records the writing process while you work: which text was typed, pasted from a source, or written with AI help. It works in Google Docs (through the Grammarly browser extension), Microsoft Word (through the desktop app) and the Grammarly Editor, and it is available on both Grammarly Free and Pro (Grammarly Authorship).
The result is a shareable report that colour-codes each sentence by source, plus a replay of the writing from first word to last. Grammarly says Authorship is opt-in, and that your school cannot see the report unless you create it and send them the link (Grammarly Authorship).
Grammarly's own explanation of the difference is useful: an AI detector analyses text "in a moment in time" and keeps no document history, while Authorship tracks version history and where text came from (Grammarly Authorship). For a student who wants to show their work, a record of the process is stronger evidence than any detector score, from any company.
| Tool | Question it answers | What it needs | Main limit |
|---|---|---|---|
| Grammarly AI detector | Does this text read like AI writing? | Only the finished text | An estimate; human text can be flagged (Grammarly) |
| Grammarly Authorship | How was this document written? | Turned on while writing | Only covers writing done in supported apps (Grammarly) |
| Plagiarism checker | Was this copied from a source? | The finished text | Says nothing about AI; see our AI detector vs plagiarism checker guide |
What the tests say about accuracy
Grammarly's claim: the RAID benchmark
Grammarly's detector page says it "achieves 99% detection accuracy and ranks #1 on RAID's independent benchmark" (Grammarly AI Detector). RAID is an academic benchmark from the University of Pennsylvania with more than 6 million AI generations across 11 models, 8 domains and 11 adversarial attacks (Dugan et al., RAID, arXiv). Detector makers run the public test set themselves and RAID scores their predictions.
Grammarly's submission, dated February 2026, is public. Across the full benchmark, attacks included, it caught about 98.2% of AI texts when only 1% of human texts may be flagged, and about 99.5% when 5% may be flagged (RAID results file for Grammarly). That is a strong result on that benchmark.
What it does not mean: that 99 out of 100 verdicts on your essays will be right. RAID measures how much AI text is caught at a fixed false-alarm rate, on its own mix of English texts. Your writing, your AI model and your editing can be very different.
A journalist's hands-on test
ZDNet's David Gewirtz runs the same five texts (two human, three by ChatGPT) through a set of detectors. In his October 2025 round, Grammarly got two of five right, and he wrote that he saw no improvement since his previous test. In the same round, Grammarly's plagiarism checker did correctly spot that his test text had been published before (ZDNet).
Two caveats, to be fair to Grammarly. Five texts is far too few to measure accuracy, and that test came months before the February 2026 RAID submission, so the detector may have changed since.
Research on detectors in general
We found no peer-reviewed study that tests Grammarly's AI detector by itself, so we will not quote one. Studies of other detectors are still a useful warning:
- A Stanford study found seven detectors flagged 61.22% of essays by non-native English writers as AI on average (Liang et al., 2023). Grammarly was not in that study, but the risk applies to any detector used on second-language English.
- A review of 14 detection tools found accuracy dropped when AI text had been manually edited or machine-paraphrased (Weber-Wulff et al., 2023).
Why Grammarly may say "AI" about your own writing
A false positive is human writing labelled as AI. The common causes apply to Grammarly as much as to any detector:
- Polished text. Writing that has been smoothed by a grammar or rewriting tool, including Grammarly's own suggestions, can take on the even style detectors associate with AI.
- Short text. A few sentences give a detector little to work with. Check whole paragraphs or the whole document.
- Formulaic writing. Definitions, abstracts, standard cover-letter phrases and legal language are predictable by nature.
- Second-language English. Simpler vocabulary and sentence structure make text more predictable (Liang et al., 2023).
- Different tools, different answers. Detectors are trained on different data and use different thresholds, so Grammarly, QuillBot and ZeroGPT can disagree on the same essay. Our reviews of QuillBot's detector and ZeroGPT explain why.
How to use a Grammarly AI score fairly
If you are a teacher or reviewer:
- Scan the whole document, not one paragraph.
- Look at which passages drove the score. A reference list or a standard introduction means something different from a high score across every section.
- Check your own rules: if Grammarly's grammar suggestions are allowed, some "AI-like" polish may be within them.
- Ask for drafts, version history or an Authorship report, and talk to the writer before deciding anything.
If you are the writer:
- Turn on Authorship, or keep Google Docs or Word version history, while you write. A record made at the time is your best evidence.
- Keep notes, sources and outlines.
- If you are flagged, ask what the score was based on and offer your drafts. Grammarly's own FAQ says a detector result should not be used alone (Grammarly).
Where ProofAI fits
ProofAI is a stand-alone AI detector built in India, with prices in rupees. It checks pasted text in the AI text detector, PDF, DOCX, DOC and TXT files up to 25 MB in the document checker, and images for signs of AI generation; video detection is coming soon. It is free for 10 checks a day without logging in and 25 with a free account.
If you already use Grammarly for writing, its detector and Authorship may be all you need. ProofAI is useful as a second opinion on a disputed result, or when you need to check whole files or images. It can be wrong too, especially on short, formulaic or heavily edited text; our published test method is on the accuracy page.
FAQ
Is Grammarly's AI detector accurate?
On the RAID benchmark it catches most AI-written text at a fixed false-alarm rate, according to the results Grammarly submitted (RAID). A small hands-on test in 2025 was less positive (ZDNet). Treat its score as a signal, not proof.
Is the Grammarly AI checker free?
Yes, there is a free AI detector on Grammarly's website; Grammarly Pro adds the AI Detector agent with more detail and rewrite suggestions (Grammarly).
What is the difference between Grammarly's AI detector and Grammarly Authorship?
The detector estimates from the finished text whether it reads like AI. Authorship records how the document was written, typed, pasted or AI-assisted, and produces a report you can choose to share (Grammarly Authorship).
Can using Grammarly make my writing look AI-generated?
It can happen with any detector: heavily polished text is more predictable, and Grammarly itself says human-written text can be flagged (Grammarly). Keep your drafts or use Authorship so you can show how you wrote it.
Does Grammarly check plagiarism and AI at the same time?
They are separate checks. A plagiarism checker compares your text with existing sources; an AI detector looks at how the text is written. Our AI detector vs plagiarism checker guide explains when you need each.
Conclusion
Grammarly's AI detector is a convenient free check with a strong benchmark result, and Grammarly is honest that it is not proof. Its most useful tool for fairness may be Authorship, because a record of how a document was written beats any guess from the final text. Whichever tool you use, check enough text, compare when it matters, and let drafts and a conversation decide.
Originally published on ProofAI.



