also searched as: is grammarly ai detector accurate · is grammarly ai checker accurate · how accurate is grammarly ai detector · how accurate is grammarly ai detection · is the grammarly ai detector accurate · grammarly ai detector accuracy · is grammarly ai detector reliable
Kinda: in small independent tests, Grammarly's AI detector cleared every human sample but missed most text from current chatbots. Its 99% claim comes from a benchmark built on older models, and Grammarly says the score is not a definitive assessment.
Trust it for a quick look at whether a draft you wrote yourself reads as machine-made.
Double-check it when the text came from a current chatbot or was rewritten afterwards, because that is where tests found it scoring low.
Don't rely on it if you are deciding whether someone cheated, or guessing what Turnitin will tell your professor.
sources checked 2026-10-13 · by dan scutaru
| who tested it | what they found |
|---|---|
| Grammarly's own claim | "99%" and "#1 on RAID" On its detector page; its help guide says the model is tuned to avoid false positives but gives no false-positive rate. source |
| RAID benchmark (UPenn), Grammarly's February 2026 entry | 99.9% of AI text caught Scored by RAID's scripts at a cutoff flagging 5% of human texts, on text from 11 models no newer than GPT-4; 98.2% with attacks at a 1% cutoff. source |
| RAID results, checked 13 October 2026 | No longer 1st On our count of the benchmark's published results files, four detectors now score higher on unedited text and more on attacked text. source |
| Pangram Labs, January 2026 (a rival detector maker) | 0 of 9 AI essays caught Nine 1,000-word essays from GPT-4o, Gemini 2.0 and Claude 3.7 Sonnet, counted as caught at 75% AI or more; all three human texts passed. source |
| Canadian Writing Centres Association blog, November 2024 | ChatGPT rewrite scored 0% The editor's own paragraph scored 0%, a ChatGPT-4o rewrite of it 0% on five runs, and a second rewrite 50%. source |
Less accurate than the 99% suggests. That figure comes from RAID, a University of Pennsylvania benchmark whose newest model is GPT-4, scored at a cutoff the benchmark sets, not the one on Grammarly's free page. The two outside tests on newer or rewritten text found the opposite failure: Grammarly let AI text through rather than flagging people.
That is the trade-off Grammarly says it chose. Why every style-based detector struggles is on our page about whether AI detectors are accurate.
To Grammarly's own detector, mostly no. Its help guide says ordinary red and blue underline corrections should typically not change the score, while rewrites from its Paraphraser and Humanizer agents will likely be flagged.
Grammarly also says its scores may differ from Turnitin, GPTZero and Copyleaks, so a clean result is not a preview of your professor's report. Whether Grammarly counts as AI under your course rules is for your syllabus to say.
Authorship logs whether each part of a document was typed, pasted, written with AI or edited by Grammarly, and builds a report with a replay of your writing that you choose whether to share. It comes with Grammarly Free and Pro and works in Google Docs with the browser extension, in Word with the desktop app, and in the Grammarly editor.
It records how text arrived, not where the ideas came from: text you retype by hand is logged as typed.
Its grammar corrections should not be, by Grammarly's account, but text rewritten by its generative features likely will be, at least by its own detector. Other detectors use other models.
As a rough first check. It caught 99.9% of older-model text on RAID but none of nine current-chatbot essays in Pangram's test, so a low score is weak evidence that a text is human.
We found no independent test of the two on the same essays, and one score does not predict the other. Turnitin's record is on our Turnitin page.