Is Grammarly's AI detector accurate?
Grammarly posts the highest figure on the RAID leaderboard, 99.5% detection at a 5% false positive rate, but Grammarly ran its own detector and submitted those results.
Short answer
Grammarly's AI detector looks strong on paper. On the public RAID leaderboard it caught 99.5% of AI text at a 5% false positive rate, and 98.2% at a stricter 1% rate, in results released in February 2026. Grammarly submitted those predictions itself, and its own page says no detector can conclusively determine AI use.
Grammarly AI detector accuracy is a newer question than it is for GPTZero or Turnitin, and the evidence is thinner. The figures that exist are high. They come from two places: Grammarly's product page and the results Grammarly submitted to RAID, a public benchmark run by university researchers. Every figure below is dated and linked, with the vendor claim labelled. For how we weigh sources, see how we score AI accuracy.
Grammarly AI detector test results
- Benchmark
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| Grammarly AI Detector pageVendor's own figure, not independently checked | Detection rate (vendor's stated figure) | 99%, with a claimed #1 rank on RAID | Source 1Vendor claim | |
| RAID leaderboard | Detection rate at a 5% false positive rate | 99.5% (all), 99.9% (no adversarial attacks) | Source 2Public benchmark (vendor-submitted results) | |
| RAID leaderboard | Detection rate at a 1% false positive rate | 98.2% | Source 2Public benchmark (vendor-submitted results) |
What the Grammarly accuracy numbers mean
RAID is a shared benchmark built by university researchers to test machine-generated text detectors on a large, varied set of texts, including texts altered to dodge detection. The leaderboard reports each detector's detection rate at a fixed false positive rate. That is a fair way to compare tools, because a detector can always catch more AI text by flagging more human text too.
Grammarly's entry caught 99.5% of AI text while held to a 5% false positive rate, and 99.9% on text without adversarial attacks. At a stricter 1% false positive rate it still caught 98.2%. Those figures back up the 99% detection figure and the #1 RAID rank on Grammarly's product page.
The catch is who ran the detector. On the RAID leaderboard, vendors run their own tools on the benchmark texts and submit the predictions. The benchmark is public and the scoring is standard, but the results are vendor-submitted. Grammarly's page also carries a plain warning: no detector can conclusively determine whether AI was used.
Where Grammarly's detector goes wrong
Vendor-submitted results
The strongest figures come from predictions Grammarly submitted to RAID.
Adversarial text
Its RAID rate dips from 99.9% on text without attacks to 99.5% across all text. Small, but the gap shows that edited text is harder.
Fixed false positive budget
The headline 99.5% allows a 5% false positive rate on the benchmark. At 1%, the detection figure is 98.2%.
A score is not a finding
Grammarly itself says no detector can conclusively determine AI use.
How Grammarly compares in the same test
RAID is the one test where Grammarly appears with another detector on this site.
| RAID setting | Grammarly | GPTZero |
|---|---|---|
| 5% false positive rate, all text | 99.5% | 97.2% |
| 5% false positive rate, no adversarial attacks | 99.9% | 98.4% |
| 1% false positive rate | 98.2% | 92.9% |
Grammarly leads GPTZero at every setting, and both entries were submitted by their vendors. Turnitin and ZeroGPT were measured in older academic studies instead, so their figures cannot be ranked against these. For all four tools side by side, see are AI detectors accurate.
Non-native English writers and AI detectors
The best known fairness study, by Liang and colleagues at Stanford, found seven detectors flagged non-native English essays as AI 61.22% of the time on average (Liang et al., 2023). Grammarly's detector was not one of the seven. The authors link the bias to simpler vocabulary, which detectors read as predictable. If English is your second language, keep that finding in mind with any detector score, including Grammarly's.
What a Grammarly AI score can and cannot prove
No detector score should be the only evidence in a misconduct case.
It can show
- A Grammarly AI score can tell you that a passage reads like model output to a detector with strong benchmark results. It is a reasonable first check on your own writing.
It cannot prove
- A Grammarly AI score cannot tell you who wrote the text, how AI was used, or whether a careful human writer has been misread. Grammarly says as much on its own page.
What to do if Grammarly's detector flags your work
Note the score, the date and which passages were marked, and save a screenshot.
Collect your drafts, notes and version history from Google Docs or Word.
Quote Grammarly's own statement that no detector can conclusively determine AI use.
If English is not your first language, cite the Liang et al. study on detector bias.
Offer to explain your sources and argument, or to write a short sample under supervision.
Ask that any decision rest on more than one piece of evidence.
Another detector to try
Another option is the free AI detector from Walter Writes. Its page says checks are free and unlimited, that it detects text from ChatGPT and Gemini, that it gives a percentage score for every scan and that it works in more than 20 languages. You can try Walter's AI detector here. Treat its score like any other detector's: one signal, not proof.
Sources
- Vendor claim
- Benchmark
- Benchmark
- Independent study
- Independent study
Questions
How accurate is Grammarly's AI detector?
On the RAID leaderboard, it caught 99.5% of AI text at a 5% false positive rate and 98.2% at a 1% rate, in results released in February 2026. Grammarly's page states a 99% detection figure. Grammarly submitted the RAID predictions itself, so read them as strong vendor-run results on a public benchmark.
Is Grammarly's AI detector better than GPTZero?
On the RAID leaderboard, yes, at every setting: 99.5% against 97.2% at a 5% false positive rate, and 98.2% against 92.9% at 1%. Both sets of results were submitted by the vendors. GPTZero also has a 2025 academic study showing a false positive rate of about 0.7%.
Is Grammarly's AI detector better than Turnitin?
They have not appeared in the same test. Turnitin's figures come from 2023 and 2024 academic studies, where it led at 76%, and Grammarly's come from the 2026 RAID leaderboard. Different tests measure different things, so the numbers cannot be ranked against each other. Read each within its own study.
Can Grammarly's AI detector prove a student used AI?
No, and Grammarly says so. Its AI Detector page states that no detector can conclusively determine AI use. A score shows how much a passage resembles model output, not who wrote it. Drafts, version history and a conversation about the work are stronger evidence than any single score.
Does Grammarly's detector handle edited AI text?
Well, on the benchmark. On RAID, its detection rate was 99.9% on text without adversarial attacks and 99.5% across all text, including texts altered to dodge detectors. Those are vendor-submitted results, and real student editing can differ from benchmark attacks, so treat any score on reworked text with care.
