Grammarly says its AI detector is 99% accurate and “#1” on an independent benchmark. We traced that claim to the source — and read the fine print Grammarly publishes right next to it.
Evidence review: built from Grammarly’s own pages, the public RAID benchmark data and published research, checked 14–15 Oct 2026. Hyptor Lab’s hands-on test is scheduled and will be added here. No affiliate links. How we test
The short answer
On a benchmark Grammarly submitted to itself, yes — around 99%. In the real world, nobody has independently tested it yet, and Grammarly’s own pages say never to rely on it alone.
Best for: a quick self-check of your own draft · Skip if: you need evidence about someone else’s writing · Better alternative: Grammarly’s own Authorship feature (free) · Confidence: Low–Medium — no independent study of this specific detector exists.
What Grammarly claims vs what the evidence shows
| Their claim | Verdict | What the evidence shows |
|---|---|---|
| “Achieves 99% detection accuracy” [1] | Holds, with caveats | Grammarly’s RAID results show 99.91% on standard text and 98.21% on attacked text at a 1% false-positive setting [5] — but Grammarly submitted those results itself, on a fixed benchmark dataset. |
| “ranks #1 on RAID’s independent benchmark” [1] | Holds, with caveats | RAID scores detectors on several metrics and settings [6]. Grammarly’s announcement gives no metric or setting [3], so “#1” depends on which view you choose and when. |
| “Designed to avoid wrongly flagging human-written text” [1] | Untested | Grammarly publishes no false-positive rate on its page, and we found no independent test of it on real human writing. |
Why this matters
Grammarly is on millions of student laptops, so its detector is often the first “AI score” a student or teacher sees. Independent research on other detectors has found high false-positive rates and bias against non-native English writers (see our AI detector evidence review). A score that sounds authoritative can do real harm if it’s treated as proof.
How we built this verdict
We quoted Grammarly’s product page, help centre and blog [1–4][7], read Grammarly’s raw submission files in the public RAID benchmark repository [5], and searched for peer-reviewed or institutional tests of the detector. We found none, so we say so rather than fill the gap with blog figures we couldn’t verify.
The 99% figure is real — but it’s self-reported
Finding: RAID is a respected academic benchmark (ACL 2024) [6]. Its public files show Grammarly’s entry was submitted from a grammarly.com contact, released 9 Feb 2026, with 99.91% accuracy on non-adversarial text and 98.21% with adversarial attacks at a 1% false-positive setting [5].
In practice: strong benchmark numbers, but the benchmark’s texts come from fixed generators — not the latest chatbots or your class’s essays. A benchmark result isn’t the same as an independent field test.
Grammarly’s own fine print tells you not to rely on it
Finding: the same page that advertises 99% says “No AI detector is 100% accurate”, that “even human-written text can be flagged as AI-generated”, that “lightly edited AI-generated text may evade detection”, and that detection “should never be used as a standalone verification method” [1]. Its help centre adds that results are “a single data point rather than the sole source of evidence” [2].
In practice: believe the fine print over the headline.
The surprise: Grammarly is quietly betting on something else
Grammarly’s blog calls AI detection “an emerging—and inexact—technology” that has been “biased against non-native English writers” [7]. Its answer is Authorship, which records how a document was written instead of guessing from the final draft — because “AI detectors only guess from the final draft” [4]. It’s available on Grammarly Free and Pro [4].
Where Grammarly’s detector genuinely helps
As a fast, private check on your own writing — for example, seeing whether a heavily AI-assisted paragraph reads as machine-written before you revise it. Its benchmark results suggest it’s competent on unedited AI text [5].
What it costs
Grammarly markets a “free AI detector” on its detector page [1]. Its plans page lists “Detect AI generated text” as included in Pro (shown at $12, billing period not stated on the page we fetched) and Enterprise, not Free [8] — likely the standalone tool versus the in-app agent. Grammarly publishes no word limits for the detector. Checked 14 Oct 2026.
Who should use it, and who shouldn’t
Use it if: you want a quick read on your own draft. Skip it if: you’re judging someone else’s work — no detector, including this one, can prove authorship. Better options: Grammarly Authorship, document version history, or simply talking through the work.
Questions people ask
Is Grammarly’s AI detector accurate?
On Grammarly’s self-submitted benchmark results, about 98–99.9% depending on the setting [5]. No independent real-world test exists yet, and Grammarly says not to rely on it alone [1].
Is Grammarly’s AI detector as accurate as Turnitin?
Unknown. We found no independent head-to-head test and no Grammarly statement comparing the two. Turnitin is sold only to institutions.
Is Grammarly’s AI detector free?
There is a free detector page [1]; the plans page lists AI detection under Pro and Enterprise [8]. Check what your account shows.
Can Grammarly’s AI detector flag human writing?
Yes — Grammarly says so itself: “even human-written text can be flagged as AI-generated” [1].
The verdict
Grammarly’s detector posts excellent numbers on a benchmark it submitted to — and Grammarly itself warns against treating them as proof, while promoting Authorship as the better answer. The decision: use it on your own drafts; for anything that matters, use Authorship or version history instead.
Change log
15 Oct 2026 — Published as an evidence review. Hyptor Lab hands-on test scheduled.
Sources
- Grammarly, Free AI Detector, accessed 14 Oct 2026 — grammarly.com
- Grammarly Support, “Does the use of Grammarly trigger AI detectors?”, accessed 14 Oct 2026 — support.grammarly.com
- Grammarly Blog, “Grammarly’s AI Detector Agent Ranks #1 in Quality”, updated 10 Feb 2026 — grammarly.com
- Grammarly Authorship, accessed 14 Oct 2026 — grammarly.com
- RAID leaderboard submission: Grammarly (metadata and results files), committed 9–10 Feb 2026 — github.com/liamdugan/raid
- Dugan L. et al., “RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors”, ACL 2024 — arxiv.org
- Grammarly Blog, “From AI Detection to Authorship”, updated 13 Aug 2024 — grammarly.com
- Grammarly Plans, accessed 14 Oct 2026 — grammarly.com
Right of reply
Grammarly can respond via our contact page; substantive responses will be published here.