Grammarly’s AI detector is a reasonable first screen, but its accuracy is not settled. Grammarly reports a top result on one public benchmark and also states, in its own help pages, that the score “should not be used as an objective source of truth.” Independent evidence on the current detector is thin, so treat the score as a starting point for review, not as proof of anything.
Key takeaways
- The detector reports a percentage: the share of scanned text that is likely AI-generated, according to Grammarly. It does not say why any text was flagged.
- Grammarly’s headline accuracy claim comes from the RAID benchmark and is Grammarly’s own characterization. It is not a rate you can expect on your own text.
- Grammarly Authorship is a different feature. It records how a document was written (typing, pasting, edit history) instead of guessing from the finished wording.
- Grammarly itself says AI detection should never be the only basis for judging originality.
What does Grammarly’s AI detector actually report?
According to Grammarly’s support article on the AI detection score, the score represents the percentage of scanned text that is likely AI-generated. It is described as an averaged estimate, and the same page says shorter passages are slightly harder to score accurately than longer ones.
Two limits stand out in that article. The score “doesn’t explain why the text may be flagged,” so a reader cannot tell which wording drove it. And because the detector uses a proprietary in-house model, its numbers can differ from those of other tools. A 30% in one product and a 12% in another are not a disagreement about the same quantity; they are two different models’ estimates.
This is the same kind of number most detectors produce. It describes a share of text that matched a pattern, not a probability that a person cheated. What an AI detection percentage means explains why.
What does Grammarly say about its own accuracy?
Grammarly’s AI detector page says the tool “achieves 99% detection accuracy and ranks #1 on RAID’s independent benchmark.” In its article on detector accuracy, Grammarly repeats the claim and adds that it builds its model “while minimizing false positives.” Neither page, as we read them, publishes a false-positive rate or describes the test conditions behind the 99% figure. The detector page also says the model was trained on text created before 2021, which is a reason to ask how it handles newer writing styles.
The same pages are notably cautious about use. Grammarly writes that detectors “are not 100% accurate,” that results are less predictable once AI text is edited or mixed with human writing, and that writers who use English as an additional language are “particularly vulnerable to misclassification.” The detector page adds that “AI detection should never be used as a standalone verification method.”
That candor is useful. A vendor claim of 99% and a vendor warning that the tool can get it wrong are both on the record, and the warning is the one to apply when you hold a result about a real person.
What is the RAID benchmark, and what does it show?
RAID is a public research benchmark described by Dugan et al. (ACL 2024). It contains a very large set of machine-generated texts across several models, domains and adversarial attacks, and it was built so that detectors could be compared on shared data. The authors evaluated 8 open-source and 4 closed-source detectors and concluded that “current detectors are easily fooled by adversarial attacks.” Detectors also struggled with changes in sampling settings and with generative models they had not seen.
Two things follow. A strong rank on a benchmark is real information about how a detector did on that benchmark’s mix of texts. It does not tell you the error rate on a particular student essay, a translated report or a lightly edited draft. And a benchmark score is a snapshot: models and writing habits change, which is why how accurate AI detectors are avoids a single universal figure.
Is there independent evidence on Grammarly’s detector?
Less than the volume of review articles suggests. We looked for peer-reviewed or neutral evaluations of the current Grammarly detector and did not find one we could check against its original source. Several third-party blogs report numbers, such as AI samples missed in a small test or accuracy dropping after editing. Many are published by competing detector vendors, use small or undisclosed samples, or cite reports we could not locate. We have not repeated those figures as findings.
What can be said with confidence is narrower. The independent RAID paper shows detectors in general lose reliability under adversarial changes and unfamiliar models. Grammarly’s own pages concede that mixed and edited text is harder. And because the product, its model and its plan structure change, an old review of “Grammarly’s detector” may describe something different from what runs today.
How is the AI detector different from Grammarly Authorship?
They answer different questions. Grammarly’s Authorship help article describes a feature that tracks the writing process in Google Docs, Microsoft Word and Grammarly’s own docs surface. It sorts text by how it entered the document, such as typed by the writer or pasted from another source. The report includes the full text color-coded by source, analytics on time spent, and a replay of typing and editing from start to finish. Attributing pasted text to a specific source needs clipboard access.
The detector looks at the finished wording and estimates how AI-like it is. Authorship looks at the process and shows what happened. Grammarly describes AI detection as a layer in paid Authorship reports that can flag text which “may have been AI-generated, even if it was not directly pasted into the document,” which is where the two features overlap.
| Question | AI detector | Authorship |
|---|---|---|
| What does it examine? | The text as it stands | How the text entered the document |
| What does it produce? | A percentage estimate | A source breakdown, report and replay |
| Can it show drafting history? | No | Yes, if it was on while writing |
| Main weakness | Cannot explain or prove authorship | Only covers writing done with it on, in supported apps |
Authorship cannot be applied after the fact to a document written elsewhere, so it helps a writer who set it up in advance. A detector helps when only the final text exists.
Who can use it, and on which plans?
Plan details differ by product surface and change, so check Grammarly’s current pages before relying on any of this. When we re-read them on 15 November 2026, the public AI detector page described a free tool that needs no sign-up to scan, and said Grammarly Pro adds an AI Detector agent that goes beyond a score. The support article says access depends on your plan: Pro users reach it through the Plagiarism agent on Windows and Mac, team plans through Microsoft Word, and Enterprise, Business and Education accounts through the AI Detector agent in docs when an administrator enables it. An earlier version of this guide said the article called detection “a paid feature”. The article does not use that wording, so we removed it. For Authorship, the help article says all plans, including free, get unlimited reports and “typed by a human” analytics, while paid plans add attribution of copied sources, including an AI-generated percentage. It also limits Word tracking to 20 pages.
How should you read a Grammarly AI score?
- Check the length. Grammarly says shorter passages are harder to score. A few sentences carry little evidence.
- Remember the number is a share, not a verdict. A percentage describes how much text matched a pattern.
- Account for tools used. Grammarly notes that rewrites from its own paraphrase and humanizer agents come from a language model, and can raise the score. Ordinary grammar fixes, translation and heavy editing also blur the line. AI-assisted vs AI-generated writing covers that gray area.
- Think about the writer. Non-native English writers and formal genres are the usual false-positive cases. See AI detection false positives.
- Look for corroboration. Drafts, version history, sources, a conversation and, where it exists, an Authorship report are stronger than any single percentage.
- Compare tools carefully. If two detectors disagree, that tells you the text is ambiguous. It is not a reason to pick the number you prefer.
Can an AI detector result prove cheating or authorship?
No. Grammarly says so about its own tool, and the research agrees. A score can justify a closer look; it cannot justify a penalty. If you are reviewing someone’s work, what to do when an AI detector flags student writing sets out a fair sequence. If your own writing was flagged, why your writing may show as AI explains common causes.
What about a second opinion?
Using more than one detector can show whether a result is stable, but it does not remove the underlying limits, and different tools measure different things. Our own checker uses the Pangram Labs API and labels passages AI-Generated, AI-Assisted or Human Written without giving reasons. The methodology page explains what it does and does not claim, and the accuracy page explains why we publish no accuracy figure. We make no claim that it is more accurate than Grammarly’s. If you already have the text, run it through our AI-generated text detector and review the flagged passages rather than relying on the overall score.
Frequently asked questions
Is Grammarly’s AI detector 99% accurate?
That is Grammarly’s claim about its performance on the RAID benchmark. The benchmark is public, but the 99% figure is Grammarly’s characterization, and the company publishes no false-positive rate for real-world text on the pages we read. Its own help pages say the detector is not 100% accurate and should not be treated as an objective source of truth.
Can Grammarly’s AI detector give false positives?
Yes, according to Grammarly. Its accuracy article names human text flagged as AI-generated as a possible error and says writers who use English as an additional language are particularly vulnerable to misclassification. Short, formal or heavily edited text is where caution matters most.
Does Grammarly Authorship detect AI?
Not in the same way. Authorship records how text entered a document and shows a source breakdown and replay. AI detection is a separate estimate from the finished wording, which Grammarly describes as a layer in paid reports.
Can my school see my Grammarly Authorship report?
Grammarly’s help article says reports can be shared through a unique link and that the replay can be excluded. Whether an instructor or institution requires one is their policy, so ask them. We have not verified how individual schools use it.