Guide

Grammarly AI Detector in 2026: How It Works, How Accurate It Is, and What It Flags

A plain-English 2026 guide to the Grammarly AI detector: the free checker and the paid AI Detector agent, what Grammarly's 99% accuracy claim is measured on, what the RAID benchmark it cites contains, what Grammarly says the detector goes easier on, how Grammarly Authorship differs from detection, and Superhuman's planned acquisition of GPTZero.

September 26, 2026 · Updated September 27, 2026 · The AI Rankings

Quick answer: The Grammarly AI detector is a free web checker at grammarly.com/ai-detector that returns a percentage estimate of how much of a passage was likely AI-generated, plus a paid “AI Detector agent” built into Grammarly Pro, Business, Education and Enterprise that marks the specific sentences it suspects. Grammarly’s headline claim is that it “achieves 99% detection accuracy and ranks #1 on RAID’s independent benchmark” — but RAID is an academic benchmark built from 2023-era models: its corpus contains no output from Claude or Gemini, two of the model families the same page says the detector catches, and no ChatGPT text newer than GPT-4. RAID’s published metric is true-positive rate at a fixed 5% false-positive rate rather than “accuracy”. Two things follow that matter more than the number. Grammarly’s own support documentation states the model is deliberately tuned to miss AI text rather than risk false accusations, and that it is “more likely to assign a higher percentage” when the text came from an outside provider such as ChatGPT or Gemini — which, by Grammarly’s own account, makes text from its own AI features less likely to score highly. And Turnitin, the detector that actually decides academic cases, explicitly says it flags Grammarly’s draft-generation and paraphrasing output, so a clean score in Grammarly is not a prediction of a clean score at your university.

This guide covers Grammarly’s detector specifically: the two products, the accuracy claim and what it is measured on, what the score does and does not cover, and how Grammarly Authorship fits alongside it. For the underlying detection mechanism across all tools, see how AI detectors work. For ranked alternatives with independent accuracy scores, see best AI detectors. For the tool that institutions actually run, see Turnitin AI detection. For Grammarly as a writing tool rather than a detector, see best AI grammar checkers. For the assistants on the other side of the desk, see best AI for students.

Who owns Grammarly now, and why it matters here

Grammarly the company no longer exists under that name. On 29 October 2025 the company rebranded itself as Superhuman, consolidating Grammarly, Coda and Superhuman Mail under one corporate identity and launching a Superhuman Suite and an assistant called Superhuman Go. The writing product kept its name: Grammarly is still Grammarly, and the company’s statement to existing customers was that “nothing changes about the products they rely on”. Throughout this guide, “Grammarly” means the product and “Superhuman” means the company that ships it.

On 23 June 2026 Superhuman announced its intent to acquire GPTZero, the best-known consumer AI detector (BusinessWire). GPTZero’s own announcement the same day said it “plans to join Superhuman” and that “nothing changes for our daily users” (GPTZero). Superhuman chief executive Shishir Mehrotra framed the deal as “bringing the most trusted writing tool and the most trusted AI detector into one platform”. No completion of that acquisition has been announced as of 26 September 2026, three months after the announcement, and the deal value was not disclosed.

If the deal closes, one company will own a leading AI writing assistant, a free AI detector, a paid AI detector agent, GPTZero’s consumer detector, and two writing-process trackers — Grammarly Authorship and GPTZero’s Replay. Superhuman’s press release describes the company as having 40 million daily users.

What the Grammarly AI detector actually is

There are two different products sharing one name, and most of the confusion about this tool comes from conflating them.

Source: Grammarly’s AI detector page and the AI Detector user guide.

The free checker returns a single score; the paid agent marks the text. Grammarly’s support documentation places the agent in Superhuman Go and in documents for paying subscribers, in the browser extension for Google Docs and the desktop clients for Microsoft Word on Pro, Plus, Business and Education plans, and in documents by default for Enterprise, Business and Education, with administrators controlling access. Grammarly’s announcement of its benchmark result refers to the agent, not the free tool, while most third-party tests run the free tier.

It is not an institutional integrity system. There is no instructor-side dashboard that scores a class’s submissions, no LMS integration that attaches an AI percentage to a gradebook entry, and no equivalent of the Turnitin AI writing indicator appearing inside a similarity report. Where Grammarly sells to institutions, the product it leads with is Authorship, covered below. The detector is a writer-facing self-check.

Grammarly publishes no length limits or language list for it. Turnitin states a 300-word floor, a 30,000-word ceiling and four supported languages. Grammarly’s detector page and support guide state none of these, and no figures for minimum length, maximum length or supported languages are published: data not available.

How Grammarly’s detector works

Grammarly runs a proprietary in-house classifier rather than a published method, and its own documentation describes the training corpus two different ways.

The support guide states: “Our model is trained on hundreds of thousands of human and AI-generated texts.” The consumer product page states: “Our detection model was trained on tens of thousands of texts, including both human-written and AI-generated text created before 2021.”

The two figures differ by an order of magnitude. Read literally, AI text “created before 2021” would predate ChatGPT’s November 2022 launch; the date may describe only the human-written half, as Turnitin does when it validates false positives against papers written before ChatGPT existed, but Grammarly does not say so. Both wordings are quoted verbatim here.

On mechanism, the product page describes the model as examining sentence structure and predictability, repetition and uniformity, metadata traces from AI tools, and comparison against known AI-generated content. Grammarly publishes no model card, no architecture statement, no per-model detection rates, and no explanation of what “metadata traces” means in a plain pasted-text workflow. For the general mechanism — perplexity, burstiness, and why trained classifiers replaced both — see how AI detectors work.

The output is a percentage plus sentence-level marking. In the agent, flagged sections are underlined in the document and shown as bars in a side panel; the free checker returns the percentage alone.

What the score means, according to Grammarly

Grammarly’s support documentation makes four statements about how to read the percentage.

It is tuned to miss AI text on purpose. In Grammarly’s words: “The AI detection model is optimized to minimize false positives, as we believe that incorrectly identifying human-generated text as AI-generated is more detrimental than failing to accurately catch AI-generated text.” This is the same trade-off Turnitin describes, and it means a low score is weak evidence that no AI was used.

It scores outside providers’ text higher. Grammarly states: “The model is also more likely to assign a higher percentage of AI-generated text in a given passage if it’s been generated by an outside LLM provider (such as ChatGPT or Gemini).” By that description, text produced by Grammarly’s own rewriting, drafting and paraphrasing features is less likely to be scored highly than the same amount of text from ChatGPT or Gemini.

There is no score-lowering feature. Grammarly states: “Currently, we do not offer a dedicated feature to reduce the AI score in user content.” Tools marketed as AI humanisers are a separate category.

The number is not a verdict. From the same guide: “The percentage result should not be used as an objective source of truth, as AI detection of any kind can be prone to errors.” The public page adds that there is “no AI detector that can conclusively or definitively determine whether AI was used”, that detection “should be just one part of a holistic approach to evaluating writing originality”, and that it “should never be used as a standalone verification method”.

Grammarly and Turnitin treat Grammarly’s AI differently

Turnitin explicitly flags Grammarly’s generative features while ignoring its grammar corrections. Turnitin states that its detector is not tuned to target spelling, grammar and punctuation corrections, and that in testing those changes were not flagged in most cases — but it explicitly excludes Grammarly’s draft generation, paraphrasing and summarising features, which “will likely be flagged as AI-generated” (Turnitin FAQs).

Combine that with Grammarly’s statement that its own model goes easier on Grammarly-generated text, and a clean score from Grammarly’s detector does not predict a clean score from Turnitin on the same document — least of all for text Grammarly’s own AI wrote. If your institution runs Turnitin, Grammarly’s checker is not a pre-submission safety net. Turnitin also does not let students self-check at all, so there is no way to close that gap by testing.

The question has reached a court. In Matter of Newby v Adelphi Univ. (2026 NY Slip Op 26021), decided 28 January 2026, the academic integrity report against the student alleged use of Grammarly, after a professor obtained a 100% Turnitin AI score. Justice Randy Sue Marber of the New York Supreme Court, Nassau County, annulled the finding as “without valid basis and devoid of reason” and ordered the record expunged (full opinion). The case is covered in detail in our Turnitin AI detection guide.

What Grammarly claims about accuracy

Grammarly’s claim, verbatim from its detector page, is that the tool “achieves 99% detection accuracy and ranks #1 on RAID’s independent benchmark, outperforming other leading AI detectors.”

The ranking claim traces to a Grammarly blog post published 10 February 2026, Grammarly’s AI Detector Agent Ranks #1 in Quality. That post says the agent “ranked #1 on RAID’s leaderboard for AI detection quality” and describes RAID as testing over 670,000 texts across writing styles and models, including attempts to fool the detectors.

The blog post announcing the result states no accuracy percentage. The 99% figure appears on the marketing page, not in the post announcing the benchmark result it is attached to. Grammarly does not publish a false-positive rate, a per-model detection rate, or the leaderboard score behind the ranking. Its own caveat on the same page is that “no AI detector is 100% accurate”.

What the RAID benchmark actually measures

RAID was published at ACL 2024 by Liam Dugan, Alyssa Hwang, Filip Trhlik, Josh Magnus Ludan, Andrew Zhu, Hainiu Xu, Daphne Ippolito and Chris Callison-Burch, and is maintained by Dugan at the University of Pennsylvania (arXiv:2405.07940, submitted 13 May 2024). The paper describes the dataset as “over 6 million generations spanning 11 models, 8 domains, 11 adversarial attacks and 4 decoding strategies”.

Four features of that benchmark change how the claim should be read.

Source: the RAID repository and the ACL paper.

The model list matters most. Grammarly’s detector page says the tool identifies text from ChatGPT, Gemini, Claude, Microsoft Copilot and Grammarly’s own AI. RAID’s corpus contains no Gemini or Claude output, and its OpenAI text comes from ChatGPT and GPT-4 as they stood in 2023 — nothing from GPT-5 or later. A benchmark result is evidence about the models in the benchmark, so a RAID ranking does not measure performance on the models most people use in 2026.

The RAID paper was written to test claims of this kind. It notes that many detectors “claim to detect machine-generated text with extremely high accuracy (99% or more)”, and its headline result is that “current detectors are easily fooled by adversarial attacks, variations in sampling strategies, repetition penalties, and unseen generative models”. Every model released after the benchmark was built is, to RAID, an unseen model.

“#1 on RAID” is also not one title. Trinka’s AI Content Detector publicly claims it “ranked #1 on the RAID Benchmark for academic text detection, achieving a score of 0.999/1.000”, and SuperAnnotate claims a #1 open-source result on the same benchmark. Because RAID scores detectors across separate domains and adversarial splits, multiple vendors can hold a genuine first place on different cuts at the same time. None of the three claims is necessarily false, and none of them means “the most accurate AI detector”. We could not retrieve the leaderboard’s per-detector figures at the time of writing, so Grammarly’s actual RAID score is recorded here as data not available.

What independent testing actually found

No peer-reviewed study has measured Grammarly’s AI detector. The strongest recent academic work in this area — the preregistered study by Marijke Van Vlasselaer, Filip Van Droogenbroeck and Bram Spruyt published in the International Journal for Educational Integrity in June 2026, which tested 160 academic papers of 4,000 words or more — evaluated GPTZero, Copyleaks, Turnitin and Pangram (Van Vlasselaer et al.). Grammarly was not among them.

The only third-party test we can find that included Grammarly was run by a competing detector vendor.

Four caveats apply to that 0 of 9. The sample is twelve documents, which is too small to establish a rate. The test was run by Pangram, which sells a competing detector and placed itself first. It used the free tier, not the paid agent that Grammarly’s benchmark claim concerns. And Pangram states its own limitation: “This test is not definitive of the validity of these tools.”

On that small set of 2024 and 2025 model output, the free Grammarly checker flagged none of the AI text and none of the human text — consistent with Grammarly’s stated tuning towards avoiding false positives.

The evidence base for this tool is thin. Grammarly publishes a benchmark ranking on a corpus that predates current models and a percentage that does not correspond to that benchmark’s metric; the independent record is one small test by a competitor. No published evidence supports a precise accuracy figure for Grammarly’s detector in 2026.

Grammarly Authorship: recording the writing process

Grammarly’s other product in this area is not a detector, and it is free.

Grammarly Authorship records how a document was written rather than guessing whether a machine wrote it. It categorises text by origin — typed by the writer, pasted from a website, generated with AI assistance, or pasted from a non-web source — and produces a shareable report that can replay the document’s construction in order (Grammarly Authorship). It works in Google Docs, Microsoft Word, Superhuman docs and Canvas (Grammarly).

Three facts determine whether it is useful to you.

It is free and available on Grammarly Free as well as Pro. Unlike the AI Detector agent, Authorship is not paywalled.

It is opt-in and the writer controls disclosure. Grammarly’s page states: “You decide when to share your Authorship Report. Grammarly doesn’t auto-send it to your professor or anyone else.” Institutions cannot pull a report a student has not shared.

Since 10 March 2026 it can name which AI agent was used. Grammarly added agent-specific attribution, so a report can distinguish whether and how a writer used particular Superhuman agents such as Reader Reactions, Citation Finder or Proofreader, rather than recording “AI” as one undifferentiated category. Grammarly reported more than 5 million Authorship reports generated since the feature entered beta in October 2024.

The limitation is shared by every opt-in process tracker. Because Authorship is opt-in and the writer decides whether to share the report, it is evidence a student can offer, not a check an institution can impose. A writer who wants to conceal AI use simply does not enable it, and an absent report proves nothing. It helps a writer show their own process; it does not catch anyone who does not volunteer a report.

Grammarly’s faculty-facing material explains the case for process evidence over detection. Its post states that detectors “can be biased against non-native English writers and neuro-diverse learners” and that “detection models frequently fail to accurately identify AI-generated text from the latest models” (Grammarly for institutions, 15 October 2024). Both points are consistent with the published research on detector bias.

The same direction of travel is visible across the category, from Turnitin Clarity on the institutional side to provider-side AI watermarking at the point of generation.

How Grammarly’s detector differs from the tools it gets compared with

This is a positioning comparison, not a ranking. For ranked accuracy across the whole market, with independent scores, see best AI detectors.

Sources: Grammarly AI Detector user guide, Turnitin FAQs, BusinessWire. Where a figure is marked “not published here”, the vendor does not state it on the pages cited — we have not substituted an estimate.

The main difference is who runs it. Grammarly’s detector is a self-check a writer runs on their own text. Turnitin’s is run by an instructor, the student cannot see it, and its result can feed a disciplinary process. A score from one says nothing reliable about the other.

What it costs

Source: Grammarly plans, read 26 September 2026. All prices in USD. Grammarly Authorship is free on every tier, including Grammarly Free.

What to do, by situation

Best for a writer who wants a free sanity check before submitting: the free Grammarly checker, with its limits understood. It costs nothing and needs no account. It does not predict what Turnitin will say, and because the tool is tuned to avoid false positives, a low score is weak evidence that no AI was used.

Best for a student who wants to be able to prove they wrote something: Grammarly Authorship, switched on before you start. It is free, it records the process as you go, and it cannot be enabled retroactively, so it has to be on before you start writing. Pair it with your platform’s own version history.

Best for a student who has already been accused: your writing process, not a counter-test. Do not run the essay through three detectors and submit the screenshots. Preserve document version history, drafts, notes, outlines and timestamps. Ask which tool produced the score and what the exact figure was. Check whether your institution’s policy permits detector output as sole evidence, — both Grammarly and Turnitin say in their own documentation that it should not be. The Adelphi judgment turned on process evidence and on the institution’s failure to consider it.

For an instructor deciding whether to rely on it: Grammarly advises against using it as sole evidence. Grammarly’s faculty-facing post states that detection models “frequently fail to accurately identify AI-generated text from the latest models” and can be biased against non-native English writers and neuro-diverse learners. For classroom tooling that does not rest on detection, see best AI for teachers.

Best for an institution that wants transparency rather than policing: process capture, and say so in the assignment brief. Authorship, Turnitin Clarity and comparable trackers produce observable behaviour rather than statistical inference. They only work when students know in advance what is expected and consent to record it, which makes this a policy design question before it is a software question.

If you write in English as a second language, or you have a diagnosed learning difference, put that in writing early. Both the UK Office of the Independent Adjudicator cases analysed by HEPI and the Adelphi judgment turned on institutions failing to consider circumstances that plausibly explain an unusual or polished style. That is a procedural failure you can name in an appeal.

Summary

• Two products: a free web checker that returns one percentage, and a paid AI Detector agent that marks suspected passages inside documents.
• The accuracy claim: “99% detection accuracy” and “#1 on RAID”. RAID’s corpus predates current models and contains no Claude or Gemini text, its published metric is not called accuracy, other vendors also claim first place on it, and Grammarly’s announcement of the ranking prints no percentage.
• The tuning: Grammarly says the model is optimised to avoid false positives, scores outside providers’ text higher, and should not be treated as an objective source of truth.
• Turnitin: flags Grammarly’s generative features, so a clean Grammarly score does not predict a clean Turnitin score.
• Authorship: a free, opt-in record of how a document was written, which the writer chooses whether to share.
• Ownership: Grammarly is a product of Superhuman, which announced a deal to acquire GPTZero on 23 June 2026; no completion had been announced by 26 September 2026.

Frequently asked questions

Does Grammarly have an AI detector?

Yes. Grammarly offers a free AI detector at grammarly.com/ai-detector that returns a percentage estimate for pasted text with no sign-up required for a basic scan, and a paid AI Detector agent included in Grammarly Pro, Plus, Business, Education and Enterprise plans that additionally underlines the sections it suspects and shows per-section bars in a side panel. Both are self-check tools for writers. Grammarly does not sell an instructor-side detection dashboard, and the product it markets to institutions is Grammarly Authorship rather than the detector.

Is Grammarly’s AI detector accurate?

Grammarly’s detector page claims it “achieves 99% detection accuracy and ranks #1 on RAID’s independent benchmark”, but that figure has not been independently verified. RAID’s corpus was built from 2023-era models and contains no Claude or Gemini output, two of the model families the same page says the tool detects; RAID’s published metric is true-positive rate at a fixed 5% false-positive rate rather than “accuracy”; and Grammarly’s own blog post announcing the ranking states no percentage. No peer-reviewed study has measured Grammarly’s detector. The only third-party test we can find that included it was published by competitor Pangram in January 2026 and scored the free tier 0 of 9 on AI text with no false positives on human text, across a sample of just twelve documents.

Is Grammarly’s AI detector free?

Yes, in its basic form. The web checker at grammarly.com/ai-detector is free and needs no account for a basic scan, returning a single percentage. The AI Detector agent, which marks individual sections inside your document, requires a paid plan — Grammarly Pro is $12 per month billed annually, or $30 billed monthly, as of 26 September 2026. Grammarly Authorship, the process tracker, is free on every tier including Grammarly Free.

Does Grammarly’s AI detector detect ChatGPT?

Grammarly’s page says the detector identifies text from ChatGPT, Gemini, Claude, Microsoft Copilot and Grammarly’s own AI tools. That is a coverage claim, not a measured per-model detection rate, and Grammarly publishes no accuracy figures broken down by model. The benchmark it cites, RAID, includes 2023-era ChatGPT and GPT-4 text but nothing from GPT-5 or later, and no Claude or Gemini output, so the ranking does not measure detection of current ChatGPT models.

Will using Grammarly get flagged as AI by Turnitin?

It depends entirely on which Grammarly features you used. Turnitin states its detector is not tuned to flag spelling, grammar and punctuation corrections, and that in testing those changes were not flagged in most cases. But Turnitin explicitly says Grammarly’s draft generation, paraphrasing and summarising features “will likely be flagged as AI-generated”. So running Grammarly’s proofreader over your own sentences is a different risk from having Grammarly rewrite or generate them. Because Turnitin does not let students self-check, you cannot test this before submitting.

Does using Grammarly count as cheating?

That is set by your institution’s policy, not by Grammarly or by any detector. Most policies distinguish between correcting your own writing and generating text, and the safe approach is to read the specific assignment brief, disclose the tools you used, and keep the drafts that show your own reasoning. Grammarly Authorship exists partly to make that disclosure straightforward, and permitted use that is documented survives a flag far better than undocumented use of any kind.

Can Grammarly detect its own AI writing?

It flags it, but by Grammarly’s own account less readily than it flags other tools’ output. Grammarly’s support documentation states the model “is also more likely to assign a higher percentage of AI-generated text in a given passage if it’s been generated by an outside LLM provider (such as ChatGPT or Gemini)”. Turnitin takes the opposite approach and explicitly includes Grammarly’s generative features in scope. If you drafted with Grammarly’s AI, Turnitin is the more likely of the two to flag it.

What is the difference between the Grammarly AI detector and Grammarly Authorship?

The detector inspects a finished piece of text and estimates whether a machine wrote it, which is statistical inference from writing style. Authorship records how the document was produced as you write it, categorising text as typed, pasted from the web, pasted from elsewhere, or AI-assisted, and can replay the construction in order. Authorship is free on all tiers, opt-in, cannot be enabled retroactively, and is shared only when the writer chooses. Since 10 March 2026 it can also identify which specific Superhuman AI agents were used rather than logging “AI” as one category.

How do I lower my Grammarly AI detection score?

There is no feature for it and Grammarly says so directly: “Currently, we do not offer a dedicated feature to reduce the AI score in user content.” If the score is high because AI generated the text, the answer is to write or substantially rewrite the passage yourself, and to disclose whatever assistance you used under your institution’s policy. If the score is high on writing that is genuinely yours, the fix is evidence rather than editing: enable Authorship, keep version history, and raise it as a false positive, which Grammarly acknowledges its detector can produce.

Why does Grammarly say my own writing is AI-generated?

All detectors produce false positives, and they cluster on writing that is formal, repetitive, structurally uniform, or produced by second-language and neuro-divergent writers — a pattern Grammarly acknowledges in its own institutional materials. Grammarly states its model is optimised to minimise false positives and that “the percentage result should not be used as an objective source of truth”, so a flag on your own work is a known failure mode of the tool rather than a finding about you. Preserve your drafts and version history, and if the score is being used against you, ask what the institution’s policy says about detector output as sole evidence.

Is Grammarly’s AI detector better than GPTZero?

There is no sound basis for ranking them against each other on current evidence, and the question may soon be moot: Superhuman, Grammarly’s parent company, announced on 23 June 2026 its intent to acquire GPTZero, with no completion announced as of 26 September 2026. GPTZero has been measured in peer-reviewed work and Grammarly has not, which makes their evidence bases non-comparable rather than one better than the other. For ranked accuracy across the market with independent scores, see our best AI detectors page.

Did Grammarly buy GPTZero?

Superhuman, the company formerly named Grammarly, announced on 23 June 2026 that it intends to acquire GPTZero, describing it as bringing “the most trusted writing tool and the most trusted AI detector into one platform”. The deal value was not disclosed and no completion has been announced as of 26 September 2026. GPTZero said at the time that it would continue as a standalone product and that “nothing changes for our daily users”. Note also that the company renamed itself from Grammarly to Superhuman on 29 October 2025 while keeping Grammarly as the name of the writing product.

Written 26 September 2026 from Grammarly’s product page, support documentation, plans page and company blog as published on that date, plus the RAID benchmark paper and repository, the BusinessWire and GPTZero acquisition announcements, and one third-party comparison whose competing commercial interest is stated in the text. Grammarly’s product page and support guide describe the detector’s training corpus differently, and both wordings are quoted verbatim. Where Grammarly publishes no figure — length limits, supported languages, per-model detection rates, false-positive rate, and the RAID leaderboard score itself — this page records “data not available” rather than estimating.

← All guides