Grammarly's AI detector and its plagiarism checker are two separate features doing two different jobs, which is where a lot of the confusion starts. Here's how the AI-detection side actually works, what Grammarly itself admits about its accuracy, and what independent testing has found when it's checked against real research benchmarks.
Trusted by students at
Grammarly's AI detector segments a document and runs each section against what Grammarly describes as a proprietary in-house model, trained on large sets of human- and AI-written text, looking for language patterns, predictability, and structure typical of tools like ChatGPT, Claude, or Gemini. It's built and scored separately from Grammarly's plagiarism checker, which compares text against existing published sources instead.
In 2024, Grammarly layered a feature called Grammarly Authorship on top of this, which reports whether a document looks human-typed, AI-generated, pasted in, or edited with Grammarly's own suggestions — a breakdown rather than a single AI-probability percentage, and it's continued to expand to more places you write (like Microsoft Word) since launch.
Grammarly's own support materials state the detector "is not 100% accurate" and gives "an averaged estimate rather than a definitive percentage" — it doesn't claim to conclusively determine whether AI was used, and it doesn't identify which specific model (ChatGPT vs. Claude vs. Gemini) might have produced a passage, only that text reads as likely AI-generated in general.
Grammarly also acknowledges bias against non-native English writing as a known driver of false positives, a limitation shared across pretty much every detector in this category, not unique to Grammarly.
Worth flagging upfront: the most detailed independent numbers here come from Originality.ai, a competing detector, testing Grammarly against the RAID research benchmark — so treat it as one data point from an interested party, not neutral ground truth. With that caveat, their published test found Grammarly scored around 22% accuracy on that benchmark (versus roughly 85% they reported for their own model), and that simple paraphrasing was enough to evade Grammarly's detection in their testing.
Separately, independent reporting from Plagiarism Today documented a more concrete flaw: running AI-generated text through Grammarly's own "Humanize" feature caused Grammarly's Authorship report to then relabel that same text as mostly human-typed — evidence that a tool's own detector and its own rewriting feature can end up working against each other.
Check your writing against Turnitin, GPTZero, Originality.ai, Copyleaks, and ZeroGPT methods at once, free, and see where they agree.
Compare Against 5 Detectors Free →