What Grammarly’s AI flag is, and isn’t
Grammarly started as a grammar and style checker, and that’s still its core product. In recent years it added an authenticity or AI-detection style feature to some of its offerings, which flags writing it judges as likely AI-influenced. That feature sits alongside spelling, grammar and clarity suggestions in the same tool, not as a separate dedicated product the way a purpose-built detector is.
Like other classifier-based detectors, this kind of flag almost certainly works by measuring surface style patterns: sentence-length variance, vocabulary choices, how predictable the phrasing is, and similar statistical features learned from comparing human and AI text samples. That is fundamentally different from checking for an actual AI provenance watermark, which is a deliberate signature some AI systems embed and which requires the generating vendor’s own key to detect reliably. A style classifier isn’t looking for that signature at all; it’s pattern-matching against writing style.
Why genuinely human writing gets caught
A style classifier’s whole approach is a problem for anyone who writes cleanly, concisely, or in a fairly neutral register, because those are exactly the qualities that also describe a lot of AI output. If you write short, direct sentences, avoid a lot of stylistic variation, or use fairly common vocabulary, a classifier trained on “AI text tends to look like this” can land on your genuinely human writing without any AI involvement at all. The same is true if you used any AI assistance for polishing, like a grammar pass or a rephrase suggestion, even on writing you originally drafted entirely yourself: a blended flag doesn’t distinguish “this document was generated” from “this document was lightly touched up”, it just measures style, and a light touch-up shifts style enough to register.
This is the structural weakness of any single opaque flag: it collapses two different questions (does this carry a provenance mark from a specific AI system, and does this read stylistically like typical AI output) into one verdict, and then reports that verdict with a confidence that sounds more certain than a style guess actually is.
What neither of us can promise, and how NeverPrompted handles the two questions differently
Neither Grammarly nor NeverPrompted, nor anyone else, can tell you with certainty whether a given AI system’s undisclosed watermark is present in a document, because that requires a key only the model vendor holds. What NeverPrompted does instead is keep the two questions Grammarly’s flag blends together separate: a watermark check, which reports honestly what we found or didn’t find under the keys we hold, and a style measurement, which looks at the concrete AI-sounding patterns in your writing (stock phrasing, flat rhythm, overused connectors) without pretending that measurement proves provenance.
We also don’t report either of those as a single number dressed up as a verdict. If your writing gets flagged somewhere as “AI-sounding” and you know you wrote it yourself, that isn’t necessarily wrong about your style, and it isn’t proof of anything about how it was produced either. Reducing the style patterns that trigger these flags is a real, checkable thing NeverPrompted can help with, entirely on your device; certifying that a specific document has seen its last flag from any given detector is not something any tool, including this one, can honestly promise.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. NeverPrompted's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.