The tells you can check for right now
Start with vocabulary. AI drafts default to a small set of safe, faintly impressive words that say very little, the word above is one of a whole family of them (this site keeps a fuller list, if you want to check yours against it). The same habit shows up in transition words strung between almost every paragraph, whether or not the ideas actually need stitching together. If you can cut a transition word and the sentence still makes sense, it was filler.
Next, look at shape. Print a page and glance at where the sentences end: if every line breaks at roughly the same length, that’s worth noticing. Human writing naturally varies, a blunt six-word sentence next to a longer one that unpacks it. AI writing tends to smooth that out into a steady, even rhythm, the same paragraph shape recurring section after section instead of actually varying with what’s being said.
Last, check for specifics. Vague, confident claims (“this approach offers significant advantages”) with no number, name, date, or concrete example attached are a strong signal. So is a total absence of hedging on things that actually deserve doubt, next to constant hedging on things that don’t; “it’s worth noting that” is the classic filler here. Real writing commits to a position somewhere and admits uncertainty somewhere else, unevenly, because that’s what actually thinking about a subject looks like.
Why your own eyes only get you halfway
People are bad at spotting statistical uniformity across a whole document. You notice one stock phrase, not that a document uses the same handful of them a dozen times, or that every paragraph runs to within a few words of the same length. Once a paragraph reads smoothly, most readers, including the person who wrote it, stop scrutinizing it. That’s exactly the blind spot polished, AI-flavored prose exploits.
There’s also a harder limit: if a passage came from a system that quietly biases its own word choices during generation, no amount of careful reading reveals that. That kind of signal, an AI watermark, is built to be invisible to a human reader and only recoverable with a keyed statistical test run against the actual text.
The reverse is true too. Careful, well-edited, or non-native English can fail the “eyeball” tells above for reasons that have nothing to do with AI, formal register and correct grammar just look regular. So a page that passes your read-through isn’t necessarily clean, and a page that trips a tell or two isn’t necessarily AI-written. You need both a specific checklist and a proper statistical test, not one or the other.
What NeverPrompted’s check adds
NeverPrompted runs two separate checks on your draft, entirely on your device. A style scan flags the exact tells above, the stock phrase, the hedge, the run of same-length sentences, with the specific passage that triggered it, so you can see and fix the actual sentence rather than stare at a single opaque score. A separate check runs a statistical test aimed at the kind of watermark some AI systems apply to their own output.
Checking is free and unlimited on the Free plan; the point is that you can run it as often as you draft, not just once at the end. Nothing is uploaded for either check: the text stays on your device the whole time.
One honest limit, stated plainly: no check, including this one, can promise a specific result against a watermark keyed by a vendor who has never shared that key. What NeverPrompted gives you is real, checkable evidence, the specific phrases and patterns, and where any statistical signal concentrates, so you can make an informed edit, not a score to chase.
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.