What these tools are actually measuring
Language models tend to pick likely next words, so their output has low “surprise”, technically, low perplexity. A classifier built to spot AI writing is largely trained to notice that low-surprise pattern: predictable word choices, predictable sentence construction, predictable structure. It isn’t reading your text for who wrote it; it’s comparing statistical patterns to examples it was trained on.
A related measure is variation across a document, sometimes called burstiness. Human writing naturally swings between simple and complex sentences, focused and rambling paragraphs, because a person’s attention and energy vary as they write. A classifier trained to notice AI output is also, by construction, trained to notice the absence of that variation. Either way, it’s pattern-matching against a labeled dataset, and it inherits every gap and bias baked into that dataset.
Why careful or non-native English trips the wire
Someone writing fluently in a second language often produces textbook-correct, grammatically regular prose, fewer idioms, fewer unusual constructions, more consistent sentence patterns. Statistically, that regularity looks closer to model output than a native speaker’s looser, messier first draft does, even though it’s entirely human and often the result of real skill and effort.
The same thing happens with heavily edited writing. Running a paragraph through a grammar or style checker nudges it toward smoother, more regular phrasing, the exact regularity a classifier is trained to notice. A careful copyedit and a model’s output can end up statistically closer to each other than either is to a rough, unedited human draft. Formal registers, academic, legal, technical writing, already sit closer to that same smoothed pattern, because both humans and models are trained on, or aiming at, the same conventions.
What a flag actually tells you, and what it doesn’t
A flag is a statistical estimate with a real false-positive rate, not a verdict on who wrote something. Treat it as a pointer at specific sentences, usually the ones with the flattest rhythm or the most stock phrasing, worth a second look, rather than proof of anything about you.
It’s also a different kind of signal from an actual AI watermark, a deliberate statistical mark some systems embed in their own output, testable with the right key. A style flag guesses from resemblance; a watermark test looks for a specific planted signal. But even a watermark check has its own honest limit: nobody outside the vendor holding that key can test it with full certainty, so no result from any tool, style-based or watermark-based, settles the question outright.
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.