How NeverPrompted Actually Works

The rewrite: what it actually changes

The rewrite is an on-device language model that specifically targets the tells covered elsewhere on this site: stock transitions, hedging phrases, showy vocabulary, a dash used as an all-purpose connector, a flat, uniform sentence rhythm across a whole document. It isn’t a generic paraphraser that shuffles synonyms; it’s aimed at the specific patterns that make writing read as AI-generated in the first place.

It works on your own draft, keeping your meaning and structure, and you choose how heavy a pass to run. A light pass might just break up a run of same-length sentences and cut a handful of stock phrases; a heavier pass rewrites more aggressively while staying anchored to what you actually said. Either way, you see and approve the result, it’s an edit, not a black box.

One honest limit: this is a style edit. The words are genuinely changed, by a model acting on your instructions, but no tool can promise a rewritten passage will clear every checker every time, because style-based classifiers keep changing what they key on, and nobody, including us, controls that.

The watermark check: what it actually tests

Some AI systems embed a deliberate statistical signature in their own output, a genuine provenance mark, not a style guess. The specific mechanism, after the method described by Kirchenbauer et al. in 2023, works by splitting a model’s vocabulary into a “green” list and a “red” list at each generation step, using a hash seeded by a secret key and the words already generated, then quietly biasing generation toward the green list. Over enough text, that produces a statistical skew that wouldn’t happen by chance.

NeverPrompted’s check runs a keyed statistical test aimed at exactly that kind of skew, entirely on-device, and reports how strong any signal is and which passages carry it. The honest catch: only the party holding the actual key used at generation time can test with full certainty. An outside test, including this one, is necessarily an estimate against a scheme it wasn’t given the key to, which is exactly why no honest tool can promise a specific result against an undisclosed vendor’s watermark. A clean result here means this test found no such signal under the keys it holds, not that the text is cleared. For someone dealing with an actual accusation who needs the deeper statistical and forensic detail behind a specific case, that’s a more specialized job than this feature is built for, and a separate product exists for exactly that side of it.

Why nothing is ever uploaded, on either feature

Text you haven’t published yet, or a document someone’s already questioning, is sensitive by nature. Sending it to a server for either job would mean someone else’s infrastructure has a copy of it. Running both the rewrite model and the watermark test on-device avoids that by how the product is built, not by a policy promise you have to trust.

In practice, that means the model doing the rewriting and the test doing the watermark check both execute locally on your machine. Your draft, and the result of checking it, stay there.