The measured bias
Published research has repeatedly found that classifier-style AI detectors flag text by non-native English writers at substantially higher rates than text by native speakers, on writing that was entirely human in both cases.
The mechanism is not mysterious. These detectors lean on fluency signals, meaning how predictable and how varied the wording is. Writing in an additional language tends to use a narrower, more standard vocabulary and more regular sentence construction, which is exactly what those detectors read as machine-like. The tool is measuring second-language writing and reporting it as AI.
Why this product measures something different
A provenance mark is not a fluency judgement. It is a statistical signature deliberately placed in text at generation time, and it is either present or absent regardless of how idiomatic the writing is.
WatermarkRemoverPro does also report a style measurement, and we are direct about what it is worth: it says how far your writing sits from a reference corpus of contemporary prose in that language. Non-native writing often sits some distance from it. That distance is not evidence of anything about how the document was produced, and the result says so in those words, on the page, so nobody can quote the number without the caveat.
Check in your own language too
If you also wrote a version in your first language, check that. Baselines exist for English, Spanish, French, German and Portuguese, and the comparison is often clarifying for a panel that has assumed fluency and authorship are the same thing.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. WatermarkRemoverPro's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.