Cross-Lingual Watermarking Audit Finds Detection Gaps Are Structural to Language Family, Not Idiosyncratic

arXiv 2608.20047·medium signalXBlueskyLinkedInCopy link

LLM watermarking schemes are almost always evaluated on English using each scheme's own detection threshold and a narrow quality measure, which hides failures that appear under multilingual deployment. The proposed framework calibrates detection thresholds empirically per deployment context, adds a threshold-independent companion measurement to separate calibration failures from detection failures, uses three disjoint quality paradigms, and decomposes cross-language disparity over a typological family partition. Applied to six watermarking schemes, three open-weight generators, and eleven languages across four scripts and eight typological families, the observed disparity is predominantly between-family, meaning the fairness gap tracks language structure rather than specific languages.