Estimating Watermark Proportions in Mixed Human-AI Text
September 1, 2026
Researchers identify that the proportion of watermarked content in mixed-source texts is unidentifiable in certain schemes. Only watermarking methods using continuous pivotal statistics allow for consistent and efficient proportion estimation.
HOW THIS AFFECTS YOU
●
builderEnsure your watermarking implementation uses continuous pivotal statistics if proportion estimation is a requirement.
●
policyDetection tools may struggle to accurately quantify AI involvement in hybrid human-AI content.