Benchmarks & Proof

Detectors flag formal writing as AI because they measure style, not authorship.

Choosing the wrong judge model can systematically favor weaker outputs over better ones.

Multi-model blending reduces token-level repetition that single models cannot avoid.

Boosting watermark strength degrades text quality and task accuracy.

Vendor claims about AI writing tools outpace what buyers can independently verify.

Token-level edits better measure watermark evasion than word counts.

Benchmarking watermark removal requires reporting z-scores against disclosed detector thresholds.

Understanding z-scores helps teams distinguish watermark evidence from proof of origin.