Are LLM Watermarks Truly Harmless? Devs Call for Open Replication and Evals

max_paperclips · x · 2026-08-12

A developer points out that the true impact of LLM watermarks remains unknown without knowing the exact method, replicating it on open models, and evaluating with and without the watermark.

While LLM text has statistical breathing room, making watermarks "mostly harmless," the author argues that zero impact is unlikely—a wrong token at a critical moment could still significantly affect the output.

Original post →

More from Research

Research channel →