AI text watermarking and quality loss

The piece examines whether text watermarking required by AI transparency rules degrades generated writing. It challenges claims from Anthropic and Google that watermarking preserves quality, arguing that their human evaluations focus on individual responses while users typically exchange multiple messages, retry answers, or compare outputs. The key distinction is between the quality of one response and the diversity of a model’s responses: a watermark can leave any single answer readable while steering repeated answers toward similar wording. The cited SynthID-Text research acknowledges this diversity-versus-detectability trade-off and warns that reduced diversity can lower the quality of systems that generate several candidates and select the best.

The author says this matters beyond automated selection. A user requesting alternate sentence endings may encounter more repetitive retries, and writers seeking creative possibilities may see a narrower range of ideas. Repeated prompts used to assess a model’s confidence can also become less informative if their answers are no longer independent. As an example, an independent test of SynthID-Text settings found that four of five watermarked descriptions of a coastal trip began with nearly the same phrase, compared with more varied openings from the unwatermarked model. The piece also notes that shared short token sequences might cause similar effects across different prompts. It concludes that watermarking sacrifices some creative range, and that proposed mitigations such as trying many times or rotating keys dilute the effects rather than eliminate the underlying trade-off.