Claude to start watermarking AI-generated text – but will it make quality worse?
TL;DR
Anthropic plans to watermark text produced by Claude, changing how the model makes small random choices as it writes. The move responds to EU rules on labelling AI-generated content. Critics question whether the intervention will degrade writing quality, since those random choices shape style and word selection. Anthropic has published no quality measurements alongside the announcement.
Nauti's Take
Verifiable labelling of AI text is real progress for anyone who has to check provenance, from newsrooms to recruiters. The risk sits in the method: changing how the model samples its small random choices touches style and word choice directly, and Anthropic has published no quality numbers yet.
Teams running Claude on client-facing copy should pull their own before-and-after samples once the rollout lands, instead of trusting the announcement.
Sources