Anthropic is watermarking Claude’s text output to make AI-generated writing detectable, according to The Decoder. The company says the system creates a statistical pattern in generated text, rather than adding visible markings or hidden characters. The Decoder reports that Anthropic’s method is based on Google’s SynthID-Text approach. In practical terms, the watermark changes the randomness source used during word selection, creating a pattern that can later be detected statistically. Anthropic says this does not affect the content, creativity, or readability of Claude’s output. That claim is now being challenged. The Decoder summarizes a critique from John Gruber, author of Daring Fireball and co-creator of Markdown, who argues that word choice cannot be treated as neutral. In his view, two synonyms do not carry exactly the same meaning, so a system that nudges Claude toward one word over another for watermarking reasons can still change the quality of the final text. Gruber’s objection is not that the watermark is visible. It is that the mechanism could alter the probabilities behind the model’s choices, sometimes making a less precise word more likely and a better one less likely. The Decoder notes that Gruber also questions whether the thumbs-up and thumbs-down rates cited in Google DeepMind’s SynthID study are a sufficient measure of text quality. The legal implications are narrower but concrete. The Decoder cites Artificial Lawyer’s assessment that Claude watermarks are mostly harmless for law firms in ordinary situations, especially where clients and courts do not object to AI assistance. The tradeoff changes when a client has explicitly barred AI use, or when a judge is skeptical of it: an AI contribution could be detectable even if the submitted document is accurate. Artificial Lawyer also raises a persistence problem. According to The Decoder’s summary, AI-marked clauses in a complex contract could carry forward into future templates, even when humans write the surrounding text. If multiple large language models are used, different watermarks could potentially overlap in a single document. Fee negotiations are another edge case. If clients argue that AI reduced the amount of human work required, a detectable AI share could become relevant in principle, The Decoder reports. That does not mean every AI-assisted matter becomes contentious, but it makes watermarking part of the operational record rather than a purely technical feature. Anthropic’s own caveat matters here: The Decoder reports that the company says watermarking is sparser in fact-heavy passages, where fewer word alternatives exist. For legal writing, where precision is central, the provided material says there are not yet empirical studies on the issue. There is also an enforcement gap. The Decoder reports that paraphrasing tools such as Declaude can strip the markings, which means watermarking may be better understood as a detection aid than a durable provenance guarantee. Who benefits: Organizations that want a way to detect AI-generated text benefit from a statistical watermark if it works as described. Clients, courts, and compliance teams may also gain leverage where AI use must be disclosed or restricted. Who's exposed: Law firms and other Claude users are exposed where AI use has been explicitly banned, questioned by a judge, or embedded into reusable templates. Writers and editors who care about exact word choice face an unresolved quality question rather than a settled technical fact.