Anthropic is embedding a text watermark into Claude output, turning AI-generated writing into something that can be statistically detected later. The company says the mark is invisible and does not change content, creativity, or readability. Critics are not convinced, and the legal sector may be one of the places where the tradeoffs become hardest to ignore.
How Claude watermarking works
The watermark does not appear as a visible symbol, label, or inserted character. Instead, Anthropic is using a method based on Google's SynthID-Text approach, which changes the randomness source used when Claude chooses words during generation.
That produces a statistical pattern in the final text. The pattern is meant to show that a passage came from an AI system, without adding anything a reader would normally see on the page.
Anthropic says this process has no effect on the substance or readability of Claude's writing. In its view, the watermark sits inside the model's word-selection process while leaving the result unchanged for practical purposes.
The rollout is tied to compliance with the EU AI Act. Anthropic is applying the watermark globally because it cannot limit the feature by region. All Claude models released after August 2 support the marking, and older models will be retrofitted in the coming months.
The quality dispute comes down to word choice
The strongest criticism in the source article centers on whether a watermark can really be neutral when it influences which words a model is more likely to choose. Blogger John Gruber, who has run Daring Fireball since 2002 and co-created Markdown, argues that small wording differences matter.
His objection is not that the watermark adds visible clutter. It is that a model generating language must constantly decide between near alternatives, and those alternatives are not truly identical. In his example, choosing between "overcast" and "grey" is not just a cosmetic switch.
Gruber's argument is that if a secret watermark key can raise the probability of one word and lower the probability of another, the system can sometimes push Claude away from the best semantic fit. He rejects Anthropic's claim that the difference is "imperceptible".
He also questions whether thumbs-up and thumbs-down rates in the SynthID study published by Google Deepmind in Nature are a strong enough measure of text quality. His point is that users may not downvote an answer simply because one word choice is weaker than another. As he puts it, a chatbot might write "bananas" instead of "pineapples" without the user flagging that as a quality failure.
Gruber even suggests that Gemini's reputation as weaker than Claude and ChatGPT could partly be connected to SynthID already being active. That remains a suggestion in the source, not a proven cause.
Why lawyers may care even when the text is accurate
Artificial Lawyer examined what Claude watermarking could mean for law firms and reached a mostly calm conclusion. Its assessment is that the markings are generally harmless, and that most clients and courts would not object to AI use. Some clients are now actively requesting it.
But the edge cases matter. If a client has explicitly banned AI use for a case, a watermark could make that contribution provable even when the document itself is factually flawless. The same issue could arise if a judge is skeptical of AI use.
In those situations, the issue is not whether the text is correct. It is whether the existence of a detectable AI contribution changes the transparency expectations around the work.
Artificial Lawyer also points out that watermarks can travel with text. A complex contract may include some AI-marked clauses while the rest was written by a human. If that contract becomes a template, those marked sections could move into future contracts as well.
The picture gets more complicated when multiple LLMs are used. Different watermarks could overlap inside the same document, making the authorship trail harder to interpret cleanly.
Fee pressure and circumvention
Claude watermarking could also affect fee negotiations. Artificial Lawyer notes that if clients argue AI made the work easier, the AI-generated share becomes verifiable in principle. That could give clients another basis for asking for discounts.
At the same time, the watermark is not presented as impossible to remove. The source article says paraphrasing tools like Declaude can strip the marking. Its developer James Padolsey criticizes the underlying EU regulation as arbitrary, arguing that it mostly affects ordinary users while doing little against deliberate circumvention.
Anthropic adds an important caveat of its own: watermarking is sparser in fact-heavy passages because there are fewer word alternatives available. For legal writing, where precision is central, that caveat matters. The source article also notes that there are not yet empirical studies on this specific legal-text question.
The result is a familiar AI governance problem. Claude watermarking may make AI output more transparent, but transparency can create new frictions in workflows that already depend on trust, precision, and clear responsibility. For law firms, the near-term task is less about panic and more about policy: knowing when AI use is allowed, when it must be disclosed, and how marked text may move through contracts and client work over time.