Anthropic’s answer to Europe’s AI transparency push is an invisible mark in Claude’s prose. Its critics see something more consequential: a system that could turn a grammar check, translation or summary into an enduring suspicion of machine authorship.

The company began applying watermarks to supported Claude models released from August 2, saying the global rollout is needed because it lacks a durable way to limit the feature by region. The marker uses subtle patterns in word selection rather than hidden characters, and is intended to signal the likelihood that Claude was involved in producing text.¹

That distinction quickly became the fault line. Anthropic says its system “doesn’t say anything about ownership or authorship,” only whether Claude may have produced or processed material.¹ It argues that lightly proofread human writing should carry little or no detectable signal because Claude selected so few of the words.

Users and developers are less convinced that employers, teachers or clients will preserve that nuance. Critics fear a positive result will be read as proof that a document was wholly AI-written, even where Claude merely polished it. Guillaume Meyer, who released an open-source removal tool days after the announcement, framed the objection sharply: “I am all for content attribution. I am against the watermarking technique.”² His argument is that watermarking treats authorship as binary; his tool and other emerging projects aim to disrupt the statistical pattern through rewriting.

The technical limits reinforce both sides of the dispute. Anthropic says light editing will probably not erase a mark, but a complete rewrite can; short, factual and code-heavy passages provide fewer usable choices for watermarking.¹ Independent researchers and legal specialists note that removal tools occupy a gray area under the AI Act, though using them to pass AI work off as human-made could create separate problems.²

Supporters counter that the ambiguity is preferable to opacity. The new marks could help readers, schools and platforms identify synthetic material—and help AI companies avoid training future models on an ever-growing pool of AI output. But Anthropic’s promise that the watermark is “indistinguishable” to readers³ does not settle the harder question: whether an invisible record of AI assistance will be treated as evidence of deception.