Anthropic’s answer to AI-era transparency is an invisible signal embedded in Claude’s prose. Its critics see something more consequential: a digital stain that could follow human work long after an AI has merely polished it.

The dispute accelerated after the EU AI Act’s transparency rules took effect on August 2. Anthropic said future Claude models would watermark text globally because it has no durable way to limit the feature by region, joining a broader industry push toward machine-readable provenance markers.¹

The company says the mark is not a hidden character or visible tag. It uses subtle, low-stakes word choices to create a pattern detectable with a key, while leaving text “indistinguishable” to readers and without practical effect on quality, creativity or readability.¹ Anthropic also stresses the detector can only estimate whether text was “partly written by Claude” — not prove human authorship or identify another model’s work.¹

That distinction has not reassured users. By midweek, some Claude subscribers said they had canceled, worried that a marker surviving translation, proofreading, code review or light editing could make client work appear AI-authored. One consultant put the fear bluntly: “Even if you edit your own work that you wrote, they add the watermark anyway.”² Anthropic says a mark signals that Claude processed content, not that it originated it.

Critics argue that is precisely the problem. Ars Technica noted that the EU exempts routine editing such as grammar correction, but a model-level system cannot reliably distinguish a comma fix from wholesale generation.³ Tech analyst Ben Thompson framed the objection as a question of authorship: treating an LLM as the author of human-directed work, he wrote, is like demanding “a ballpoint pen advertise itself as the author.”⁴

Others see a useful, if imperfect, counterweight to synthetic content. Supporters say provenance signals could help audiences judge whether writing reflects a person’s own thinking and reduce the risk of AI systems being trained on an expanding pool of AI-made material.⁵ The argument now turns on whether Claude’s watermark clarifies that boundary — or blurs it further.