Anthropic’s push to make Claude’s output traceable has opened a harder question than whether AI should be labeled: what happens when the label follows work that was mostly human to begin with?

The dispute began after the EU’s AI transparency obligations took effect on August 2. Anthropic said new models would carry machine-readable marks from launch, with older models to follow; the system is being applied globally because the company says it cannot reliably limit it by region. Text marks are designed to survive copying and possibly some editing, while supported files will use C2PA provenance metadata.¹

Anthropic’s case is that the watermark is not a visible stamp or a hidden character. Instead, Claude uses a keyed pattern in otherwise low-stakes word choices—the sort of choice between “overcast” and “grey” that should not alter meaning. The company says the result is imperceptible, adds no identifying information, and allows only a probabilistic answer to whether Claude was involved. “Watermarking does not impact the quality of Claude’s output,” Anthropic said, adding that readers cannot distinguish marked text from unmarked text.²

But critics argue the distinction between generated and merely processed work may vanish once a detection result is shown to an employer, client or teacher. Ars Technica noted that model-level marking could catch material touched only by proofreading, translation or a minor edit—uses the EU rules explicitly treat differently from substantial generation. A detected mark, Anthropic acknowledges, means only that content “may have been processed by Claude,” not that Claude authored it.³

That ambiguity has already produced a consumer backlash. Some Claude subscribers say they cancelled over fears that watermarked code or documents could prompt authorship disputes or breach workplace rules; Anthropic says it has not observed a broader increase in cancellations. One consultant summed up the anxiety bluntly: “Even if you edit your own work that you wrote, they add the watermark anyway.”⁴

The company plans a detection API, and insists light proofreading may leave too little Claude-selected text to register. Yet the central tension remains: a tool intended to make AI use legible could also make ordinary human work look suspect.