Histoire
août 13, 2026
Claude’s Invisible Watermarks May Blur the Line Between AI and Human Writing
Anthropic’s global watermarking push aims to satisfy EU transparency rules and expose AI ghostwriting. But critics warn the marks could follow lightly edited human work, disappear under rewrites and deepen confusion over what “AI-generated” really means.
Anthropic is preparing to make Claude leave an invisible trace on its output worldwide, turning Europe’s AI transparency rules into a new test of whether technology can distinguish ghostwriting from a grammar check.
The clock started on August 2, when the EU AI Act’s new transparency obligations took effect. Anthropic says every new Claude model released from that point will mark generated text from launch, while supported files will carry C2PA provenance metadata. The policy will apply globally, across Claude’s consumer products, API and cloud distribution—not merely inside Europe.1
Anthropic’s case is straightforward: the marks are meant to give users and platforms a better way to identify AI involvement. “Text from supported Claude models, including output from Claude Code, will carry an invisible watermark,” the company said, adding that it plans to release a detection API.2 The company says the watermark does not alter a response’s meaning, quality or readability.
But the rollout exposes a sharper problem than a simple label can solve. Watermarking is applied at the model level, meaning a human-written press release cleaned up, translated or formatted by Claude may receive the same signal as a fully AI-drafted document. Anthropic acknowledges that output can carry a mark even when “the underlying ideas, text, or data originated from another source.”3
That distinction matters to schools, publishers and employers, where a detected mark could be treated as proof of authorship rather than evidence of AI assistance. Meanwhile, the system is not a lock: substantial rewriting, mixing text, translation or paraphrasing can weaken detection, while metadata in images and video can be stripped away.4
The result is an awkward new norm. Watermarks could make it harder to secretly submit AI-written homework or pass off a novel as wholly human work. Yet for the far more routine uses of AI—emails, meeting summaries and internal documents—the stigma may be limited. The central question is no longer only whether Claude touched the text, but whether people will mistake that touch for authorship.5