Anthropic’s new invisible markers promise to make Claude less of a silent collaborator. But the same system designed to expose undisclosed AI output could also put a machine-readable label on ordinary workplace polishing.

The shift follows the EU AI Act’s transparency requirements, which took effect on August 2 and require generated or manipulated material to be identifiable by other systems. Anthropic says new Claude models released after that date will automatically mark generated text and files, while older models are being brought into line during the compliance period.¹

The company’s approach has two tracks. Claude-generated files will carry C2PA provenance metadata, while text will receive an “imperceptible watermark” embedded at the model level. Anthropic says the mark travels when text is copied and pasted and “may persist through some editing,” across products including its consumer chatbot, API and cloud-hosted services.²

That breadth is the point for transparency advocates: publishers, schools and platforms could gain another tool to investigate work passed off as wholly human. It also arrives amid mounting pressure on platforms to label AI material, with Substack partnering with detector Pangram and other major AI companies signing on to the EU code.³

Yet the disclosure test is far messier than a simple authorship verdict. Anthropic acknowledges that proofreading, formatting or translating a human-written document can leave a detectable signal. Meanwhile, heavy rewriting, short passages, mixing text from other sources or translation can weaken detection—or remove it altogether.¹

That creates two competing anxieties. Students and novelists may face sharper scrutiny, particularly where authorship is central. But for workers using AI to draft emails, meeting summaries and internal documents, the real question may be whether being “Claude-handed” is actually shameful. One commentator argues most users will simply accept the disclosure: routine office writing is not judged by the same authorship standards as a term paper or a prize novel.⁴

The watermark, then, may make AI assistance more visible without making it easier to judge. A detected mark does not prove Claude wrote the work; no mark does not prove it did not.