Anthropic will watermark all Claude text and images to comply with EU AI Act

Anthropic just committed to embedding invisible, machine-readable watermarks into every piece of text and images Claude generates. The move isn’t altruism — it’s compliance with the EU AI Act’s transparency requirements, which kicked in August 2nd and give existing products a four-month grace period to fall in line.

For images, Anthropic is adopting the C2PA provenance standard already used by Adobe, OpenAI, and Google. That metadata will be digitally signed and applied to supported files. The text watermarking is more mysterious. Anthropic describes it as an “imperceptible watermark” woven directly into Claude’s output — something that survives copy-paste and some light editing — but won’t name the system or explain how it works yet. Technical documentation is promised later.

The watermarking applies globally across all Claude surfaces: the API, the consumer chatbot, Claude Code, Cowork, and Tag. Even when you access Claude through AWS, Google Cloud, or Microsoft Foundry, the marks follow. That universality matters because it closes a loophole enterprise users might have exploited to dodge detection.

But here’s where the skepticism kicks in. C2PA metadata is notoriously fragile — platforms routinely strip it during upload, sometimes by accident. And Anthropic itself admits these systems are far from infallible, noting that unmarked content could still originate from generative AI. The text watermarking’s robustness is completely unproven. If it can’t survive a paraphrase or a round of editing in Google Docs, its practical value for platforms trying to detect AI-generated spam or misinformation collapses. Detection tools are supposedly coming, but Anthropic hasn’t said whether existing C2PA detectors like Google’s Gemini will work with Claude’s output. I’ve asked for clarification. The fanfiction community, which has been building its own Claude-detection systems for AO3, may end up being the canary in the coal mine here.