Anthropic adopts EU AI Act code, watermarks Claude output

Anthropic has committed to the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content and is turning on machine-readable marking across Claude, according to a Claude support article. Any Claude model released on or after August 2, 2026 supports marking at launch, and models launched before that date are in progress.

The mechanics split by output type. Text carries an imperceptible watermark that, per Anthropic, does not change 'the meaning, quality, or readability of Claude's response' and persists through copying and pasting, though heavy editing can degrade it. Generated files in .svg, .png, or .jpg formats receive signed metadata following the Coalition for Content Provenance and Authenticity (C2PA) standard. Coverage extends across the Claude Platform (API), Claude, Claude Code, Claude Cowork, Claude Tag, and cloud partners including AWS, Google Cloud, and Microsoft Foundry.

The announcement slots into a wider EU compliance push our tracker has followed closely since Brussels made AI content labels and watermarks compulsory on August 2. It also joins 357 Anthropic items we have logged over the last 90 days. Anthropic positions itself as an early public signatory of the Article 50(2) code in both roles the regulation names, as a generative AI model and as a system provider.

Anthropic's own caveats deserve close attention. The company states plainly that 'a detected mark provides a signal that content was processed by Claude, but is not fully conclusive,' and that marks may be absent if content was edited, paraphrased, or generated by pre-marking models. The support page does not disclose the watermarking algorithm, publish a false-positive rate, or give a dated timeline for bringing older Claude models into compliance.

For publishers, platform trust and safety teams, and regulators, the asymmetry matters. A C2PA-signed PNG from Claude gives you something concrete to key policy on; an unmarked essay gives you nothing at all. The useful posture is to treat marking as a positive signal only, and to design any enforcement or moderation flow around that limitation from the first day it ships.