Anthropic is embedding machine-readable watermarks into text and attaching signed provenance metadata to images produced by its Claude models. The move satisfies the European Union's AI Act transparency requirements that took effect on August 2, 2026. The company says new Claude models launched on or after that date will mark content from day one, while older models will receive marking support during a four-month transition period.

The system uses two complementary techniques. For text, an imperceptible watermark is woven directly into the generated output at the model level. This means the signal travels with the content when copied and pasted and may survive some degree of editing. For image files —.png,.jpg, and.svg — Anthropic applies digitally signed provenance metadata following the C2PA open standard, the same framework used by Adobe, OpenAI, and Google.

What's new

  • Text watermarking: Embedded at model level across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag; persists through copy/paste and some editing.
  • Image provenance: C2PA-signed metadata attached to supported file types; signals tampering if metadata is stripped.
  • Cloud coverage: Watermarks apply when supported models run on AWS, Google Cloud, or Microsoft Foundry; C2PA metadata support varies by platform.
  • Detection tools: Anthropic plans to publish technical documentation enabling users and third parties to detect both watermark types.
  • Rollout timeline: Models launched on or after August 2, 2026 mark at launch; pre-August 2 models are being updated during the EU's grace period.

Anthropic has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content as both a model and system provider. The marking applies worldwide, not only in the EU, and covers every surface where supported Claude models operate. Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have also pledged to follow the EU code.

The company is hedging on reliability. A detected mark indicates the content may have been processed by Claude — not that Claude authored it. Proofreading, translation, or summarization of human-written text can still carry the watermark. Conversely, the absence of a mark does not prove human authorship; heavy editing, short passages, format conversion, or use of older models can all strip or prevent detection. C2PA metadata is also fragile: routine actions such as uploading files to social platforms routinely strip it. Anthropic has declined to disclose the technical underpinnings of its text watermarking method, and it remains unclear how much editing is required to remove the signal.

Why it matters

The EU rules require AI-generated or edited content to be marked in a machine-readable way that other systems can identify. Anthropic's model-layer approach means the signal follows the content regardless of which Claude product or cloud partner serves it. That design matters for platforms and publishers building detection pipelines: they can scan for a single watermark format instead of integrating per-product.

Beyond compliance, the move reflects a broader industry shift. AI music platform Suno recently committed to watermarking tracks after a wave of legal challenges, and newsletter service Substack partnered with Pangram to flag AI-generated writing. Substack's CEO Chris Best highlighted "Claudefishing" — the practice of using AI to generate content that passes as human — as a growing concern. These parallel efforts suggest watermarking is becoming a baseline expectation rather than a differentiator.

Our take

Model-level watermarking is the right architectural choice for interoperability, but Anthropic's refusal to document the text method makes independent robustness assessment impossible. Until detection tools ship and researchers can stress-test persistence against paraphrasing, translation, and adversarial stripping, the system remains a compliance signal more than a reliable provenance guarantee.

Sources