Anthropic has started embedding invisible watermarks into all AI-processed content from its Claude models. The global rollout is triggered by the European Union’s AI Act, which requires AI system providers to label generated or manipulated material. The company confirmed that all new models released after August 2 will carry the marks from day one, and older models must comply by December 2026.

What You Need to Know

The watermark, internally called Scarlet Letter, is machine-readable and invisible to users. It applies not only to fully generated content but also to any text or media processed by Claude, including minor edits like grammar fixes. Anthropic says it will eventually share detection tools, but for now the marks cannot be independently verified. The policy covers all Claude models globally, not just those in the EU.

Watermarking Mechanics

Anthropic’s approach applies watermarks at the model level, meaning any content that passes through Claude receives a digital stamp. Text outputs carry embedded watermarks that are invisible to human readers but detectable by automated systems. Other generated files, such as images or audio, include digitally signed provenance metadata where the format supports it.

  • Text outputs: Receive an embedded machine-readable watermark that does not alter readability.
  • Images and audio: Get digitally signed provenance metadata where file formats support it.
  • All content types: Marks are applied even when Claude performs only minor edits, like grammar correction.

Regulatory and Practical Implications

The EU’s AI Act requires watermarking for AI-generated outputs, but it carves out exemptions for cases where AI performs an assistive function for standard editing. Anthropic has acknowledged this gray area, stating that its model-level approach cannot distinguish wholesale generation from a small comma fix. As a result, Claude may stamp content that the law was designed to leave unmarked. The true extent of overapplication will remain unclear until the company releases a detection tool that third parties can test.

Why This Matters

This global watermarking policy sets a precedent for AI transparency, but it also raises questions about user privacy and the practicality of blanket labeling. For developers and businesses using Claude, the marks could create unexpected metadata in edited documents. Regulators will need to evaluate whether the overreach undermines the AI Act’s intent to avoid penalizing routine edits. Anthropic’s decision to apply the watermark universally, rather than limiting it to fully synthetic content, may pressure other AI providers to adopt similarly broad measures. The Scarlet Letter system, once its detection tools are released, will likely become a reference point for how effectively watermarking can be audited.

What Comes Next

Anthropic plans to publish technical details on detecting the watermarks, fulfilling the EU’s requirement for support. Independent researchers and watchdogs will then be able to examine how thoroughly the marks survive editing and transformation. The company’s approach may influence how other AI firms, particularly those releasing models after the August deadline, structure their compliance. The European Commission has not yet commented on whether Anthropic’s universal application meets legal expectations.