Anthropic adds invisible watermarks to Claude AI text to meet EU rules

Anthropic has begun embedding machine-readable, imperceptible watermarks into text generated by its Claude AI models. The rollout, which started for new models launched on or after August 2, 2026, aims to make synthetic content identifiable to other systems regardless of where the text is published.

These watermarks are embedded directly at the model level, ensuring they apply across every Claude product and user surface. Because the watermark is integrated into the text itself, it is designed to travel with the content during copy-pasting and may persist even after some editing. This deployment extends beyond Anthropic's own interfaces to include models accessed via Microsoft Foundry, Google Cloud, and AWS, although the company noted that signed provenance metadata for files may vary depending on the specific platform used.

Regulatory Pressure from the EU

The initiative is a direct response to the European Union's AI Act, specifically Article 50 and the associated Code of Practice. This regulatory framework mandates that AI-generated or edited content be marked in a machine-readable format to prevent deception and facilitate the identification of synthetic media. By implementing these markers, Anthropic aligns its global output with the EU's transparency requirements, joining other major AI labs in the effort to curb the spread of undetected synthetic content.

A Shift in AI Detection

This move represents a fundamental shift in how AI-generated text is identified, moving from probabilistic detection to deterministic detection. Traditional AI detectors typically guess whether a text is synthetic based on stylistic patterns and linguistic probability, a method often prone to false positives. In contrast, watermarking looks for a deliberate, embedded marker, providing a more reliable method of verification.

This change significantly complicates the ability to pass off AI-generated text as human-written. The implications are broad, impacting academic integrity in education and the broader fight against "AI slop" and the proliferation of AI-generated misinformation across the web.

Future Rollouts and Detection

While the current rollout focuses on models released after August 2, 2026, Anthropic plans to extend watermarking support to its older models. Additionally, the company intends to publish the technical detection methods for these watermarks, allowing third-party systems to verify the origin of the text. Observers will be watching to see how resilient these markers remain against aggressive rewriting and whether other major LLM providers adopt similar deterministic standards to satisfy global regulatory demands.