Claude Is Getting AI Watermarks as Anthropic Complies With EU AI Act

Anthropic will add watermarks to text generated by Claude and its other models as part of its compliance with the European Union’s AI Act transparency requirements, according to an updated company support page. Models released after August 2 will automatically mark generated text and files, with the system operating across Claude products including the Claude platform API, Claude, Claude Code, Claude Cowork and Claude Tag.

The change follows the EU AI Act’s Transparency Code taking effect on August 2. The rules require providers to make AI-generated or AI-edited material identifiable by other systems. Anthropic is implementing that requirement differently depending on the type of output. Generated files will use C2PA, an open standard for recording content provenance. For text, the watermark is built into model output rather than being limited to a particular Claude application or interface.

That model-level approach means developers using Claude through Anthropic’s API will receive watermarked output just as users interacting directly with Claude products will. Anthropic also said it intends to bring the technology to older models, although it did not specify when that expansion will happen.

The company says the text watermark is designed to remain attached to content as it moves between applications. It may also withstand at least some modifications after generation. “Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from,” Anthropic’s support page says.

One significant detail remains unresolved: Anthropic has not explained how extensively Claude-generated text can be edited before the watermark disappears. The company has been asked for clarification on that point.

Anthropic is not alone in committing to the EU framework. Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia have also agreed to follow the code. Other AI-related services have recently introduced their own approaches to identifying generated material. AI music platform Suno said last week that it would begin marking music created through its service following legal challenges. Substack, meanwhile, partnered with Pangram last month to identify AI-generated content. Substack CEO Chris Best has also criticized “Claudefishing,” a term for using AI-generated material as authored content.

For Anthropic, the implementation places the identifying mechanism directly within its models rather than restricting it to individual Claude products. That makes the watermark applicable across the company’s different interfaces and API-based uses, while the unanswered question of how well it survives subsequent editing will determine how persistent that identification is in practice.

This analysis is based on reporting from TechCrunch.

Image courtesy of Anthropic.

This article was generated with AI assistance and reviewed for accuracy and quality.