Anthropic Details Claude Text Watermarking to Comply With EU AI Act
Anthropic has published a blog post outlining a new text watermarking system for its Claude chatbot. The implementation is designed to comply with the EU AI Act’s Transparency Code.
On Friday, August 15, 2026, artificial intelligence company Anthropic published a detailed blog post explaining how it intends to watermark text generated by its flagship chatbot, Claude. This technological development is being implemented specifically to comply with the EU AI Act and its associated Transparency Code regulations.
As part of this rollout, Anthropic plans to utilize the SynthID-Text approach originally outlined by Google DeepMind back in 2024. In addition to embedding these markers into generated text, the company also plans to release a dedicated watermark detection API for verifying content origin.
Addressing performance concerns, the company stated that "Watermarking does not impact the quality of Claude’s output." Elaborating further on the user experience, Anthropic noted, "To a reader, a watermarked response is indistinguishable from an unwatermarked one." However, the company added a caveat regarding heavily edited material, explaining: "In the latter case, of course, it’s arguable whether the text can any longer be described as AI-generated."
Discussing the technical limitations regarding vocabulary, Anthropic noted that the system can apply markers across "nearly all the words" in standard prose. Conversely, the company cautioned that in certain contexts, "there’s very little (if anything) for the watermark to attach to." Addressing programming applications specifically, the organization stated: "Having said that, in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code." They concluded regarding programming output: "But by definition, it will have a negligible effect on the actual code produced."
The broader discourse surrounding AI transparency features has drawn commentary from various industry observers and platforms. Reports from Business Insider have referenced "dozens" of instances, while reactions across the internet include a notable perspective from a Reddit poster who remarked, "The only reason you wouldn’t want this is to lie to people."
These developments involve numerous stakeholders across the technology sector, though official statements and technical documents primarily center on Anthropic, Google DeepMind, and regulatory frameworks within the European Union. Other entities mentioned in the broader reporting landscape include TechCrunch, Lorenzo Franceschi-Bicchierai, Anthony Ha, Lucas Ropek, Ivan Mehta, Russell Brandom, Aisha Malik, Zack Whittaker, Pangram, Reddit, Apple, Delta, and YouTube.
