Anthropic to Add Invisible Watermarks to Claude Text Under EU AI Act
Anthropic says it will add an invisible watermark to responses generated by its Claude model as it moves to comply with the European Union’s AI Act—an effort aimed at making machine-produced text easier to identify, even if the mark can’t be seen by readers.
Anthropic’s invisible watermark for Claude
The company announced a new approach tied to the EU’s AI Act requirements, which call for different AI models to identify content created by them using machine-readable methods. In practice, Anthropic plans to embed an “invisible” watermark directly into Claude’s generated text.
Anthropic’s position is that the watermark won’t be detectable by users and won’t change the meaning, quality, or readability of Claude’s output. It also claims the watermark is part of the text itself, so it should travel with the content when copied and pasted into another location.
How the company says detection will work
Anthropic says the watermark is based on imperceptible statistical patterns integrated during response generation. Because of that, dedicated tools can later analyze a document and determine whether it was produced by Claude. The company frames this as an important step toward improving transparency around AI usage as academic work, publications, and documents increasingly incorporate assistance from AI systems.
In its explanation, Anthropic emphasized that since the watermark is embedded within the text, it moves with copied content rather than being lost when users transfer the response elsewhere.
Experts warn the mark may be bypassed
Despite the company’s confidence, the initiative has drawn skepticism from experts who question how reliable watermarking will remain in real-world conditions. They argue that relatively simple alterations—such as rewriting passages, paraphrasing the content, or copying and pasting into certain environments—could make the watermark much harder to detect or potentially eliminate it altogether.
The concern reopens a broader debate over how effective these systems will be as AI generation tools and counter-measures continue to evolve.
Part of a wider industry push for traceability
This move lands as major AI companies, including OpenAI and Google, as well as Anthropic itself, pursue methods to distinguish human-authored material from AI-generated text. Even while effectiveness may vary, the decision reflects growing pressure on the sector to provide tools that support content traceability and help reduce misinformation and fraudulent uses of AI-generated material.
