US company Anthropic has announced the addition of an invisible watermark to the text outputs of its Claude models in response to the transparency requirements of the European AI Act, but acknowledged that the watermark may become undetectable when text is heavily edited, translated or combined with other writing.
According to a report by US technology website TechCrunch, the announcement came after the European rules entered into force on August 2, 2026. The watermark applies to models released after the law took effect, while Anthropic plans to add it gradually to older models.
The transparency requirements under the European law oblige companies to label content generated or modified using artificial intelligence, allowing other systems and tools to identify it.
Embedded watermark and signed provenance data
Anthropic said the new watermark will not be visible to the naked eye and will cover text outputs across Claude tools, including external tools that use the model's application programming interface. Under this approach, the watermark remains with the text when it is copied and pasted directly from the model.
According to the company's help center, the identification mechanism consists of a watermark embedded in the text and signed provenance data within files produced by the model. However, the presence of the watermark does not prove that the content was written entirely by Claude; it indicates that the content passed through the model. Likewise, its absence does not confirm that the text was written entirely by a human.
Detection tools not yet available
Anthropic has not yet released the tools needed to read the watermark, nor has it disclosed the details of how the technology works. The company said it would soon provide tools to detect content generated using Claude, while external tools currently cannot identify the watermark, according to the help center.
The UAE's The National News reported that conventional AI detectors cannot detect Anthropic's new watermark because they look for previously known indicators.
Limitations raise questions about effectiveness
The company acknowledges that the watermark does not work with short texts and may become undetectable after heavy editing, translation or combining the text with other writing. These limitations raise questions about its effectiveness, as manual rewriting, converting screenshots into editable text or using paraphrasing tools could strip away the watermark and provenance data.
The National News report pointed to concerns that platforms could treat watermarked content differently, even though the watermark does not provide conclusive evidence that the text was generated entirely by artificial intelligence.
Similar technologies and compliance with European rules
Anthropic is not the first company to use technology to label AI-generated content. Google began using its SynthID system with content generated by its models in 2023 and made the technology open source to support the development of other detection tools. Users can use Gemini to check whether content carries a SynthID watermark.
According to TechCrunch, Meta, Microsoft and OpenAI have announced their commitment to the new European law, along with specialized tools such as Suno for generating music tracks and Synthesia for generating video clips.
The viability of Anthropic's technology ultimately depends on its ability to withstand modifications and on the company's provision of reliable tools for reading the watermark.
