Anthropic, the US-based artificial intelligence company behind Claude, will introduce an invisible watermarking system for text processed by its chatbot, in a move aimed at meeting new European Union requirements for identifying AI-generated content.
The watermark, announced Monday in San Francisco, will not be visible to users and will remain embedded even when text is copied and pasted. Anthropic said the technology will be rolled out globally rather than being limited to Europe.
However, the company stressed that detecting a Claude watermark does not necessarily mean Claude originally generated the content.
“Claude may not be the original author,” Anthropic said, explaining that users commonly rely on the chatbot to proofread, translate, summarise or convert material. In such cases, the output may carry a Claude watermark even when the underlying text, ideas or data came from elsewhere.
The company also warned that very short texts or content that has been heavily edited may not retain a detectable watermark.
For supported file formats such as JPG, Claude will attach digitally signed provenance metadata based on an open standard designed to track the origin and history of digital content.
Anthropic said Claude models launched on or after August 2 will support the watermarking system, while older models will also be upgraded.
The move comes as the EU introduces rules requiring AI-generated content to carry machine-readable markings. Companies that fail to comply could face significant penalties.
Anthropic was among around 190 organisations that signed the EU’s voluntary AI code of practice last month, alongside major technology companies including Google, Meta, Microsoft, Mistral and OpenAI.
Google has separately developed SynthID, its proprietary digital watermarking and detection technology, and has urged other major technology companies to adopt similar standards.
Anthropic, however, has yet to release a watermark-detection tool, despite the EU code recommending such a capability.