Anthropic, the American AI firm, has announced plans to implement AI-generated text watermarks in its chatbot Claude to meet new European Union rules on AI-generated content.
The watermark will not be visible to users and will not disappear when text is copied and pasted.
But Anthropic reiterated that the watermark doesn’t imply that the content was generated by Claude in the first place.
The users could use Claude to proofread, translate, summarise or convert content, meaning that Claude could also give the mark to text that was created by other sources or by humans.
Read more: Google plans huge $40 billion investment in anthropic
Likewise, if you have a very short text or heavily edited material, there will be no recognisable watermark.
Anthropic said the technology would, therefore, serve as a way to suggest rather than to prove, that the content was generated by Claude, not the individual author.
If the file type is supported, like JPG image files, Claude will add digitally signed provenance metadata, which is an open standard to trace the history of digital content.
These changes are made to Claude models released on or after August 2, with earlier models also being upgraded.
While the measures are being implemented in order to comply with EU requirements, Anthropic noted it intends to implement the watermarking system globally.
The new AI regulations outlined in the EU mandate that any content created by AI be marked with machine-readable information, and that companies could be fined for breaches.
The EU’s voluntary code of good practice was recently signed by 190 organisations, including some huge tech firms such as Google, Meta, Microsoft, Mistral and OpenAI, as well as Anthropic.
Google has also created its own digital watermarking and detection algorithm.
But, despite the EU code encouraging watermark detection tools, Anthropic is still to introduce one.
Also read: US order forces Anthropic to disable advanced AI models