Claude to watermark AI ‘slop’ – including edited human writing

Anthropic, a leading AI company based in Silicon Valley, is set to introduce invisible watermarks to text generated by its Claude AI models as part of EU regulations aimed at identifying AI-generated content. From August 2nd, Claude models launched in the EU will include machine-readable marks, and Anthropic is working to extend this capability to older models.

However, the company has clarified that the watermarks will also apply to human-written text if it has been edited or processed by Claude, which could complicate efforts to determine the true origin of the content.

While the watermarks are designed to remain intact when text is copied and pasted, and may survive some forms of editing, they may not always be detectable, especially with heavy editing or paraphrasing. Additionally, the watermarks will be applied to images and other files processed by Claude, represented as signed information showing the processing of the file.

Anthropic has also announced that it will provide tools for users and third parties to check for its watermarks, though technical details are yet to be published. The company stresses that the presence of a watermark does not definitively prove that Claude was the original author, as human writers could potentially have their work edited by Claude afterwards.

This move aligns with growing pressure from regulators and publishers to make synthetic content more identifiable in the face of a surge in machine-generated material online. Despite the EU AI Act requiring this marking only for AI systems performing an assistive function for standard editing or where the material is substantially altered or its meaning changed, Anthropic's decision to apply watermarks to all supported Claude output could potentially cover a broader range of content than the legislation strictly mandates.