Claude Now Leaves Machine-Readable Marks on AI-Generated Content

Anthropic's Claude AI tool is now adding invisible watermarks and provenance data to its generated text and files, in response to new EU transparency rules. This change took effect on August 2, 2024. The markings are applied to text and file outputs differently, allowing various systems to potentially detect their presence. Supported models place an imperceptible watermark in generated text, while file outputs can contain digitally signed metadata, creating a provenance record without visible labels.

This marking extends beyond the main Claude chatbot to the company's API and services like Claude Tag in Slack. However, the durability of these markings can diminish with changes to the wording, substantial rewriting, or translation. Detection tools and further technical guidance are still under development. Even if a detection is successful, it does not definitively prove authorship, as human-written material sent through the AI tool can still come back marked.

Organizations handling AI-generated content should maintain documentation of the AI's involvement in the workflow and adhere to any disclosure requirements. AI-use policies should outline what disclosures are necessary and how to handle disputed findings. Multinational teams should also take into account EU transparency requirements, as vendor changes made for Europe may impact workflows outside of Europe.

As detection tools improve, organizations will need to determine the weight a marker carries in disputes over authorship or disclosure.