Anthropic is set to implement invisible watermarks on text produced by its Claude AI models. This change, which began with models launched in the EU from 2 August, is part of new EU rules aimed at making AI-generated content easier to identify.
The Silicon Valley AI company confirmed on Tuesday that these machine-readable marks will apply worldwide wherever supported Claude models, including Claude, Claude Code, and its developer platform, are available. The embedded watermarks are designed to persist through copying and pasting, and potentially some editing.
However, the system will also mark some text originally created by humans if it has subsequently been edited or processed by Claude. Anthropic acknowledged that the presence of a watermark does not definitively mean Claude authored the original material, as it could appear on an article proofread or translated by the AI.
Anthropic has cautioned against using the marks as conclusive proof of AI authorship, noting that heavy editing or paraphrasing can make watermarks undetectable. Short passages may also lack sufficient text for a reliable signal, and metadata on images can disappear during file conversion.