Anthropic Watermarks All Claude Outputs Globally to Comply With EU AI Act, Enabling Binary AI Detection

The company behind Claude is launching a system to identify machine generated text through a pair of new detection tools and an API. Anthropic is adopting a global watermarking initiative to satisfy a voluntary Code of Practice under the EU AI Act, electing to apply the measure to all users worldwide rather than limiting it to European markets.

The watermarking method employs word choice patterns instead of hidden characters to ensure outputs remain indistinguishable to human readers. While the system applies to new models immediately, existing models will transition by December. A full rewrite of the generated content is required to remove the patterns, as the markers are permanent and not removable through standard editing.