Claude watermarks all processed text to comply with EU AI Act

The EU AI Act, which applies to all AI models released after August 2, requires providers to watermark AI-generated or manipulated audio, image, text, and video outputs, with compliance required by December 2026 for previously released models. Anthropic's approach of watermarking all processed content—even minor edits—reflects a deliberate choice to go further than the law mandates. The regulation itself exempts assistive editing functions (such as grammar correction) and content that doesn't substantially alter the user's text, but a watermark applied at the model level cannot distinguish between a full rewrite and a comma fix. This mismatch between technical capability and legal intent creates a problem: users and downstream readers may interpret any watermark as signifying wholly generated content, when in fact it may only indicate that an AI touched the text. The watermark's weakness—that it is trivially easy to defeat through copying, pasting into another tool, or editing metadata—further undermines its reliability as a durably informative signal. Anthropic acknowledges that a detected mark "is not fully conclusive" and that "the content may have been processed by Claude," leaving significant room for misinterpretation, especially among the general public who may not grasp the difference between processing and generation.