Anthropic will begin embedding an invisible, machine-readable watermark in text produced by its Claude models released on or after Aug. 2, a move aimed at making AI output easier to identify and aligning with new transparency rules under the EU AI Act. The model-level marker is designed to persist through copy-and-paste and some light editing, though extensive rewrites or translations may strip it out; images processed by Claude will also carry indicators of processing and signs of tampering. The push comes as platforms and publishers confront a surge of low-quality “AI slop,” with Substack rolling out reader-triggered detection tools and YouTube tightening monetization rules for templated, AI-driven channels. While watermarking is unlikely to stop determined actors, Anthropic’s approach reflects mounting pressure from regulators and users for clear signals that distinguish human work from machine-assisted content—without penalizing benign uses like proofreading or translation.
Related article:
