Anthropic has explained how it plans to watermark text generated by Claude as AI companies prepare for transparency requirements under the EU AI Act’s Code of Practice. The company says the watermark will be invisible to readers but detectable by parties with the right key.
The method relies on small choices between equally acceptable words. By nudging those choices into a statistical pattern, Claude can leave a signal in generated text without changing its apparent quality. Anthropic says it will use the SynthID-Text approach described by Google DeepMind and plans to release a detection API.
The company also spelled out limits. Light editing may not fully remove the watermark, but a complete rewrite can. Text that Claude only proofreads may carry little or no watermark because most of the words still come from the human author.
Code is another edge case. Because working code often leaves fewer arbitrary word choices, Anthropic says watermarking should have little effect on generated code beyond optional areas such as comments. The result is a transparency tool, not a universal detector for every AI-assisted document.