Claude is about to start watermarking text it merely edits, not just text it writes.
Anthropic said in a support article that it is rolling out machine-readable watermarks across its models to comply with the EU AI Act, which requires providers to mark AI-generated or manipulated audio, image, text, and video. The rule applies to any model released after August 2, with a grace period until December 2026 for models already on the market. Anthropic says all new models will carry the marks globally, not just in the EU, from the day they launch. Text gets an invisible embedded watermark; other generated files get digitally signed provenance metadata where the format supports it.
The EU's rule carves out an exception for assistive edits, like grammar correction, that don't substantially change a user's meaning, so Anthropic didn't technically have to watermark those. It's applying the watermark anyway, to essentially anything the model processes, wherever the format allows it. That's a wider definition of AI involvement than the law strictly demands, and it means routine edits get quietly tagged right alongside fully generated text.
For now the mark is invisible to readers, so the practical effect is mostly future-proofing, until someone builds a tool that reads it.