Claude's new Scarlet Letter watermark is invisible — for now

Per Ars Technica, Anthropic is rolling out machine-readable watermarks on all content processed by its models to comply with Europe's AI Act, which mandates that AI system providers mark generated or manipulated content. The watermarks will be embedded in text outputs and included as signed metadata in other file types, invisible to users but detectable by machines. All new models released globally will carry these marks from day one. Here's the tension: Anthropic is watermarking everything processed, even though the EU law exempts assistive functions like grammar correction. At the model level, the company can't distinguish a full generation from a minor edit, so Claude may end up marking exactly the kinds of assistance the law was written to leave alone. The watermarks aren't yet testable. Anthropic hasn't released a detection tool, so for now the invisible mark remains unproven.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton