Claude's new Scarlet Letter watermark is invisible — for now
Ars Technica Ashley Belanger ● Covered by 5 sources
Claude will soon watermark not just AI-made text, but even edited text. The marks are invisible, and Anthropic says it’s doing it to satisfy the EU AI Act.
Based on reporting by Ars Technica, Ashley Belanger — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic is about to start tagging content handled by its models with machine-readable watermarks. That includes more than freshly generated output. It also covers material the system merely processes, which is where this gets a little odd.
The company says the change is meant to line up with the European Union’s AI Act. That law requires providers to watermark AI-generated or manipulated audio, image, text and video. It applies to models released after August 2, with a grace period running until December 2026 for older models to be updated.
Anthropic says all new models it offers globally, not just in Europe, will mark AI-generated content from day one. Text will carry embedded watermarks that users won’t see. Other generated files will include digitally signed provenance metadata where the format supports it.
But the company is going further than the EU seems to demand in some cases. Anthropic says it is applying watermarks to all processed content where possible, even when the model is only doing an assistive edit like grammar correction, or when it doesn’t substantially change the user’s words or meaning. A model-level watermark can’t really tell the difference between a full rewrite and a comma fix, so Claude may end up stamping material the law appears to leave alone.
How strict this really is won’t be clear until Anthropic releases a detection tool that outsiders can test. The company says it plans to share details on how to detect the marks later, partly to provide the technical support the EU law calls for.
My take — AI-written commentary, not fact-checked reporting
This is the usual regulatory overcorrection: if a rule is hard to interpret, a company marks everything and calls it compliance. It’s safer for Anthropic, tidier for lawyers, and mildly absurd for anyone who thought “editing” and “generation” were different things. The AI industry keeps proving that the fastest way to simplify a messy policy is to make the product less useful.
Read more about this at: Ars Technica