Overview
- Anthropic began embedding a machine-readable, invisible watermark into outputs of new Claude models since early August and is extending the feature to older models during the EU AI Act transition.
- The text signal is designed to travel with copied text and survive minor edits while file and image provenance will use signed C2PA metadata that can be lost when files are converted or screenshotted.
- Anthropic warns a detected watermark shows Claude processed a text but does not prove sole authorship, and it says heavy rewrites or very short excerpts can break or be unreliable for detection.
- Several professional users have publicly canceled or moved away from Claude citing fears that marked code or client deliverables could raise authorship or contractual issues, though Anthropic says it has not seen a measurable surge in cancellations.
- The practical effects will hinge on the detection tools and technical details Anthropic plans to publish, which will shape real-world legal, compliance, and workflow consequences for businesses and creators.