Anthropic to watermark all Claude-processed content

Anthropic will watermark content processed by any of its models, not just text generated from scratch, to comply with the European Union’s AI Act.

The company said in a support article that all new models offered globally — not just in the EU — will carry machine-readable watermarks from release. Text outputs will carry embedded, invisible marks; other generated files will include digitally signed provenance metadata where supported, using the C2PA approach for non-text content.

The EU AI Act requires providers of AI systems to watermark AI-generated or manipulated audio, image, text, and video outputs. It applies to models released after August 2, 2026, with a grace period until December 2026 for previously released models.

The watermarks will affect anyone using Claude for proofreading, translation, summarization, or file conversion, even when the underlying ideas or text came from another source. That goes beyond the law’s assistive-function exemption for standard editing or grammar correction that does not substantially alter text or meaning.

A detected mark is not fully conclusive, the company said. It signals that content was processed by Claude, but the absence of a detected mark does not mean content was not AI-generated or processed.

Marks may also appear on content Claude did not generate. They will travel with copied-and-pasted text, persist through some editing, but could be destroyed if pasted into another chatbot that edits the text.

Anthropic said watermarks will not work on some platforms or features that do not support them. It plans eventually to share detection details to provide technical support required by EU law.

The company did not immediately respond to Ars Technica’s questions about a timeline for detection details, testing results on false negatives or positives, or how watermarks may conflict with standard editing exemptions.

Violations under the AI Act can carry fines up to 15 million euros or 3 percent of a company’s worldwide annual revenue. Anthropic said it will continue working on watermarks and detection methods that meet EU demands.

Ars Editor-in-Chief Ken Fisher contributed to this report.