Anthropic will apply invisible watermarks to all text and file outputs generated by its Claude artificial intelligence models worldwide, the company said, in a move designed to meet transparency obligations under the European Union's AI Act.
The policy, which took effect on 2 August, covers the consumer app, the Claude API, and models accessed through AWS, Google Cloud and Microsoft Foundry. Text will carry an imperceptible embedded watermark, while supported file types such as images will include digitally signed provenance metadata using the C2PA standard. The company led by Dario Amodei said it plans to release a detection tool in the coming weeks.
The change follows the activation of Article 50 of the AI Act, which requires providers of generative AI systems to mark outputs in machine-readable formats. Non-compliance can trigger fines of up to €15 million or 3% of global annual turnover. However, research has shown that watermarks can be removed through subsequent editing, and a method presented at the ACL 2026 conference aims to address the vulnerability by tying watermarks to textual meaning.