Anthropic’s latest Claude models will embed machine-readable provenance data to meet the EU’s new transparency regulations, signalling a significant step towards combating AI misuse and boosting content accountability.
Anthropic has said its latest Claude models will begin attaching machine-readable provenance data to content used in the European Union, as the bloc’s new transparency rules start to bite. The move comes as Brussels presses AI developers to make it clearer when users are dealing with machine-generated material and how that content was produced or altered.
The company said models released on or after 2 August will support embedded metadata marking, including digital provenance signals based on Coalition for Content Provenance and Authenticity standards. Anthropic said the labels will be applied at the model level, meaning they should travel with text created through Claude across supported products and cloud partners, including Amazon Web Services, Google Cloud and Microsoft Foundry.
That places Anthropic among more than 200 signatories to the European Union’s Code of Practice on Transparency of AI-Generated Content, a voluntary framework tied to the AI Act. OpenAI, Mistral, Meta and Microsoft are also among the signatories. The European rules are designed to make synthetic content easier to identify as generative AI systems become more capable of producing convincing text, images, audio and video.
The limits of watermarking remain significant. Anthropic said the markers may disappear if text is heavily edited, paraphrased or translated, if the output is too small, or if file metadata is stripped during conversion or by taking screenshots. The company also noted that some human-written material may carry Claude markers if it has been processed through the system for reading or summarising. Research on EU AI Act compliance has similarly found that watermarking remains technically uneven, with recent academic work arguing that no current method fully meets the law’s requirements for reliability, interoperability, effectiveness and robustness.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
