Commitment to Transparency in AI
Anthropic has pledged to comply with the new transparency requirements set forth by the European Union. The company recently announced that all Claude models released from 2 August onwards will incorporate a labelling system. This initiative will enable audiences, platforms, and third-party services to identify content generated by these artificial intelligence (AI) systems.
This measure aims to fulfil Article 50 (2) of the EU’s AI Act, which, as of earlier this month, mandates technology companies to clearly label content related to public interest matters that has been created or manipulated using this technology. Anthropic stated, “As AI-generated content becomes increasingly prevalent, greater transparency and information regarding its origins can provide users with valuable context about the information they consume. To promote transparency and adhere to our legal obligations, Anthropic is working to include machine-readable labels in the content generated by Claude,” as detailed in their support page.
The labelling system announced by Anthropic will be implemented by default in its forthcoming AI models. However, the company is still in the process of ensuring compatibility with its existing algorithms and products, as the law provides a four-month grace period for compliance.
Understanding the Labelling Mechanism
Anthropic has clarified that the labelling will vary according to the format of the generated content. For instance, images or graphics created by AI will feature metadata that indicates their origin, based on the open standard established by the Coalition for Content Provenance and Authenticity (C2PA).
On the other hand, text materials will be identified through invisible watermarks that will remain intact even when the content is copied and pasted onto third-party platforms. These watermarks may endure after specific edits as well. “The watermark will be applied at the model level, meaning it will be present regardless of the Claude product or the original source of the text. You won’t see it, and it does not alter the meaning, quality, or readability of Claude’s response,” the company elaborated.
Furthermore, Anthropic is developing a mechanism that will allow users and third parties to detect the watermarks and provenance metadata that will identify content created by Claude. The company has promised to provide further details regarding this mechanism in the near future.
Global Application of Machine-Readable Labels
The machine-readable labels will be globally applied to compatible Claude models, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. They will also be present when accessing Claude models via AWS, Google Cloud, or Microsoft Foundry.
However, Anthropic acknowledges that its labelling system has certain limitations. For example, if a system detects a Claude label, it indicates that the content may have been processed by the model, but it does not necessarily prove that the AI originally generated the ideas, information, or text.
The company also noted that there may be instances where the system does not include a detectable label for content generated or processed by Claude automatically. This might occur if the material was produced by a model released before 2 August, if the text was edited, paraphrased, translated, or significantly mixed with other content, or if the segment is too short to detect a reliable signal. Additionally, the metadata of a file may be lost when converting, saving, or taking a screenshot of the content.
Ongoing Developments and Regulatory Compliance
Consequently, Anthropic cautions that the presence of a label should be regarded as an indication of provenance rather than definitive proof of human or artificial authorship. With this announcement, Anthropic positions itself as one of the first companies to express its commitment to adjusting its products in line with the EU’s current AI Act.
Nevertheless, the measures outlined only partially address the requirements stipulated by the regulation. For instance, the regulatory framework mandates that AI-generated content must combine invisible labels and metadata, along with visible icons to indicate its provenance. It also requires interactive AI systems, such as chatbots, to inform users explicitly from the first interaction that they are engaging with artificial assistants.
It is possible that the company will soon announce modifications regarding these aspects, as it is also a signatory of the Code of Good Practice related to the EU’s AI Act.