Anthropic Introduces Watermarking in AI Outputs to Comply with EU AI Act
This week, Anthropic announced that all Claude products released from 2 August 2026 onwards will feature machine readable watermarks in AI generated content. The company has until 2 December 2026 to implement watermarking for older Claude models.
The introduction of watermarking is in response to the EU AI Act, specifically Article 50(2). This provision mandates that providers of AI systems generating synthetic content must ensure that outputs are marked in a machine readable format, clearly indicating whether they are artificially generated or manipulated.
The EU AI Act aims to promote trustworthy artificial intelligence while safeguarding fundamental rights as outlined in the Charter of Fundamental Rights of the European Union. Anthropic has assured users that the watermark will be invisible and will not compromise the quality or readability of the output produced by its large language models.
The watermark may be embedded within the text of the output or included in the metadata of audio files, images, or videos. Anthropic has also indicated its commitment to assisting users and other stakeholders in detecting whether a piece of content contains a watermark, although details on the detection process remain forthcoming.
This development could have significant implications for various industries, including the legal sector, where issues of confidentiality, accountability, and data provenance are critical. Notably, Anthropic plans to implement this feature globally, demonstrating the influence of EU regulations on international practices, akin to the global adoption of standards following the General Data Protection Regulation.
However, there are limitations to the watermarking feature. Users may be able to extract watermarks from metadata through processes such as resaving and reformatting. Additionally, when using Claude for tasks like proofreading or summarising, the original text may not be generated by the AI, even if a watermark is present in the output.
The absence of a watermark does not necessarily indicate that AI was not involved in the content's creation. The EU AI Act acknowledges that the watermarking obligation does not apply when AI systems perform an assistive function for standard editing or do not significantly alter the input data.
Article 99 of the EU AI Act outlines penalties for non compliance, which can include administrative fines of up to €15 million or up to 3% of total worldwide annual turnover, whichever is higher. Other AI companies, such as Google DeepMind, have also developed tools that embed invisible watermarks in various content types.
In addition to the legal obligations, a voluntary Code of Practice on Transparency of AI generated Content has been developed, with over 190 organisations, including major players like Google, Meta, OpenAI, and Microsoft, having signed on.