started · updated
Anthropic implements invisible watermarks in Claude AI to meet EU transparency rules
Anthropic has announced the implementation of invisible, machine-readable watermarks and digitally signed provenance metadata for its Claude AI models. This initiative is designed to ensure compliance with the European Union AI Act, specifically Article 50, which mandates transparency for AI-generated or manipulated content.
For text outputs, Anthropic is embedding imperceptible watermarks directly into the model's responses. These marks are designed to be invisible to users and do not alter the meaning, quality, or readability of the text. Crucially, the watermarks are intended to persist even when content is copied and pasted into other applications. However, the company noted that substantial editing, paraphrasing, or translation may render the watermark undetectable.
For supported file formats, including .svg, .png, and .jpg, Anthropic will use the Coalition for Content Provenance and Authenticity (C2PA) open standard to attach signed provenance metadata. This allows for the verification of a file's origin and detection of subsequent alterations.
The rollout applies to all Claude products and platforms globally, including the Claude API, Claude Code, Claude Cowork, and Claude Tag, as well as through cloud providers like AWS, Google Cloud, and Microsoft Foundry. While new models launched in the EU on or after August 2, 2026, will feature these capabilities from day one, Anthropic is working to update existing models during a designated transition period.
Entities
ACHNET · ACHNET Inc. · AWS · Anthropic · C2PA · Claude · Coalition for Content Provenance and Authenticity · EU AI Act · European Commission · European Union · European Union · German Federal Cabinet
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 4 SOURCES] Fines for non-compliance can reach up to 15 million euros or 3 percent of global turnover. jornalempresasenegocios.com.br · exame.com · www.b2bnn.com
- [○ 1 SOURCE] The Norwegian Tax Administration has detected AI-generated false attachments used in tax returns. www.ht.no
- [○ 1 SOURCE] The draft AI Migration Administration Act (KIMVG) would allow agencies to use personal data to train AI models for migration processing. verfassungsblog.de
- [● 2 SOURCES] AI systems used for recruitment and candidate evaluation are classified as high-risk. persoblogger.de
- [● 3 SOURCES] The compliance deadline for high-risk AI systems has been moved to December 2, 2027. persoblogger.de
- [● 9 SOURCES] Article 50 of the EU AI Act requires chatbots to disclose they are artificial intelligence unless obvious from context. jornalempresasenegocios.com.br · www.iww.de · exame.com · bitcoinethereumnews.com · blog.medientage.de · +1 more
- [○ 1 SOURCE] Kenya's ODPC issued new guidelines to address AI-related data protection challenges like algorithmic opacity and bias. www.businessdailyafrica.com
- [● 6 SOURCES] AI-generated or modified images, audio, and video that appear real must be clearly labeled as such. www.iww.de · jornalempresasenegocios.com.br · exame.com · blog.medientage.de
- [○ 1 SOURCE] ACHNET released compliance documentation for employers using high-risk AI in recruitment following the EU's deadline extension.
- [○ 1 SOURCE] The German Federal Cabinet approved a draft law allowing agencies to use personal data to train AI models for migration processing. verfassungsblog.de
- [○ 1 SOURCE] Bloomsbury Publishing and other entities will receive nearly $19 million from the settlement.
- [● 2 SOURCES] The court ruled that training language models on copyrighted material can constitute fair use under U.S. law. www.diariodealmeria.es