Anthropic has announced that its new models will include machine-readable watermarks in text generated by Claude to comply with European Union transparency regulations. This change is significant for communications teams that utilize Claude for tasks such as editing or formatting human-written press releases, as these documents will now carry an AI signature.
Starting from August 2, for models launched in the EU, Anthropic will implement two types of markings: - Text watermarks, which embed patterns into the generated text that are claimed to be imperceptible. - File metadata, which will include digital signatures confirming that the media files were processed by Claude.
However, Anthropic has noted two limitations regarding its detection technology: - Content that is AI-assisted may still appear AI-generated, even if Claude was only used for proofreading or formatting. - If the text is heavily rewritten, mixed with other content, or too short, the watermarks may not be detectable.
These changes are intended to meet Article 50 of the EU AI Act, which requires disclosures for AI-generated or manipulated content. Other AI providers, such as OpenAI, are also working on compliance measures, although their current focus is primarily on watermarking images and audio rather than text. Additionally, digital platforms are enhancing their AI identification efforts, with LinkedIn testing a detection feature and Substack integrating AI detection tools.
AI providers have until December 2 to ensure that legacy models comply with EU regulations, which will clarify how major players in the industry plan to mark AI-generated content.