Anthropic Adds Watermarks to Claude Content Amid Deepfake Concerns
AI company Anthropic will implement machine-readable watermarks across its Claude products to differentiate AI-generated content from human work, addressing the growing issue of deepfakes.
Artificial intelligence company Anthropic has announced it will make machine-readable watermarks a default feature across all its Claude products worldwide. This move addresses rising concerns over the proliferation of deepfakes and other AI-generated content, which complicates the distinction between authentic and synthetic information.
The new policy will apply to Claude models launched in the European Union on or after August 2, 2026, with older models gaining support during a transition period. The watermarks are hidden markers readable by machines but not directly by human eyes, designed to persist with the content regardless of its subsequent use. For audio, images, and other file types, Anthropic plans to use signed metadata, in addition to embedded text watermarks, adhering to the C2PA standard.
Competitors like Google have already introduced similar technologies. Google's SynthID system watermarks text, images, and audio generated by its Gemini model. OpenAI previously developed a watermarking tool but did not deploy it in ChatGPT, citing concerns about potential false accusations and user backlash. Anthropic acknowledges that a watermark is not conclusive proof of AI generation, as content can be edited afterward.
According to Amitabh Kumar, CEO of Contrails AI, the implementation of watermarks is a minimum requirement but he highlights that the current approach has gaps. Unclear details remain regarding watermark persistence in digital copies and the availability of detection tools for third parties. In India, deepfakes have already caused significant harm, including fabricated videos of political figures during elections and an increase in financial fraud.
It remains uncertain if Anthropic's solution will avoid the challenges faced by its competitors. The effectiveness of watermarks depends on open detection tools, interoperability between standards, and whether companies treat marking as more than just a compliance checkbox. Anthropic has pledged to share technical guidance on its approach, and the industry will be watching how quickly these promises translate into functional markings.