Anthropic Fights Deepfakes: New Watermarking Tool to Authenticate AI Text

Anthropic, a leading developer of artificial intelligence models, has officially announced its decision to implement watermarking on text generated by its systems, including the Claude suite of models. This strategic move is undertaken to ensure compliance with stringent European regulations, specifically the EU AI Act's Transparency Code, which officially came into effect on August 2. The company publicly confirmed this policy update through a revised support page, highlighting a broader industry trend towards greater accountability and clarity in distinguishing AI-produced content.

As per Anthropic's updated directives, all AI models released subsequent to August 2 will be equipped with integrated technology designed to apply watermarks to both computer-generated text and digital files. For file-based outputs, Anthropic is leveraging the C2PA open standard, a recognized framework dedicated to enhancing content authenticity and traceability. This adoption of a standardized approach is aimed at promoting interoperability and facilitating the reliable identification of AI-generated materials across diverse platforms and systems.

A critical characteristic of Anthropic's watermarking solution is its inherent persistence and pervasive application. The company has clarified that the watermark is embedded directly within the textual content, ensuring its continued presence even when users copy and paste the text into different environments. Furthermore, the watermark is engineered to withstand certain levels of editing, though the precise extent of modification required to remove it has not been detailed. This model-level implementation guarantees that the watermark will be present regardless of the specific Claude product or interface from which the text originates, encompassing offerings like the Claude platform API, Claude, Claude Code, Claude Cowork, and Claude Tag.

Anthropic's commitment to watermarking reflects a wider initiative across the AI industry to mitigate user concerns and avert potential regulatory scrutiny stemming from the increasing volume of AI-generated content. Numerous other prominent entities within the artificial intelligence and digital content sectors are similarly engaged in efforts to mark AI-produced materials. For example, Suno, an AI music platform, recently declared its intention to watermark tracks created on its platform in the wake of several legal challenges. Concurrently, the newsletter service Substack has forged a partnership with Pangram to identify AI-generated content, with Substack's CEO, Chris Best, having previously coined the term "Claudefishing" to describe the practice of individuals using AI to create deceptive content. Beyond Anthropic, a collective of influential companies, including Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia, has also publicly pledged its adherence to the transparency provisions stipulated by the EU's code.