Anthropic revealed that it will start embedding an imperceptible watermark in the output of its Claude language model from 2 August. The hidden signal travels with copied or pasted text and survives modest editing, allowing downstream tools to detect that the content originated from the AI system.
How the watermark works
The watermark is a machine-readable pattern woven directly into the model's generation process. It does not alter the readability or quality of the text, and it is invisible to human eyes. When users interact with Claude via the chatbot, the API, or specialised tools such as Claude Code, the signal is attached automatically. Images produced by Claude will also carry a similar marker that indicates the file was processed by the model and flags any subsequent tampering.
Alignment with the EU AI Act
The timing coincides with the enforcement of the European Union's AI Act, which entered into force on 2 August. The legislation, overseen by the European Commission, requires generative-AI providers to make synthetic output detectable and machine-readable. By embedding a watermark at the model level, Anthropic aims to meet these new transparency obligations across the bloc.
Why the move matters
Detecting AI-generated text has proved difficult because written content is frequently copied, paraphrased, translated or merged with human-written material. Previous watermarking attempts have focused mainly on images, leaving text largely unmarked. Anthropic's approach does not claim that a whole document was produced by Claude; it merely signals that the model contributed to the text, even if the user only asked for a proofread or a translation. This could help platforms, publishers and regulators differentiate between low-quality AI spam and legitimate human-AI collaboration.
Industry reaction and next steps
Other tech firms are also tackling the surge of AI-generated "slop" on social media. Platforms such as Substack and YouTube have introduced tools and policies to flag or demonetise mass-produced AI content. However, experts warn that a watermark can be stripped away by heavy rewriting or translation, so it is not a panacea. Anthropic says it will extend the marking to older Claude models in future releases.
For now, the hidden watermark offers a practical way for developers of detection tools to comply with EU law and for users to verify the provenance of text they encounter online. As the AI ecosystem evolves, the balance between seamless AI assistance and transparent attribution will likely shape future regulatory and industry standards.