Anthropic deploys invisible watermarking for Claude to meet EU AI Act rules
The AI company is implementing a statistical signature in token selection to identify machine-generated text globally.
Anthropic is implementing an invisible, machine-readable watermark for all text generated by its Claude AI to comply with the European Union's AI Act. The move signals a shift toward systemic content provenance as regulators demand clearer distinctions between human and synthetic media.
The system works by modifying the randomness of token selection during the generation process. Rather than adding hidden characters or metadata, the AI creates a statistical signature within the text itself. According to Anthropic, the process uses "low-stakes choices" that occur repeatedly across a response to leave a detectable pattern without altering the appearance, quality, or speed of the output.
Technical Foundation
The implementation is based on Google DeepMind's SynthID-Text approach, which injects a signature directly into the token sampling procedure. Because the watermark is embedded in the probability of word choice rather than added as an external tag, it remains invisible to human readers. Anthropic has confirmed that the watermark will be applied globally to all Claude-generated text, including outputs delivered via its API.
Regulatory Pressure
The primary driver for this deployment is the EU AI Act, which mandates that AI providers mark generated content to ensure transparency and easier identification. While the legal requirement is centered on the European market, Anthropic is deploying the feature worldwide. This follows a broader industry trend where major providers, including Google and OpenAI, are exploring similar methods to track the origin of synthetic content.
Industry Implications
This transition from relying on detectable linguistic patterns—such as repetitive phrasing—to using cryptographic-style signatures represents a significant escalation in the effort to combat AI-generated misinformation and plagiarism. By embedding a machine-readable identity into the text, AI companies aim to make it more difficult for users to pass off synthetic work as human-authored.
The Path Forward
Despite the technical sophistication of the SynthID-Text approach, the long-term effectiveness of text watermarking remains a subject of industry debate. Experts note that while the statistical signature is robust, complete rewrites of the generated text by a human or another AI can still strip the watermark. Observers will be watching to see if this method provides a durable solution for provenance or if it serves primarily as a regulatory checkbox.