Anthropic shares more details about how Claude’s new watermarks will work

Anthropic Details Claude's Watermarking Plan

Anthropic has released specific details about its plan to watermark text generated by its Claude chatbot, a move intended to align with transparency requirements in regulations like the EU AI Act’s Transparency Code. The announcement follows a mixed reaction from users, some of whom have reportedly canceled subscriptions in protest, highlighting the growing tension between AI developers’ compliance obligations and user expectations for unrestricted content generation.

The Technical Implementation

The company confirmed it will use the SynthID-Text approach developed by Google DeepMind. This technique embeds a statistical pattern into the text during generation by making specific, low-stakes word choices—such as picking “grey” instead of “overcast”—that are imperceptible to a human reader but can be identified by a corresponding detection tool. Anthropic asserts this process does not degrade the quality of Claude's output and plans to release a watermark detection API for verification.

• Editing Resistance: The watermark is designed to persist through light editing, though a complete rewrite will remove it.
• Code Impact: Functional code will have a minimal watermark, as the model has less freedom in choosing syntax. Watermarks may still appear in more arbitrary sections, like code comments.
• Partial Generation: For text only partially edited by Claude, the watermark's detectability will depend on the extent of the AI's contributions.

A New Industry Standard

This initiative is not unique to Anthropic. The company noted that other major model developers have also signed the same Code of Practice, signaling an industry-wide shift toward embedding provenance directly into AI-generated content. This method represents a move away from less reliable, pattern-based AI detectors and toward a more robust, cryptographic-style approach to content verification. As regulatory pressures mount, such built-in traceability systems are likely to become a standard feature across all major foundation models.

Anthropic's adoption of SynthID-Text isn't just a compliance checkbox; it's a strategic move to establish a technical standard for AI content provenance. As regulators demand more transparency, the race is on for model providers to implement robust, difficult-to-spoof watermarking that can become the industry norm, moving beyond fallible stylistic detectors.