Anthropic adopts Google's SynthID-Text to watermark Claude outputs for EU AI Act compliance

Anthropic has revealed how it plans to watermark text generated by its Claude models, adopting a version of Google DeepMind's open-source SynthID-Text approach. The move is designed to comply with the European Union's AI Act, which requires synthetic content—including text—to carry machine-readable marks that identify it as AI-generated. The company announced the mechanism on Friday, alongside C2PA support for Claude-processed images.

Watermarking works by subtly influencing the model's word choices in low-stakes situations. As Anthropic explains in its blog post, when a model chooses between equally plausible words—like "overcast" or "grey" after "The weather today was cold and…"—the selection is normally made by a random number generator. With SynthID-Text, that randomness is instead derived from a cryptographic key and preceding words, leaving an invisible pattern that can be detected by anyone holding the key. The reader sees no difference, and the semantic content remains unchanged.

Anthropic stresses that the watermarking will not increase costs for users or degrade output quality. "It doesn't matter much to the reader which of these latter two words the model ultimately chooses," the company notes, underscoring the low-impact nature of the mechanism.

This places Anthropic alongside other major AI developers responding to the EU's transparency mandates. Google's Gemini has supported SynthID-Text since 2024, and while OpenAI has not yet detailed text watermarking plans for ChatGPT in its AI Act compliance roadmap, it too will be subject to the requirements.

For developers building on Claude, the practical implications are minimal: API integrations will not need to change, and outputs will remain semantically equivalent. The primary effect is auditability—organizations with the appropriate key can verify whether a given text originated from Claude, which could be useful for content provenance and trust workflows.

As the AI Act continues to reshape the landscape, text watermarking is likely to become a standard feature across major models. For now, Anthropic's adoption of SynthID-Text signals a pragmatic, interoperable approach—one that aligns with an existing open standard rather than a proprietary system.