Anthropic clarified Friday how it will embed invisible watermarks in text produced by its Claude model, a step aimed at complying with the European Union’s AI Act. The company described the new system as a "version of the SynthID‑Text approach," an open‑source watermarking technique originally created by Google DeepMind.
The EU regulation requires synthetic audio, image, video and text to carry machine‑readable marks that allow detectors to flag content as artificially generated. Anthropic’s answer is to embed a subtle pattern in the word‑choice process of Claude’s output. The pattern is undetectable to readers but can be identified by anyone with the appropriate detection key.
In practice, the watermark relies on low‑stakes lexical decisions that occur throughout a paragraph. For example, after the phrase "The weather today was cold and…" a model might randomly choose between "overcast" or "grey," both reasonable continuations. When watermarking is active, the random selection draws from a deterministic source that incorporates a secret key and the preceding words, leaving a traceable signature across the text.
Anthropic emphasizes that the watermark will not increase Claude’s pricing or degrade the quality of its responses. The company assures users that the changes remain transparent to the end‑user while providing regulators with a reliable detection method.
The move mirrors steps taken by other AI developers. Google has employed SynthID‑Text for its Gemini chatbot since 2024, and OpenAI, though silent on specific plans, faces the same compliance deadline under the AI Act. Anthropic also announced support for C2PA standards in Claude‑processed images, extending its transparency measures beyond text.
Industry observers note that the EU’s rules could reshape how AI products are built and deployed worldwide. By adopting an open‑source watermarking framework, Anthropic positions itself to meet regulatory demands without reinventing the technology from scratch. The approach also signals a broader industry shift toward embedding provenance signals directly into generative models.
Anthropic’s explanation underscores a growing consensus: compliance can be baked into AI systems without sacrificing performance. As the AI Act’s enforcement timeline approaches, more developers are likely to follow suit, integrating similar invisible markers to assure regulators and users alike that synthetic content can be reliably identified.
Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.