Anthropic has shared fresh detail on how it plans to stamp invisible watermarks into Claude’s text output. The change is meant to satisfy Europe’s AI transparency rules. It lands as the EU AI Act pushes major developers to label synthetic content.
The company says its text marking system is “a version of the SynthID-Text approach.” Google DeepMind built that open-source tool. It creates detectable patterns by nudging how the model chooses its words.
The EU AI Act requires machine-readable marks on synthetic output. That covers audio, images, video, and text alike. The marks must let people spot content that was generated or manipulated by AI.
Anthropic offered a plain example to explain the mechanism. Take the sentence “The weather today was cold and…” The next word is unlikely to be “sugary.” But “overcast” or “grey” both fit fine. Readers rarely care which one lands. A random number normally settles that choice.
Watermarking reuses those low-stakes decisions. Each one leaves a small trace across a full response. The pattern stays invisible to readers. Anyone holding the right key can detect it later.
Anthropic says the feature will not push up prices. It also will not change the quality or content of Claude’s answers. The update arrives alongside C2PA support for Claude-edited images. Together, the two moves help Anthropic meet its EU AI Act duties.
Google’s Gemini chatbot has supported SynthID-Text since 2024. OpenAI has not detailed text watermarking plans for ChatGPT yet, but the same law applies to it as well.
The wider goal is to curb AI-generated spam and misinformation. Watermarks give platforms a way to flag machine-written text without wrecking the reading experience. That push matters more as AI-generated research papers flood peer review. Anthropic remains a central player in that debate, as recent public clashes over AI safety have shown.
Source: The Verge
