Anthropic is embedding "imperceptible" machine-readable watermarks directly into text generated by new Claude models, making copied AI output easier to trace as schools, publishers and other organizations grapple with undisclosed machine-written content.
Claude Watermarks Follow Text Across Platforms
IPO-bound Anthropic on Monday said Claude models launched on or after Aug. 2 will support marking from launch. The watermark "doesn’t change the meaning, quality, or readability" of Claude’s response, travels when users copy and paste text and "may persist through some editing."
The marking applies worldwide across supported models used through Claude, Claude Code, Claude Cowork, Claude Tag and Anthropic’s API, as well as through AWS, Google Cloud and Microsoft Foundry. Anthropic is also working to retrofit older models and plans detection tools for third parties.
The move follows Anthropic’s commitment to the European Union AI Act’s transparency framework. EU rules require generative-AI providers to make synthetic output identifiable in a machine-readable format, aiming to give users clearer signals about content provenance.
Schools And Publishers Gain New Detection Tool
The technology could matter particularly in education. Anthropic recently launched Claude for Teachers, expanding the chatbot deeper into classrooms as educators debate responsible AI use. Watermarking could give schools another signal when investigating whether students submitted generated work as their own.
Publishing faces similar pressures over AI authorship and training data. Anthropic recently secured final approval for a $1.5 billion copyright settlement with authors, highlighting how generative AI continues to collide with questions of authorship, ownership and disclosure.
Detection Limits Leave Room for Workarounds
That said, the watermark is not foolproof. Anthropic says heavy editing, paraphrasing, translation or mixing Claude output with other writing can destroy the detectable signal. Very short passages may also provide too little material for reliable detection.
A detected mark also does not prove Claude originally wrote the material. Using Claude merely to proofread, translate or summarize human-authored text can leave a mark, Anthropic said.
Anthropic is not alone. Alphabet Inc. (NASDAQ:GOOG)(NASDAQ:GOOGL) subsidiary Google DeepMind already embeds its SynthID watermark into Gemini-generated text by subtly adjusting token probabilities without visibly changing output quality.
Image via Shutterstock
Login to comment