How Claude’s Text Watermark Works An upcoming update to Claude’s text generation process will introduce a watermarking system designed to help identify content generated by the model. This change aligns with regulatory requirements under the EU AI Act and reflects broader industry efforts to address transparency and accountability in AI-generated text. The watermarking method, based on techniques developed by Google DeepMind, operates by subtly altering the randomness used to select words during text generation. While the final output remains indistinguishable to readers, the watermark allows for probabilistic detection of Claude’s involvement through specific patterns embedded in the text. The watermarking process leverages low-stakes word choices that occur naturally during text generation. For example, when a model like Claude selects a word such as “overcast” or “grey” to complete a sentence like “The weather today was cold and…,” the choice is typically influenced by randomness. Watermarking introduces a different source of randomness, using a cryptographic key and preceding words to determine the next token. This creates a detectable pattern for those with access to the key, without altering the model’s output quality or readability. The randomness remains inherent, but the method ensures that certain sequences of words can be analyzed to assess the likelihood of Claude’s involvement. Key aspects of the watermarking approach include its non-intrusive nature and its focus on probabilistic detection rather than definitive identification. Internal testing and controlled studies have shown no measurable impact on the quality, creativity, or readability of Claude’s outputs.#claude #google_deepmind #eu_ai_act #synthid_text #nature_paper
