Anthropic Unveils the Hidden Watermarking Technique in Claude’s AI-Generated Text

While Google and OpenAI have implemented SynthID watermarks for AI-generated images to verify authenticity, text remains a challenge. Recently, Anthropic introduced functionality within Claude to include watermarks in generated text, providing further explanation about the mechanism.

According to a blog post by Anthropic, Claude’s watermarking technique is based on SynthID-Text by Google DeepMind. The company emphasizes that the watermark is invisible to readers and does not significantly alter the content or quality of the text. There are no hidden characters involved, it doesn’t require extra tokens, and it cannot be traced back to any individual, organization, or chat.

The company describes how AI models typically generate one word at a time, selecting the next based on prior text. For instance, after the phrase “The weather today was cold and…”, a reasonable following word might be “overcast” rather than “sugary.”

Anthropic explains that this kind of word selection has little impact on readers, so it employs a random number generator for the word choice process:

Low-stakes choices are utilized throughout a piece of text to create a detectable pattern in Claude’s responses. This pattern remains unnoticed by readers but can be identified by someone possessing the appropriate key. Although watermarking introduces a unique source of randomness, it still employs random word selection.

Claude maintains neutrality regarding specific word choices, with the watermarking process not influencing the model to consider uncharacteristic words.

However, the company admits that this watermarking approach is not a foolproof method for identifying generated text:

With our key, detection will indicate “What is the probability this text was partially composed by Claude?” It does not ascertain if the text is human-written, nor can it determine if it belongs to another AI, as different AIs will have distinct watermarking methods and keys. Detection becomes more reliable with longer texts, due to the increased number of word variations.

This means that the system is tailored to work most effectively on longer texts. Additionally, factual passages experience diminished watermarking effectiveness since they entail fewer options for low-stakes word selections. For example, in the statement “Isaac Newton’s most famous work was titled Principia…”, the only accurate continuation would be “Mathematica.”

The same principle applies when users ask Claude to proofread their texts; watermarks will only be present in the adjustments made (like punctuation or grammar). Moreover, generated code will have limited watermarking opportunities due to its specific requirements.

Anthropic plans to introduce a watermark detection API shortly, allowing verification of whether text was generated by Claude. Ultimately, there is hope that major AI platforms like Gemini and ChatGPT will adopt similar text watermarking techniques, akin to the utility of SynthID in AI image detection.