Anthropic explains how Claude’s invisible text watermarks will work
Key Points:
- The sentence "The weather today was cold and..." is unlikely to be followed by an unrelated word like "sugary," but likely to be followed by words like "overcast" or "grey," which convey similar meanings.
- In such cases, the choice between similar words is typically random and does not significantly affect the sentence's meaning for the reader.
- Watermarking leverages these low-stakes word choices to embed a hidden pattern in AI-generated text, which is imperceptible to readers but detectable by those with a specific key.
- Instead of using a standard random number generator, watermarking uses a key combined with preceding words to determine the next word choice, ensuring the pattern is consistently encoded.