It’s super easy to remove the metadata from the image, but just having it in there makes Anthropic comply with the legal requirements. They're not an image generation platform but the future of trusted visuals online is really going to be a whitelist approach of "assume false unless proven true by cryptographically sealed metadata" rather than using metadata to show that it's AI. This text example is directionally accurate, making it easy to understand how the watermark is really just a manipulation of which still-random set of flexible outputs get chosen, but it’s not quite accurate to how the actual system works. In reality, it needs to be able to evaluate a block of text without knowing the prompt or a specific pattern that will be followed, so the text carries its own seeds. The "random" number at each step is a hash of the secret key plus the last four words just written: seed = hash(key + previous 4 tokens) The seed feeds a pseudorandom function g(word, seed) that assigns a secret score (effectively a thumbs-up or thumbs-down) to *every word in the entire vocabulary* at that step. The generation step then takes the candidate words through a knockout tournament where thumbs-up words tend to win (the paper uses 30 layers of this). The detector then asks a question that requires no knowledge of the alternatives: take each word that actually appears, recompute its g-score from the key and the preceding four words, and check: is this text suspiciously full of thumbs-up words? Human-written or unwatermarked text should average 50% thumbs-ups; watermarked text should run consistently above that. The final “Planted Probability, Is This AI Score” is the mean across the whole text. But any time that there’s one word that *really works better* as an output in a given slot, the watermarking can’t really work there without forcing a lower-quality output, hence the type and length of output text significantly impacting things. #Science #ai #technology #claude — @distilledscience TikTok