🖍️ Watermarking described with crayons.
Wait, watermarking? Claude's newest models watermark (aka leave a fingerprint behind that can be detected after the fact) the text they output. Opus 5, 5.5, and Fable 5.1 all do, with older ones coming later as Anthropic works to meet the standards required by the EU AI Act.
Nothing gets added to the output, though: no hidden characters, no extra tokens, and nothing that points back to you. The pattern is in which word Claude picks, but the trick is to do that without changing the odds of which word Claude picks! 🤯
Picture a class full of kids coloring a park scene. The pictures all look basically the same, but each chose the colors they used themselves. Except for one kid! Instead of picking colors based on vibe, she had a rule. She counts the letters in the name of whatever she is coloring. An odd number of letters means pick a darker color; an even number means pick a lighter one. So, "Road" might become a light gray. Of course, if only one color suits her needs, she picks that one even if it doesn't follow the rule.
Her resulting picture looks like everyone else's because both crayons she chose already looked right. But a teacher who knows the secret rule could go through each picture and check. All the other kids' drawings would match the rule about half the time by pure luck (on average). But this kid... her drawing would hit way more often, and the more spots the teacher checks in the image, the more sure they get.
The colors are like the words the model already likes; you (probably) won't get a green sky. At least you won't get a green sky any more often than any other kid 😉. The letter-count rule is like a secret function that only Anthropic knows.
I really wanted to understand this, so I built a watermarking scheme on my laptop to see it happen! My writing scores about 0.5, but a watermarked paragraph scored 0.668. It might sound like a small difference, but statistically, it is a huge difference, over 8 standard deviations away from human-authored text.
Thanks to Tim Rawson for helping put these together every week. The coloring example was, to my surprise, Jaws' idea. I rewrote it a lot, Claude drafted the rest, and then I rewrote that too. There might be a watermark in this post, but it is minimal.