Play
Be the Watermark
Claude now hides a watermark in the words it picks. According to Anthropic, it’s a version of SynthID-Text, a method Google DeepMind published in 2024. The model doesn’t change. What changes is how the software running it chooses between near-equal words.
Here’s the idea. For each word, a secret key and the few words before it decide a coin flip. When the AI has a real choice between words, it picks one whose coin comes up heads. Ordinary writing comes up heads about half the time. Watermarked writing comes up heads more often, and a detector with the key can count them.
In this game, you do the AI’s job. Swap words in the story until enough coins come up heads, then run the detector.
On a sunny day, the tabby rested by the creek. A duck glided over the water. The dog ran to the beach and barked. The toad jumped onto a rock. The bird sang in the tree.
Press Step to check one word, or Run to check the whole story. Words with a dashed box can be swapped.
How to play
Open a word’s menu to see whether each choice comes up heads or tails in that spot with the current key. Swap some tails words for heads words and run the check again. When at least 60% of the 32 coin flips come up heads (20 of 32), the detector shows “Watermark detected.”
- Work left to right. Each coin flip uses a word and the four words before it, so a swap can change the next four flips too. It never changes the flips before it.
- Watch the small words. A swap can cost you heads on a word without a menu, like “the” or “water.” The only fix is to change a word before it.
- Try the wrong key. Once you get the badge, change the last key digit to 6 and run again. Your heads count will probably fall back toward half, even though your watermark is still there. The detector needs the right key.
How the coin flip works
Every word gets a token ID: its position, counting from 0, in an alphabetical list of the 75 words on this page. To flip a word’s coin, multiply the IDs of the four words before it and the word itself by the five key digits, then add them up. If the sum ends in 5 through 9, it’s heads. If it ends in 0 through 4, it’s tails. The first four words don’t have enough words before them, so the check starts at word five.
Real watermarks are more involved. Token IDs aren’t alphabetical, the key is much bigger, and the words are mixed with a hash function instead of a sum. And a story this short proves little: 32 fair coin flips reach 60% heads about one time in nine. Over hundreds of words, that kind of streak becomes very unlikely by chance.
Why this matters for your business
Right now, only the AI companies and a few organizations they’ve approved can check text for these watermarks. If you need to know where content came from, provenance has to be built into your own systems. That’s what Kilroy AI does: our founder co-chairs the C2PA AI/ML Task Force, and Free2PA brings C2PA-style content credentials to AI agents.
Request a free AI readiness call
This game is adapted from “Be the Watermark” by Andrew Stellman, from his O’Reilly Radar article “Claude Watermarks Your Words, and You Can’t Check Them (Yet)” and his upcoming O’Reilly book Be the Model.