Can You Spot Where the AI Left a Watermark?
This post was written by a human.
As we discussed a few days ago, Claude now watermarks its text output. It’s not hidden characters or the em dash, but the choice between words that mean the same thing.
This a game teaches you to see the slots where that choice happens.
The game is: find the word the model was free to swap.
Five sentences, ten seconds each, faster is worth more.
Spot the Slot
In each sentence, one word or phrase could be swapped for something that means exactly the same thing. The rest carries facts, names or numbers that can’t move without changing the meaning.
Click the swappable one. Ten seconds. Five rounds. It gets harder.
Click the word the model could have swapped without changing the meaning.
0
Here’s the catch
You just found the slots. What you still cannot do, and neither can I, is read which option the watermark actually favoured. That takes a key only the model’s maker holds.
So a watermark can tell them a machine was involved. It cannot tell anyone whether a human read the thing, checked the facts, or took responsibility for publishing it. Under Article 50(4), that second part is the one carrying your legal duty.
A watermark proves a machine was involved. It says nothing about whether a human took responsibility.
That gap is the whole reason Art50 exists. The mark on the text, if it survives editing at all, tells an auditor one thing. A signed, timestamped record of who reviewed the piece and who stands behind it tells them the thing they actually asked about.
Seal the human review. Free until 2 December 2026.
