Anthropic Explains New Watermarking for Claude and Its Limits

Anthropic has put a hidden mark on text written by Claude, its helpful robot assistant.

This mark is like an invisible fingerprint that shows the text was made by a machine.

The mark is not added when Claude writes computer code.

Code has to be exactly right, so changing even one word could break it.

Adding a mark there might make the code stop working.

However, the mark can appear in comments inside code.

Comments are little notes for humans that do not change how the program runs.

The words in comments can be chosen more freely, so the mark fits there safely.