Claude’s watermark “mystery” is solved

When Anthropic announced “invisible” watermarking for Claude, it left one fairly important detail unanswered:

What technology would actually create the watermark?

Now we know. They confirmed Claude’s text watermark will use a version of SynthID-Text, the method published by Google DeepMind in 2024.

It does not add hidden characters or metadata to the writing. Instead, it creates a detectable pattern through the low-stakes word choices Claude makes while generating text.

Anthropic also confirmed all the weaknesses.

Short passages may not provide enough signal. Factual writing and light proofreading leave fewer opportunities to add the watermark. A complete rewrite can remove it.

And even when detected, the watermark only suggests that Claude was involved.

It does not mean Claude did the thinking, the research, found the angle and wrote the whole thing.

Back to arguing about em dashes.

If you’d like to put AI to useful work and create more leverage, without the drama, let’s talk.