Anthropic says new Claude models will embed invisible watermarks in all generated text, everywhere Claude is offered.
The watermark is part of the text, it isn't metadata: "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing."
This starts with models launched on or after August 2, 2026, under an EU AI Act code Anthropic signed. Anthropic is still working on adding it to current models. The rollout is worldwide.
Date: August 10th, 2026 4:24 PM Author: gunneratttt (gunneratttt)
Seems like the lib instinct to "do something" even if it's just performative. All people need to do is figure out how the watermark is embedded into the text and then change a couple words. Meanwhile, over time these "can't hurt, why not?" type laws increase compliance costs makes everything more expensive and harder to compete internationally.
Date: August 10th, 2026 4:27 PM Author: federal volcanic crater
You know how it says in this article it doesn't rely on metadata? It's because the way that it works is they put in an algorithmic bias toward selecting certain tokens at certain spots that supposedly don't change the meaning of the output, and that makes it traceable as coming from that model. But if you rewrite parts of it or put it through a translator and back it destroys it. Another thing is they were worried it would have a DITHPARTE IMPACT on ESL speakers who can't desloppify AI outputs well. I'm sure they also considered that people won't want to use it anymore if they know its easily traced from claude.