Creators Discuss Claims of Claude Watermarks From 2026
Posts examine an assertion that Anthropic will watermark future Claude outputs.
Alex Volkov noted EU AI Act rules on LLM watermarking and said Anthropic began complying on August 2 while OpenAI has not. Andrew Curran described Google's token-bias method in Gemini since 2024. Multiple users shared a claim that Anthropic will embed invisible watermarks in Claude models launched on or after August 2 2026 that survive copy and paste. Emad Mostaque and others retweeted explanations of the KGW technique used by labs. Daniel Jeffries raised concerns about edits to legal documents. Pliny the Liberator and Peter Steinberger flagged the same post without added details. Elie Bakouch asked whether the method could detect training on Claude outputs.
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…
Reactions from ranked influencers
103 postsGemini has also been doing this since 2024. Google uses a secret key to bias the model toward certain tokens, creating a detectable statistical signature - an invisible pattern embedded in all text generated by Gemini. It is unknown whether OpenAI does something similar.
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…
the watermark will basically be the statistical signature/word choice of the models Ant claims this won't affect quality but this seems non-obvious given that they'll be constraining the model's sampling & biasing token choice and potentially making the model less…
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…
🚨 JUST IN: Claude models will now have invisible watermarks embedded in ALL text, and ALL metadata attached to files…
Hey, this might be the first positive thing out of EU AI regulations: forced Ant to put fingerprints in Claude models. I'm actually mildly shocked they didn't do that yet, given all their fearmongering??
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…
Startup idea: Rephrase output to preserve the meaning while obliterating the watermarks. This might not become that big in itself, but it could make a great beachhead.
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…
Defeated with cheap rewriting, definitely defeated if you train a rewriter against a reconstructed detector. Amusing GAN project actually But at least will help Pangram to deal with the low effort slop
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…