August 18, 2026, (Inside AI) — Anthropic has revealed the technical foundation of its new text watermarking method for Claude, confirming it is based on Google DeepMind's SynthID-Text research and does not hide invisible Unicode characters. The company says the watermark is imperceptible and has no practical impact on output quality.
In a blog post published on August 14, Anthropic detailed how the watermark will work for text longer than 200 tokens, roughly 150 words. The watermark survives copying, pasting, and even some editing. The company also plans to attach cryptographically signed provenance data to AI-generated images.
Anthropic's move is not voluntary. It responds to Article 50(2) of the EU AI Act, which took effect on August 2, 2026. The law mandates transparency for AI-generated content. Anthropic is part of a group of providers, including OpenAI, Google, Meta, and Microsoft, that signed the EU's Code of Practice on transparency.
"We use a method of watermarking that does not have any practical impact on the quality or content of Claude's outputs [...] Watermarking doesn't change the meaning or experience for the person reading it, but if you wanted to check after the fact whether the text was likely generated by Claude, the watermark allows you to do so," Anthropic, blog post.
The company said it is working on a watermark detection API for launch in the coming months. Existing Claude models will also receive watermarks soon.
Token choices leave a detectable statistical fingerprint #
Large language models generate text by selecting tokens from a probability distribution. Each token choice is semi-random. SynthID-Text assigns a score to each token based on previous tokens. The sampling strategy might pick the token with the highest SynthID score among the top five most likely options.
Anthropic's version adds a cryptographic key. The key influences token selection in a way that is invisible to readers but detectable later. Anyone with the key can check whether token choices consistently align with what Claude would have chosen using that key.
For example, in the sentence "The weather today was cold and...", Claude might choose between 'overcast' and 'grey'. If the key favors 'overcast', and hundreds of such choices align, the watermark is detected. The company says the model choice remains random and non-biased. Anthropic cited internal testing showing no impact on content, creativity, or readability. It also referenced a Google DeepMind study where Gemini users gave thumbs-up or thumbs-down ratings. The study found no statistically significant differences from the unwatermarked model. A controlled test showed human raters saw no quality difference side-by-side.
Critics question the cost of constrained word choice #
Not everyone is convinced. John Gruber of Daring Fireball called the technique a "perverse adulteration" of writing. He argued that even small shifts in word choice matter.
"I want any LLM I use to choose the very best, most precise words at every single decision point," John Gruber, Daring Fireball.
Anthropic's watermark has real limits. It cannot distinguish "Claude wrote this" from "Claude heavily edited this." It does not confirm if text was human-written or generated by another AI model. The key-centric technique only applies to Anthropic's own models.
Watermarking is less reliable on factual passages where few alternative token choices exist without hurting accuracy. It also struggles with proofreading and AI-generated code, where exact output is required and there is little room for watermark-friendly choices.
Translated text using Claude will also carry a watermark. The company denied hiding invisible non-printing Unicode characters, a method some reports initially speculated about. It claimed negligible impact on model speed and token consumption.
Anthropic said watermarks will not contain personally identifiable information linking text to a user, organization, or chat session. The detection API is expected in the coming months.