Anthropic is adding an invisible watermark to all text created by future Claude models. If your team uses AI for publishing, education, or regulated records, this change affects how you track and prove AI usage under the European Union AI Act.
On August 14, 2026, the company published a technical guide in Anthropic — How Claude’s text watermark works. Anthropic is rolling out this feature worldwide because it cannot easily limit it by region yet.
How the hidden pattern works
Anthropic built its system using Google DeepMind’s SynthID-Text method. The signal is completely invisible. It does not add special letters, extra tokens, or visible tags to your output.
Instead, the model slightly adjusts the math it uses to pick words when several options make equal sense. A checker tool uses a secret key to scan these subtle choices. It then calculates the probability that Claude helped write the text.
Here is what you get:
– No extra cost. Anthropic says watermarking adds no extra tokens and will not raise serving prices.
– No personal data. The watermark contains no information about user accounts, organizations, or chats.
– Detection API coming. Anthropic plans to launch a detector API soon, though final details are still pending.
What the vendor claims — and what nobody verified yet
Anthropic says internal tests showed no real drop in writing quality, citing a separate study on SynthID-Text. Keep in mind: these claims come directly from Anthropic’s internal tests and cited research. Nobody outside the company has checked every Claude model, language, or temperature setting yet.
Where the detection breaks down
A positive check only means Claude was likely involved at some point. It cannot prove human authorship. It also cannot tell Claude apart from another AI, or show whether Claude drafted or edited the text.
Detection gets much weaker on:
– Short text snippets
– Coding tasks
– Strict factual writing with few alternative word choices
– Light proofreading and full rewrites
Also, text watermarking is completely separate from Anthropic’s C2PA metadata for image files.
What you should do next
Do not rely on scores alone. Never treat this watermark as a single proof of plagiarism or policy violations.
When Anthropic releases its detector API, test it on your own workflows. Generate long samples with Claude, edit them, translate them, and see where the signal breaks. Until then, keep clear internal records of prompts, source files, and human reviews.