Anthropic explained on 14 August 2026 how it watermarks text written by Claude, and updated the post on 1 September with details of who can use its detection service. The short version: the watermark is a statistical signal that Claude was probably involved, not a universal authorship test, and the detector is not something the public can run.
Anthropic says the watermark works by nudging Claude’s choice between equally good words. Nothing is added to the text, there are no hidden characters, and the watermark carries no information that could trace a passage to a person, organisation or chat. Detection estimates the likelihood that Claude was involved, and works less well on short passages, factual text, code and text Claude only lightly edited.
The detection API is in private preview for eligible organisations under EU law, such as regulators, media, fact-checkers and researchers, and for enterprises with their own compliance obligations. Anthropic is watermarking to comply with the EU AI Act and applies it globally.
How the Watermark Works
Language models write one word at a time, choosing among candidates. Anthropic says its method takes advantage of the moments where two or more choices would be equally good and nudges the selection in a way a detector with the key can later recognise. Where there is only one right answer, such as completing a sum or naming a book, there is nothing to nudge, so the watermark is sparser on factual passages and on code. The company says the effect is not distinguishable to readers, adds no tokens and does not change the price.
Source: Anthropic: How Claude’s text watermark works, read 10 September 2026.
What a Result Means
Four Questions, and What Anthropic Says
Was Claude involved? Detection estimates the likelihood.
- Confidence rises with the length of the passage
Did Claude write all of it? The watermark cannot say.
- Lightly edited human text may carry too little watermark to detect
Does no detected watermark prove a human wrote it? No.
- Short, factual, edited or code-heavy text may be undetectable by Anthropic’s own description; earlier models did not carry the watermark
Will a short passage give a reliable answer? Not reliably.
- Anthropic says detection does not work well on small samples
Two limits matter for anyone reading a detection result. Anthropic says the watermark attaches only to words Claude chooses, so when Claude proofreads a person’s writing the changes may not be enough to make its involvement detectable. And the post says future Claude models will carry the watermark, so text from earlier models should not be assumed to carry it.
Source: Anthropic: How Claude’s text watermark works, read 10 September 2026.
Who Can Run the Detector?
Anthropic says the detection API is in private preview. It is available to organisations that need it under EU law, which the post lists as regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, and to enterprises with similar compliance obligations. Anthropic says it plans to expand access over time. That is different from a public checker anyone can assume they have.
Source: Anthropic: How Claude’s text watermark works, read 10 September 2026.
Why It Matters for Writers and Editors
If someone sends you a detector result, ask which tool produced it and for the exact passage submitted. Keep that beside the dated draft and source notes. It gives you something concrete to discuss instead of arguing over whether a paragraph sounds like AI. A detection result still does not establish whether the claims in it are true.
Before Acting on a Detection Score
Even knowing that a tool was involved would not establish whether a passage is correct, fairly attributed or suitable to publish. A practical review still has to check the claims.
Sources and How This Was Checked
Every technical description is attributed to Anthropic’s post dated 14 August 2026 and updated 1 September 2026, read on 10 September 2026. Tech X Telco did not test the detector or measure its error rates, and the inference that a missing watermark does not prove human authorship is drawn from the limitations Anthropic itself describes.
Related on Tech X Telco: Claude Fable 5.1 and Mythos 5.1: what changed and who can use them?, on the models that will carry it.
Michel Elijah covers technology, streaming and online security for Tech X Telco. He writes practical how-to guides on everything from email troubleshooting to spotting the latest scams doing the rounds in Australia, with a focus on clear steps anyone can follow.
Spotted something wrong? Corrections are recorded in the open. Report a correction
