Anthropic will begin adding invisible watermarks to Claude texts so that machines can tell that the content has passed through the company’s AI. This will be applied to Claude models released starting August 2, 2026.
Anthropic will begin adding invisible watermarks to Claude texts so that machines can tell that the content has passed through the company’s AI. This will be applied to Claude models released starting August 2, 2026.
The company announced this on its website.
The innovation is related to the requirements of the AI Act, a law that sets rules for the development and use of artificial intelligence in the EU. However, Anthropic has decided not to limit it to Europe, so the labels will work for users around the world.
Visually, a person will not be able to see such a watermark, but a machine will read it. Claude will imperceptibly embed the marking directly into the text. According to Anthropic, since the mark becomes part of the text itself, it will even be able to survive some editing. The marked text will be received in all the company’s products: in the regular Claude, Claude Code, Claude Cowork and others. Images will also be marked.
However, the watermark found will not mean that Claude wrote the text from scratch. For example, a person can write an article themselves, and Claude can only be asked to translate, shorten, or edit it. The result can still receive a label. So it would be more correct to perceive it as an indicator that «this text was processed by Claude,» rather than as evidence of AI authorship.
But the absence of a tag does not mean that the text was definitely written by a human. For example, after heavy editing, mixing with other text, or translation, the text may lose its watermark. The same can happen with files: the origin information may disappear after the file is converted, re-saved, or a screenshot is taken.
Ordinary users will also have access to detection tools. In addition, media, fact-checkers, researchers, and law enforcement will be able to use them free of charge and without restrictions on the number of requests. In the future, the EU wants to simplify the verification itself. By February 2027, signatories to the European Code of Practice on AI Content Transparency, which includes Anthropic and other large technology companies, must make their verification systems more compatible so that the user does not have to run content through several different detectors.
Older Claude models will remain without them for now. For AI models that were already on the market before the new requirements came into effect, the EU has provided a transition period until December 2, 2026. But Anthropic says it is already working on adding the marking to older models.
The idea of invisible labels in AI texts is not new. For example, back in 2024, Google introduced SynthID Text, a technology that leaves a statistically significant signal in the text generated by a model.
Comment hidden for violating commenting rules.