The Era of Transparency: Anthropic Embeds Watermarks in Claude's Text
Have you ever wondered whether the text you are reading was written by a person or by artificial intelligence? This question, which until recently was a matter of guesswork, is now becoming a technical fact. Anthropic, one of the main players in AI, has announced a significant step: embedding invisible watermarks in text produced by the new Claude models [1, 2].
Why Is This Happening Now?
Anthropic's move is no coincidence. It comes as part of preparing for strict European regulation, and in particular the EU AI Act, which took effect in August 2026 [1, 3]. The law requires technology companies to ensure full transparency about content created by a machine, in order to fight disinformation and protect users [2].
How Does It Work in Practice?
Unlike visual watermarks on images, the watermarks in Claude's text are "invisible" to the human eye [1, 5]. They are embedded directly in the structure of the generated text, without harming the quality, readability, or meaning of the answers [1].
Beyond text, Anthropic also adds provenance metadata (Provenance) to supported files (such as images and certain text files), using cryptographic signatures [1, 5]. This means it will be possible to identify not only whether the content was created by Claude, but also whether it was manipulated or edited afterward [1].
Is It Really Effective?
Even now, before the technology has been fully deployed, tools have popped up online that try to "clean" these watermarks [3]. It is a classic cat-and-mouse game: on one side, regulators and companies want transparency, and on the other, users fear tracking or the use of watermarks to verify "original work" in academic institutions or workplaces [6].
It is important to note that Anthropic has already started offering an API for detecting watermarks in a limited Preview version, intended for regulatory bodies and media outlets. This shows that the goal is not only to "mark" the content but to enable active verification of it [8].
The Bottom Line
Embedding watermarks is a necessary step on the way to building trust in artificial intelligence systems. As models get smarter (like Anthropic's new Fable and Mythos series [8]), the ability to tell human creation from machine creation becomes critical.
For us, the developers and tech people, this means the future of AI is more transparent. We will see more tools that verify content, and less room for haze. Will it solve all the reliability problems? Probably not, but it definitely sets a new bar in the industry.
Additional sources to explore: