Anthropic explains how watermarking works in Claude's texts
According to Anthropic, light editing probably will not completely remove the watermark from text generated by Claude, and the company is preparing a tool to detect it. This also applies to offers, marketing content, and company documents.
On Friday, August 14, Anthropic published a post explaining how watermarking works in texts generated by Claude. The company is introducing it to meet the requirements of the EU Code of Practice on the transparency of AI-generated content. The code is linked to Article 50 of the AI Act, the EU regulation on artificial intelligence. The obligation to label such content in an identifiable way took effect on August 2, 2026. Details are provided by Anthropic (opens in a new tab) and TechCrunch (opens in a new tab).
The watermark is created through choices that do not matter for the meaning, for example between two synonyms describing cloudy weather. The model then chooses words according to a pattern determined by a key and a few preceding words. A reader will not notice it, but anyone with the key will detect it. Anthropic uses the SynthID Text method, described by Google DeepMind in 2024. The company also announces an API (application programming interface) for detecting the watermark. At first, it will be available in a limited version to, among others, regulators, media outlets, researchers, universities, and companies with compliance obligations, meaning compliance with regulations.
According to Anthropic, light editing probably will not remove the watermark completely. Only rewriting the text, with every word replaced, will do that. The watermark carries over with the text when it is copied and pasted. When Claude only edits text written by a human, the watermark has few words in which it could be embedded. In code, it appears mainly in comments. The company assures that the watermark does not contain data identifying a person, organization, or specific conversation. Watermarking works worldwide because, as Anthropic says, it cannot currently be permanently restricted to a single region.
The announcement triggered criticism from some users. Business Insider, cited by TechCrunch, reported that dozens of people on X declared they would cancel their subscriptions. The code was also signed by, among others, Google, Meta, Microsoft, and OpenAI, so similar watermarks will appear in other tools.
For companies, this means that offers, marketing texts, and documents prepared with Claude and only edited by a human may remain recognizable as AI-generated for entities that have access to the tool detecting the watermark.