Anthropic’s decision to add an invisible watermark to Claude-generated text sparked a new online debate, with users discussing whether the mark can be removed. A suggestion making the rounds involves generating text in languages such as Dutch or German and then translating it into English. Current reporting suggests that simply changing the language is not a reliable way to eliminate Claude’s watermark.
Anthropic introduced the system as part of its efforts to improve transparency around AI-generated content. The watermark is designed to be machine-readable rather than visible to people. It works by subtly influencing word choices during generation, creating a statistical pattern that can later be detected.
The idea behind the language-switching claim is fairly straightforward. If Claude generates content in one language and another system subsequently translates it, the wording changes substantially. Some users have therefore speculated that the original watermark might disappear during translation.
However, that does not automatically mean the content becomes undetectable. Anthropic’s watermarking approach is based on a version of Google DeepMind’s SynthID-Text technology, which influences the model’s token choices rather than inserting a hidden character or obvious code into the text. This distinction is important because the watermark is embedded in statistical patterns within the generated language.
Recent reports indicate that Claude-generated translations can themselves carry a watermark, since Claude still selects the words in the translated output. In other words, changing the language does not necessarily remove the underlying signal if Claude is responsible for generating the new version.
Researchers and developers have also been experimenting with paraphrasing, rewriting, and other transformations. Some forms of substantial rewriting may weaken or eliminate a watermark, but that is different from proving that a simple language switch reliably defeats detection.
This distinction matters since a watermark detector is not necessarily answering the question, “Was every sentence originally written by Claude?” Anthropic said its watermark is intended to indicate that Claude processed or generated text, rather than establish definitive authorship.
The move is closely linked to the European Union’s AI Act, which introduces transparency requirements for providers of generative AI systems. Anthropic has said its watermarking approach is being deployed globally rather than being restricted to European users.
The company is also expected to provide a detection API, giving developers another way to identify Claude-marked content.
Also Read: Anthropic’s Mythos 5 Created Fake Developers in Cybersecurity Test
The controversy ultimately goes beyond whether a particular trick can bypass a watermark. Critics argue that AI-assisted writing can include human-created material that is merely edited, translated, or proofread by an AI system. A watermark could therefore indicate AI involvement without necessarily proving that the machine created the underlying ideas or original work.
For now, the Dutch-to-English or German-to-English theory should not be treated as a guaranteed method for watermark removal. The technology is still evolving, and so are attempts to test its limits. What is clear is that AI companies and developers are entering a new phase where proving the origin of digital content is becoming increasingly important.