AI Watermark War Escalates: Anthropic Adds Invisible Marks to Claude Content

Artificial intelligence companies are increasingly turning to digital watermarking to identify machine-generated content, but developers are already working on ways to bypass these invisible markers. Anthropic has announced that future Claude-generated content will carry invisible, machine-readable watermarks. The move is linked to transparency obligations under Article 50(2) of the EU AI Act, which came into effect on August 2, 2026.

The watermarking approach is based on Google DeepMind’s SynthID-Text research. Instead of inserting visible symbols or hidden Unicode characters, it subtly influences the model’s choice of words and tokens so that a machine with the appropriate detection mechanism can identify a pattern. Anthropic says the technique should not affect the meaning, readability or quality of Claude’s responses.

However, the technology has quickly triggered a cat-and-mouse race. Developer Guillaume Meyer released an open-source tool that attempts to remove Claude’s watermark by using other AI models to rewrite content. The project attracted significant attention on GitHub, while other developers have explored rewriting, sentence restructuring and translation-based approaches. The effectiveness of these methods cannot yet be conclusively established because Anthropic’s official detection system is still being developed.

A major concern is the possibility of false positives. Anthropic itself says the watermark would indicate that Claude was likely involved in processing content, but it cannot establish whether Claude originally wrote the material, merely edited it, or whether another AI system generated it. Experts therefore warn that relying too heavily on AI detection could unfairly affect students, researchers or job applicants.