Claude Adds Invisible Watermarks to All AI Text

Anthropic has begun embedding invisible watermarks into all text generated by its Claude models. This system works at the model level, meaning the mark stays attached whether the text is copied from the app, an API, or integrated platforms like Slack. The company says these marks will soon be readable by detection tools, which will help distinguish human writing from machine output.

While the specific method remains private, experts suggest it likely uses a green list approach. The AI selects words from a specific list in a way that looks random to the human eye but follows a detectable pattern. However, Anthropic notes that these detectors have limitations. Short snippets of text might not provide enough data for an accurate read, and heavy editing or paraphrasing could effectively strip the watermark away.

The update comes as AI developers scramble to comply with the EU AI Act. Since Anthropic signed the Code of Practice on Transparency, they must ensure their AI content is clearly marked by December 2, 2026. While some users fear this could lead to more aggressive policing, it is worth noting that a positive detection does not necessarily prove cheating, as many people use Claude to summarize or polish their own existing work.

Looking ahead, the market will need to see how well these detection tools actually perform. Because this move is largely driven by regulatory pressure, other AI labs are expected to follow suit. For now, the watermarks remain a silent feature until the official detection tools are released to the public.