Anthropic will watermark Claude-generated text

Anthropic has explained how text watermarking will work in future Claude models. As Anthropic wrote in its technical overview, the company is adding the system to comply with EU AI Act transparency rules for AI-generated content.

What will change

The watermark is not a visible label and does not add hidden Unicode characters. Instead, it slightly changes how Claude makes low-impact word choices during generation. Across a long enough text, those choices create a statistical pattern that can later be checked with the right detection key.

Anthropic says the approach is based on SynthID-Text, the watermarking method published by Google DeepMind. The company also says the method should not affect output quality, speed, token usage, or pricing.

Consequences for users

• Anthropic plans to provide a watermark detection API.
• Short passages may be harder to detect because there are fewer word choices to analyze.
• Light editing may leave enough signal, while a full rewrite can remove it.
• Code should contain less watermarking because working code leaves fewer arbitrary choices.
• The watermark is not designed to identify a specific user, company, or chat.

As TechCrunch reported, some Claude users reacted negatively because watermarking affects not only fully generated text but potentially translations or heavily edited content. That makes the real issue less about the watermark itself and more about how schools, platforms, employers, and publishers decide to use detection tools.