Anthropic reassures on AI watermarking system amid user concerns

Anthropic addresses backlash over its new invisible text watermarking system for Claude, emphasising its non-intrusive design and compliance with EU regulations, while acknowledging its limitations in detecting AI-generated content.

Anthropic has moved to calm concerns over its new text watermarking system, after users criticised the plan to tag Claude’s outputs. In a blog post published on Friday, the company said the mechanism is designed to create an invisible pattern in word choice rather than alter the meaning, tone or readability of generated text. Anthropic also said the system is meant to satisfy obligations under the EU AI Act, which requires clearer marking of AI-generated content.

The company says the approach is based on Google’s SynthID Text and works by changing the source of randomness used when Claude selects words. Anthropic argues that this does not make the output less creative or less readable, pointing to internal testing in which it says it saw no material change in quality. It used a Monopoly analogy to explain that the pattern is hidden in the sequence of choices, not in the visible text itself.

Even so, Anthropic acknowledged clear limits. Its detector can only estimate the probability that text came from Claude; it cannot prove whether text was written by a human or by another AI system. The company also said watermarking is less effective on short passages, factual snippets and code fragments, where there are fewer word choices to alter. In code, it said the method may apply to comments but should have negligible effect on the actual programme.

Anthropic also addressed privacy concerns, saying the watermark and its key do not reveal anything about a user, their organisation or their conversations. It added that the system is being launched globally rather than only in Europe because it does not yet have a reliable way to limit it by region.