Anthropic watermarks Claude text, sparking user backlash and workarounds

Anthropic has begun embedding invisible watermarks into text generated by its Claude AI models, a move designed to comply with the European Union's AI Act transparency rules. The company says the watermark is imperceptible to readers but detectable by a key, and it does not affect output quality Source: The Verge. However, the announcement has triggered a swift backlash, with some users canceling subscriptions and developers creating tools to strip the marks Source: Gizmodo, The Next Web.

How it works

Anthropic's watermarking, based on Google DeepMind's SynthID-Text method, subtly biases Claude's word choices in low-stakes situations—like picking between "overcast" and "grey"—to create a statistical pattern that can be verified with a key Source: Engadget. The company plans to release a detection API and notes that light editing won't remove the mark, but a complete rewrite will Source: TechCrunch.

The backlash

Within hours of the announcement, developer Guillaume Meyer published an open-source tool to remove watermarks, which quickly went viral on GitHub Source: Wired. Critics argue the watermarking is a blunt instrument that fails to distinguish between AI-generated and AI-edited text, potentially harming users Source: The Next Web. Some users have canceled subscriptions, and Google Trends shows a 60% rise in searches for "AI watermark remover" Source: The Next Web.

What's next

Anthropic defends the feature as compliance, not surveillance, and says other major model developers have signed the same code of practice and will implement their own watermarks Source: TechCrunch. The long-term impact on AI adoption and content provenance remains to be seen.