- Published on
Claude adds watermarks to all AI text, and users worldwide are furious!
- Authors
- Name
- DeepHorus
AI company Anthropic recently announced that it will forcibly add machine-readable watermarks to all text generated by its Claude models, in response to requirements of relevant provisions of the EU Artificial Intelligence Act. The policy applies to all users globally, not just those in the EU, which has sparked widespread controversy and has been criticized as a "one-size-fits-all" and paternalistic governance approach.
Article 50(2) of the EU AI Act stipulates that from August 2, 2026, all AI systems that generate synthetic text and other content must ensure that outputs include detectable machine-readable markers. Anthropic signed the EU AI Generated Content Transparency Code of Practice in July and decided to build watermarks into all new model versions from August 2, with text generated by existing models also gradually to be added.
Unlike the "character watermark" used for specific users earlier this year, this time it adopts statistical watermarking technology based on Google's SynthID-Text. This technology, through a preset key, makes the AI tend to choose specific synonym combinations when generating text, forming machine-recognizable statistical features. Anthropic claims that the watermark adds no extra content, consumes no additional tokens, and carries no identity information, making it impossible to trace back to specific users.
However, Anthropic has "added extra layers" on its own in terms of complianceāthe EU Act targets only EU users, but the company decided to roll out globally without offering a choice. This approach quickly sparked user dissatisfaction. A developer released an open-source tool specifically to strip AI watermarks within five days, which has gained over 11,000 stars on GitHub. Ironically, parts of the tool's source code were written by Claude itself.
A deeper controversy lies in the applicable scenarios of the watermark. Whether it's translation, polishing a resume, or checking grammar, all Claude outputs may be labeled as "AI content." Technology analyst Ben Thompson sharply pointed out that if humans complete the research, opinions, and first draft, and only let the model polish, labeling the final work as AI content is equivalent to "listing the pencil as the author because the article was written with a pencil."
Adobe Photoshop provides a reference example: its C2PA watermark is only added when using Firefly generative tools; manual cropping, color adjustment, and other operations do not trigger it. In contrast, Claude's scenario-agnostic policy may lead companies, schools, and publishers to simply equate "watermark presence" with "purely AI-generated," causing extensive collateral damage. Moreover, statistical watermarks are not foolproof; they can be bypassed by rewriting with other models, which is also the reason OpenAI previously abandoned the full deployment of ChatGPT watermarks.
Anthropic's mandatory watermark policy reveals that in an era when generative AI has become a super tool, defining "who wrote it" with a single watermark is outdated. The industry generally agrees that AI content needs to be labeled, but equating probabilistic watermarks directly with "AI works" and adopting a globally mandatory approach goes beyond reasonable boundaries. How to strike a balance between transparency and user autonomy remains an urgent issue to be solved.
