Labeling AI-generated content: why these circumvention tools could backfire on you

Anthropic recently announced the introduction of a labeling system applied to certain of its AI models, such as Claude Code, Claude Cowork, and Claude Tag. Specifically, this involves watermarks embedded in the text and signed provenance metadata. Thus, content modified or generated by artificial intelligence is labeled as such.With this new system, Anthropic’s aim is to comply with Article 50 of the AI Act, which requires providers of AI systems to establish a mechanism for labeling content generated or modified by AI.

Anthropic announces a labeling system as services for circumventing it multiply

As is often the case when rules, constraints, or prohibitions are put in place, some people quickly seek to circumvent them. Indeed, since Anthropic announced its labeling system, services claiming to be able to remove it have proliferated online. Examples include the GitHub projects watermarks-remover, claude-watermark-cleaner, remove-ai-watermarks, and noai-watermark. On paper, their promises are clear: remove watermarks and metadata applied by Anthropic to content generated or modified by one of the company’s models. In practice, however, these promises may not hold up.

Why these sites may not be reliable

First and foremost, it should be recalled that Anthropic has not explained how its labeling system works. It is therefore difficult to develop a service capable of circumventing it effectively. In this respect, some of these services simply do not seem to work. Pasquale Pillitteri, a cybersecurity researcher, tested several of these tools and found that one of themstill left information hidden in the text.But the real difficulty for these services lies in removing the watermark itself. It is embedded in the model’s word choices, rather than in the form of hidden characters. Therefore, to try to remove it, it would be necessary to rewrite a large part of the text using another AI model.

More generally, it is advisable to be cautious with some of these sites that appear out of nowhere and promise flawless results. It is also worth recalling that Anthropic itself acknowledges that its labeling system has “its limitations”. We now need to wait to learn more about how it works in order to determine to what extent it is actually possible to circumvent it. In any case, for Guillaume Meyer, founder of Memo and developer of the Watermarks Remover tool, this move by Anthropic is simply a “pure compliance requirement to remain in the EU market”.

👉🏻 Follow the latest tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.

Source: Bleeping Computer