Shortly after Anthropic disclosed the functioning of the invisible watermark intended to track texts produced and modified by Claude, developers began making available tools that allegedly can weaken or eliminate this marking.
The first code was developed by programmer Guillaume Meyer and quickly gained popularity, accumulating over one hundred thousand contributions on GitHub. Meyer stated on the X platform that he created the remover only as an experiment and that his goal is not to combat content identification, but rather the methodology of the watermark itself.
This Claude detection mechanism is being implemented in line with the European Union's Artificial Intelligence Act, which established the obligation to label content generated by artificial intelligence. Currently, approximately one hundred and ninety organizations have adhered to the code of conduct, and it is expected that textual marking will become a standard in the sector by the end of this year.
Anthropic's watermark operates through a statistical pattern established during the word selection process. When Claude needs to choose between synonyms, the system uses a key to direct part of these decisions, thus creating a detectable signature.
The company itself emphasized that even minimal modifications have the potential to weaken this mark. This is where Meyer's tool focuses: the text is rewritten, including multiple times, using another language model that does not employ the same marking.
Several other methods quickly emerged, following similar processes, all based on synonym changes and sentence reorganization. Since detection depends on a specific pattern, such adjustments may be sufficient to decrease the probability of the text being classified as synthetic, which could be achieved with a simple human review.
The technology used by Anthropic is based on SynthID-Text, a method published by Google DeepMind two years ago. This system acts on so-called 'low-risk choices': when two words have similar meanings in a sentence, the AI selects one of them following a predefined rule. The intention is to keep the text natural for the reader, while the sequence can be identified by those who possess the verification key.
Anthropic clarifies that its verification indicates a probability and cannot accurately distinguish whether a text was entirely generated by AI or merely edited by the model. However, it is precisely the potential 'false positives' of this type of system that cause apprehension among developers and certain professionals.
There is a fear that educational institutions, universities, recruiters, or clients may start using detectors as a factual evaluation criterion, which could harm individuals who use AI for specific tasks or who simply do not use it.
From the users' perspective, another criticism raised is that the implementation of this technology may reduce the overall quality of content produced by Claude, given that the result must adhere to this marking standard. In its description of the new technology, Anthropic assures that internal tests did not indicate any decline in quality.