Anthropic added watermarks to Claude — developers immediately released "erasers"
Anthropic's decision to watermark texts generated by Claude sparked a new surge of interest in services for removing such marks. Some tools appeared within days of the company's announcement and have already garnered thousands of users, writes devby.io.
Business Insider reports that interest in the query "AI watermark remover" in Google Trends in the US increased by 60% in a week. Simultaneously, several projects emerged promising to remove or weaken watermarks in texts and files.
One of them is the open-source Watermarks Remover, launched by French entrepreneur Guillaume Meyer. According to him, he assembled the first version in about five hours. The project has already received over 14,000 stars on GitHub.
The tool removes hidden symbols and metadata, then rewrites the text to disrupt the statistical pattern of word choice on which Anthropic's watermark is based. However, the developer does not guarantee that the watermark will be completely removed.
"I am totally in favor of indicating content origin. I am specifically against this watermark technology, and that's an important distinction," Meyer stated.
In his opinion, such an approach oversimplifies the issue of authorship, as the same watermark can appear in both entirely generated text and human-edited material.
Anthropic previously emphasized that its watermarking is not intended to determine authorship. It merely allows assessing the probability that Claude was involved in creating the text.
Another developer, Ansh Aneja, released his own tool for removing Claude watermarks on the day the feature was announced. Later, he published a local open-source version called MarkScrub. According to him, 8,500 people used the service within the first 24 hours.
The effectiveness of these solutions has not yet been confirmed by independent tests. ETH Zurich researcher Thibault Glaaigen notes that watermarks fundamentally have a limitation: if a text is rewritten strongly enough, the statistical pattern can be broken.
The legal status of such tools in the EU remains uncertain. The AI Act requires model providers to make watermarks resistant to removal and alteration, but it does not directly prohibit third parties from creating tools for their removal.
However, using such services to pass off AI-generated text as human-created may violate Anthropic's rules. The company's policy prohibits presenting model outputs as human-created.