Dario Amodei’s Anthropic recently introduced a new way to identify text created by its Claude AI models. The company is using hidden watermarks in AI-generated text, but a new report now claims that developers have already started looking for ways to remove or break these markers.
According to Wired, unlike a normal watermark, Anthropic’s system cannot be seen by people. The company uses patterns in the way Claude chooses and generates words. These patterns can help identify whether a piece of text was created by Claude.
The move is part of Anthropic’s effort to make AI-generated content easier to identify. The US-based AI startup is also working on a detection system that could allow users and organisations to check whether text was generated by its AI models.
However, the new technology has already attracted attention from developers who are trying to find ways around it.
Some developers are experimenting with other AI models to rewrite Claude-generated text. By changing the wording and sentence structure, they can potentially disturb the patterns used by the watermark, the report said.
“This has created a new challenge for AI companies. As companies develop better ways to identify AI-generated content, developers can also build tools designed to make that content harder to detect,” it added.
Anthropic’s watermarking system is also raising questions about how AI-generated content should be treated.
People do not always use AI to write something from scratch. They may use Claude to correct grammar, improve sentences, translate text or rewrite an existing piece of content.
In such cases, it can become difficult to decide whether the final text should be considered AI-generated. This could be important for students, writers and professionals who regularly use AI tools as part of their work.
The development comes after governments around the world pushing Big Tech companies to make AI-generated content easier to identify. New transparency requirements in the European Union are expected to increase pressure on AI companies like OpenAI and Anthropic to develop reliable ways of marking or detecting content created by their models.
