Anthropic has for the first time detailed how its built-in text labeling system works for content generated by its Claude neural network. The feature is being introduced to comply with the European Union's Artificial Intelligence Act, which requires AI-generated content to be marked. Users have expressed concerns that the technology could reveal the use of the AI, but developers have assured that the labeling will not affect the quality of responses, reports infohub.kz.
As explained by Anthropic, the watermark is created through statistical word selection in places where it has minimal impact on the meaning of the text. The company uses Google DeepMind's SynthID-Text algorithm, which leaves a mathematical pattern invisible to the average reader. To detect such markers, the company plans to release a dedicated API tool.
The company also stressed that the technology does not increase the cost of requests and does not affect code generation. However, a complete rewrite of the text could remove the watermark entirely.