Anthropic has implemented AI watermarking for its Claude chatbot to comply with the EU AI Act's transparency requirements. The feature embeds invisible signals into AI-generated text, but has drawn criticism from some users concerned about being caught using the tool.
The implementation of AI watermarking by a major AI provider like Anthropic signals a significant step towards regulatory compliance and transparency in the rapidly evolving AI landscape, while also highlighting user concerns about the practical implications and ethical considerations of such technologies.
Anthropic has implemented AI watermarking for its Claude chatbot to comply with the European Union's AI Act, which mandates the labeling of AI-generated content. The technique embeds invisible signals into text, derived from a 2024 Google DeepMind paper, allowing for the identification of AI-generated outputs.
While the watermarking aims to satisfy regulatory requirements, it has sparked discontent among some users. Critics argue the feature is unethical, will unfairly flag average users who use AI for tasks like reorganizing paragraphs or summarizing transcripts, and is hypocritical given the training data used for AI models. Conversely, many users support the watermarking as a necessary tool for transparency and to mitigate the risks associated with AI-generated content.
Anthropic plans to offer an API for checking text for Claude's watermark and noted that the feature does not alter user ownership rights. However, the watermark's effectiveness is limited; it indicates Claude's involvement rather than general AI generation, may not apply to factual passages or code, and can be removed by a complete rewrite.