Anthropic has provided additional technical details about the watermarking system it recently introduced for Claude, according to TechCrunch AI. The company addressed key questions about how the technology functions in practice and its limitations.
According to the report, the watermarking works by subtly influencing Claude’s word choice patterns during text generation in a way that can be detected by Anthropic’s verification tools but remains invisible to human readers. However, the company acknowledged that the watermarks can be degraded or removed through editing. TechCrunch AI reports that modifications to the generated text, such as paraphrasing or significant rewrites, may reduce the detectability of the watermark.
Regarding code generation, Anthropic explained that the watermarking system also applies to code output from Claude. The company indicated that the watermarks are designed to work across different types of content, including programming languages, though the specific implementation details for code were not fully elaborated in the available reporting. The disclosure comes as AI companies face increasing pressure to develop methods for identifying AI-generated content amid concerns about misinformation and academic integrity.