Key facts
- Claude AI acknowledges watermarking limitations.
- Watermarks can be removed through editing or paraphrasing.
- Claude AI supports watermarking as part of a broader system.
- Claude AI notes watermarking limitations against malicious actors.
- John Gruber criticizes Anthropic's AI watermarking plan.
- John Gruber calls the plan 'patently offensive'.
- Gruber argues watermarking compromises quality and precision of AI writing.
Anthropic's AI chatbot, Claude, has acknowledged significant limitations regarding the effectiveness of watermarking AI-generated text. The AI model itself stated that watermarking is not a foolproof solution because the marks can be easily removed through editing or paraphrasing. While Claude supports the concept of watermarking as a component within a broader system of AI safety and security measures, it emphasized that the technique is vulnerable to malicious actors who could deliberately strip the watermarks.
Tech blogger John Gruber has strongly criticized Anthropic's plan to implement watermarks on text generated by Claude. Gruber described the proposal as "patently offensive" and argued that such watermarking would compromise the quality and precision of AI-generated writing. His criticism centers on the potential degradation of the output, suggesting that the process of embedding and detecting watermarks could interfere with the integrity of the text itself.