tech
Claude to start watermarking AI-generated text
Anthropic says it will change way chatbot makes small, random choices, to comply with EU regulation
TL;DR
- Anthropic is updating its Claude AI model to watermark generated text to comply with an EU regulation effective December.
- The watermarking is intended to be undetectable to the average reader but will alter the random choices AI models make in generating text.
- Commentator John Gruber argues this will 'adulterate' writing and force less precise word choices, negatively impacting quality.
- Professor Steven Murdoch believes the change will have no noticeable impact, as AI models already rely heavily on randomness.
- Watermarking may also help prevent 'model collapse' caused by AI training on AI-generated content.