Anthropic's Claude Watermark Has a Loophole on Older Models
Anthropic is adding a text watermark to four more Claude models to meet an EU AI Act deadline in December. On two older models, developers can skip the watermark by changing one sampling setting. The report says Anthropic is working to close that gap.
Quick glossaryAnthropic, EU AI Act, watermark, sampling parameter
- Anthropic
- An AI company based in the United States, best known for its Claude chatbot and AI models.
- EU AI Act
- A European Union law that sets rules for how artificial intelligence can be built and used.
- watermark
- A hidden marker added to AI-generated text so that people can later check where it came from.
- sampling parameter
- A hidden setting that controls how randomly an AI picks its next words when writing an answer.
Burning Potato score: 60%