Can You Ever Prove Something Was Written by AI?
中文摘要
Anthropic 开始为 Claude 的输出添加水印。作者通过构建并攻击自定义水印，揭示了证明 AI 生成内容的实际挑战。
English Summary
Anthropic is watermarking Claude's text. An author built and attacked their own watermark, uncovering insights into the difficulty of proving AI authorship.
Original Excerpt
Anthropic just started watermarking everything Claude writes. So I built a watermark myself, attacked it, and found the number nobody is… Continue reading on Medium »