返回全部动态
Anthropic 为 Claude 输出加水印，批评者质疑权衡
原标题：Anthropic watermarks Claude's output, but critics question the tradeoffs
AI 摘要
Anthropic 正在为 Claude 的输出添加水印，通过统计模式标记 AI 生成的文本，以符合欧盟法规。批评者认为这会影响文本质量，因为模型会基于水印密钥而非语义选择词汇。法律行业对此反应不一，认为在特定情况下可能引发透明度问题。水印可通过改写工具去除，且 Anthropic 在全球范围内应用该功能。
以上摘要由 AI 生成，可能存在误差。事实请以原文为准。
正文节选
Anthropic watermarks Claude's output, but critics question the tradeoffs Key Points - Anthropic is embedding a watermark in Claude that tags AI text through statistical patterns in word choice to comply with EU regulations. - Critics say text quality suffers because the model picks words based on the watermark key rather than meaning. Tools like Declaude can strip the marking, though. - For law firms, the system is mostly a non-issue. But permanently detectable AI text could get awkward during f
发布时间：2026-08-17 19:43
抓取时间：2026-08-17 19:51
来源机构：THE DECODER