OpenAI 为遵守 EU AI Act，未来几周将对欧盟区内 ChatGPT 和 Codex 的合格文本输出嵌入基于措辞的隐形水印，API 客户即日起可全球为部分模型开启文本水印。作者质疑该方向，指出 400-token 测试中替换 25% 同义词会使检测率从约 92% 降至 17%，且水印无法区分人类贡献程度，经 ChatGPT 编辑的自写文本也可能携带水印。
作者在报道 OpenAI 欧盟文本水印计划之外，补充了同义词替换使检测率从约 92% 降至 17% 的局限，提供了评估该政策的参考角度。
OpenAI will start changing ChatGPT’s word choices in the EU to make its text detectable, and I don’t like this direction.
Over the coming weeks, ChatGPT and Codex outputs will receive invisible text watermarks in response to the EU AI Act. The pattern is embedded in the wording itself.
OpenAI says its tests show no meaningful quality loss. Still, I’m dislike with detection requirements influencing how a writing tool phrases my text.
Especially given the limitations: in one test on 400-token passages, replacing 25% of words with synonyms cut watermark detection from about 92% to 17%.
The watermark also cannot distinguish how much a human contributed. Even text you wrote yourself and had ChatGPT edit can carry it.
The EU is once again concerning itself with the really important things (sarcasm).
We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology. Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model. In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Customers using our API can turn on text watermarking for select models worldwide today.在 X 查看被引用的帖子
来源：Chubby♨️ · x.com