OpenAI宣布将为在欧盟生成的ChatGPT和Codex文本添加隐形水印，以符合欧盟AI法案的透明度要求1。该法案的相关规定将于8月2日生效，水印功能将在未来数周内向欧盟用户推出1。这项技术通过微妙改变模型的词汇选择来实现，使文本检测器能够识别内容是否由AI生成1。OpenAI同时发布了名为textGrain的技术报告，该方法是与宾夕法尼亚大学和耶鲁大学研究人员共同开发的1。
水印的检测效果存在一定局限性1。在未经处理的文本中，检测率可达92%，但若替换其中10%词汇的同义词，检测率将下降至66%1。OpenAI强调这一功能不会识别用户身份，模型性能也不会出现明显变化1。API开发者可从即日起选择启用该功能，但全球范围内默认保持关闭状态1。OpenAI仅向获批的研究机构提供检测器访问权限1。
OpenAI announced plans to embed invisible watermarks into text generated by ChatGPT and Codex within the European Union, marking a compliance effort ahead of the EU AI Act's transparency requirements taking effect on August 2 1. The watermarks function by subtly altering the model's vocabulary choices and will be detectable by specialized tools, though they are not designed to identify individual users and do not measurably impact model performance 1.
The feature will roll out to EU-based ChatGPT and Codex users over the coming weeks, while API developers can opt into the functionality starting immediately, with the capability remaining disabled by default globally 1. OpenAI has released a technical report on the watermarking method, called textGrain, developed in collaboration with researchers from the University of Pennsylvania and Yale University 1. The company has indicated that access to detection tools will be restricted to approved research institutions 1.
However, the watermarking approach has limitations: replacing just 10 percent of words with synonyms can reduce detection accuracy from 92 percent to 66 percent 1. This announcement follows Anthropic's move two months prior to implement watermarking across all Claude-generated text on a global basis 1. OpenAI had previously delayed releasing this feature in 2024, citing concerns that users might migrate to competitors that do not employ watermarking 1.
评论
还没有评论，欢迎留下第一条。