查找相似内容

Proposed Chinese AI Safety Standards: A Closer Look Hey there, AI & Law community! On October 11, the National Information Security Standardization Technical Committee in China released a draft document outlining precise regulations for evaluating generative AI models. Unlike the often vague AI regulations, this document provides a clear blueprint for compliance. This standards proposal sets forth rigorous criteria for assessing AI data sources and their content. The document covers topics like training data diversity, moderation, and prohibited content. It emphasizes the need for diversified training corpora and the assessment of data quality. If more than 5% of data is "illegal and negative information," the corpus is flagged for future training. The proposal also suggests that AI companies employ moderators to enhance generated content quality, aligning with national policies and third-party complaints. This implies a potential expansion of the human-driven moderation and censorship workforce in the AI era. Companies are tasked with identifying hundreds of keywords for flagging unsafe or banned content, with separate categories for political and discriminative content. They must also generate more than 2,000 prompts, ensuring fewer than 10% of responses breach the rules. Interestingly, the document encourages subtler censorship measures, such as not refusing to answer sensitive prompts but allowing AI models to respond to specific, non-sensitive inquiries. It's crucial to clarify that these standards are not laws, and non-compliance doesn't result in penalties. However, proposals like these can significantly influence future regulations or work alongside them. The standards receive input from tech experts hired by companies, giving corporations like Huawei, Alibaba, and Tencent a say in shaping these regulations. Their influence could have far-reaching implications for the global AI industry and how AI technologies are regulated worldwide. #AISafety#AIRegulations#GenerativeAI#ContentModeration#ChineseTech#AIInfluence#GlobalAI