生成式人工智能的价值对齐比较研究——基于2012-2023年十大国际新闻生成评论的实验
A Comparative Study of Value Alignment in Generative Artificial Intelligence:Experiments on the Top Ten International News from 2012 to 2023
胡正荣 1闫佳琦2
作者信息
- 1. 中国社会科学院新闻与传播研究所,北京 100021;中国社会科学院大学新闻传播学院,北京 102488
- 2. 中国社会科学院新闻与传播研究所,北京 100021
- 折叠
摘要
作为生成式人工智能良序发展的基础,价值对齐日渐成为关涉人类社会生存安全的重大现实问题.基于对话交互式大语言模型,本文旨在回答价值对齐何为、何在、何以可能的系列问题.研究通过设置不同价值观作为提示词,得到GPT4-turbo、文心一言4.0基于2012-2023年120条十大国际新闻生成的共720条新闻评论,发现中美大模型原生价值存在鲜明的在地化差异,中国大模型原生价值更能体现中国主流倡导的价值导向,美国大模型在展演模式下的价值对齐能力更强.研究进一步探讨了价值鸿沟、价值仿真、价值展演、智能体价值问题,扩展了价值对齐议题的衍生思考空间.
Abstract
Value alignment,which is the basis for the orderly development of generative artificial intelligence,has increasingly become a major practical issue related to the survival and security of human society.Based on the conversational interactive large language model,the research aims to answer a series of questions about what value alignment is,where it is,and how it is possible.By setting different values as prompt words,the study obtained a total of 720 news comments generated by GPT4-turbo and ERNIE Bot 4.0 based on 120 top ten international news items from 2012 to 2023 It was found that there are clear localized differences in the native value of large models in China and the United States,and that American large models have a stronger ability to align their values in the performance mode.Furthermore,the research explored derivative discussion spaces such as value gap,value simulation,value performance,and agent value.
关键词
价值对齐/生成式人工智能/大语言模型Key words
value alignment/generative artificial intelligence/large language model引用本文复制引用
基金项目
国家社会科学基金重大招标项目(22ZDA088)
出版年
2024
NETL
NSTL
万方数据