道德物化:大模型人机价值对齐的技术伦理进路
Moralizing Technology:Human-machine Value Alignment and Ethic in Large Language Models
袁雨晴 1陈昌凤1
作者信息
- 1. 清华大学新闻与传播学院 北京100084
- 折叠
摘要
以ChatGPT、Sora为代表的大模型在语料库、人工标注编码等方面存在着政治、性别、种族等方面的偏见,使得人机价值对齐问题成为关注的焦点.新一代人工智能带来的价值观和伦理问题,已经成为关乎人类未来的紧迫而重大的论题.人机价值对齐有何思想源流?有哪些研究的面向?有怎样的理论路径和实践指导?本文从现实语境和理论基础出发,从道德物化的新技术伦理进路中探究人机价值对齐问题.道德物化通过发掘技术的道德能动性,将抽象的道德观念嵌入具体的技术结构与情境中,提供了一种内部的、技术伴随性的伦理思路,从而寻求人机价值对齐的三大路径:价值敏感性设计、构建人工智能道德体以及强化算法审计.
Abstract
Generative Artificial Intelligence such as ChatGPT and Sora have attracted world-wide attention for its exceptional capability of high-quality responses to human inquiries.However,political,gender,and racial biases embedded in these models raise concerns about human-machine value alignment.Based on the technological context and theoretical foundation,we propose an ethi-cal approach to explore technology agency by embedding abstract moral concepts into specific techno-logical structures and situations.From an internal,accompanying perspective,value-sensitive de-sign,artificial intelligence agent construction,and algorithm auditing can be applied to reach hu-man-machine alignment.
关键词
大语言模型/人机交互/价值对齐/智能伦理/道德物化Key words
large language models/human-machine interaction/value alignment/intelligent ethic/moralizing technology引用本文复制引用
基金项目
国家社科基金重大项目(18ZDA307)
清华大学-丰田联合研究基金项目(QH202211)
出版年
2024
NETL
NSTL
万方数据