人工智能的自学习能力、就业冲击与法定工作时间

白军飞¹　朱秋博²　罗丽¹

（1. 中国农业大学；2. 首都经济贸易大学）

摘要：近年来，自动化、工业机器人和人工智能对就业的影响受到国内外学术界的高度关注，然而现有研究往往将三者混为一谈。本文在深入比较人工智能与前三次工业革命的技术特征之后认为，人工智能的自学习能力使其与传统的自动化、工业机器人存在本质区别。据此，本文将自学习能力引入阿西莫格鲁和雷斯特雷波（Acemoglu & Restrepo）的任务模型，从理论上重新考察人工智能对就业的影响。结果显示，考虑自学习能力后，人工智能对就业影响的理论预期比不考虑该能力时更加严峻，失业和收入极化问题可能有所加剧，资本与劳动者的关系将面临新的挑战。进一步，本文首次提出通过有计划地缩短法定工作时间来缓解人工智能对就业冲击的应对策略，并从理论上论证了该策略在兼顾效率与公平上的优势。

关键词：人工智能；自学习能力；就业；法定工作时间；任务模型

作者简介：白军飞，中国农业大学经济管理学院教授、博士生导师，北京，100083；朱秋博，首都经济贸易大学经济学院副教授，通信作者，北京，100070；罗丽，中国农业大学经济管理学院博士后。

基金项目：中国农业大学2115人才培育发展支持计划

引用格式：白军飞，朱秋博，罗丽. 人工智能的自学习能力、就业冲击与法定工作时间[J]. 经济与管理研究，2026，47（7）：110-124.

Self-Learning Capability of Artificial Intelligence, Employment Impact, and Statutory Working Hours

BAI Junfei¹, ZHU Qiubo², LUO Li¹

(1. China Agricultural University, Beijing 100083;

2. Capital University of Economics and Business, Beijing 100070)

Abstract: In recent years, the impact of automation, industrial robots, and artificial intelligence (AI) on employment has attracted widespread attention. However, existing research often conflates these three factors, treating AI as merely an extension of previous automation technologies. After conducting an in-depth comparison between AI and the technological characteristics of the three industrial revolutions, this paper argues that the self-learning capability of AI sets it apart from automation and industrial robots. Unlike earlier technologies, which mainly replaced human labor in routine and repetitive tasks, AI is capable of learning from data and continuously improving, thereby performing increasingly complex cognitive and non-routine tasks. This fundamental difference has profound implications for the labor market.

Accordingly, this paper introduces the self-learning capability into Acemoglu and Restrepo’s task-based model, aiming to re-examine the impact of AI on employment theoretically. It modifies the model by relaxing the assumption that labor has a comparative advantage in newly created tasks and assumes that AI can rapidly learn and master newly created tasks, potentially at a lower cost than retraining workers. This modification captures the unique feature of AI: its capability to continuously expand its comparative advantage across tasks. The results indicate that when accounting for self-learning capability, theoretical expectations regarding the impact of AI on employment are more severe than those obtained without considering this capability. On the one hand, the displacement effect of automation is no longer counterbalanced by the reinstatement effect from the creation of new tasks. On the other hand, unemployment and polarization may intensify, and the relationship between capital and labor will face new challenges.

In addition to the theoretical analysis, this paper reviews existing policy proposals to mitigate the negative employment effects of AI. These proposals fall into two categories: slowing down AI adoption through taxes or regulations and establishing comprehensive income support mechanisms. However, both approaches have significant drawbacks and are not conducive to the growth of social demand. Slowing down AI innovation may hinder technological progress and reduce long-term productivity gains, while income support mechanisms face challenges related to fiscal feasibility. In contrast, this paper proposes a novel strategy: a planned reduction of statutory working hours. This approach does not impede AI research and development but instead uses market mechanisms to redistribute work among a larger number of workers. By shortening the average working time, firms are incentivized to retain or even hire more workers to maintain output levels, thereby mitigating the displacement effect of AI. Furthermore, reducing working hours can enhance workers’ well-being by allowing more leisure time and reducing unemployment-related social costs. Therefore, in balancing efficiency and equity, this strategy offers distinct advantages over previous proposals.

Keywords: artificial intelligence; self-learning capability; employment; statutory working hour; task-based model

下载全文