AI Affect the State Transition of the Economic Society： A Markov Game Framework

Abstract: [Purpose] Drawing on the "task-based model" of Acemoglu and Restrepo and the research perspective of "AI economics" from Mei Guoping's team, this paper integrates the task-based model with a stochastic game framework, attempting to construct a "four-state, two-agent" dynamic stochastic game system based on the Markov Decision Process (MDP) to analyze the interaction mechanism between artificial intelligence (AI) technological progress and labor market structure. [Methods] Taking the margin of automation and the human capital structure as state variables, the task space is discretized—through prior approaches such as "event forecasting and causal inference"—into four stages: the "traditional economy period," the "transition pain period," the "human-machine collaboration period," and the "post-AI era." By jointly solving the one-period payoff matrix, the state transition matrix, and the Bellman equation, the strategic interaction between firms and workers is characterized as a Markov Decision Process (MDP). Furthermore, regulatory instruments such as "training subsidies, automation tax rates, and improvements in matching efficiency" are introduced to analyze their structural effects on the "payoff matrix" and the "state transition matrix," and theoretical construction is carried out based on "a stochastic game framework." [Results] In the AI era, the impact of AI and related technologies on labor market structure exhibits significant stage-dependent and nonlinear features, with multiple equilibria and path dependence. From the perspective of "event forecasting and causal inference": the social system may converge to a high-efficiency "human-machine collaboration" equilibrium; it may also fall into a "low-level trap" due to problems such as "technology–skill mismatch"; and extreme scenarios such as "system collapse" caused by "excessive automation" are also possible. Among these, the "displacement effect" is most pronounced during the "transition pain period," while the combined forces of the "productivity effect, the deepening-automation effect, and the recovery effect" determine the direction of state transitions. [Limitations] The discretization of the state space simplifies the continuous process of technological evolution; parameter calibration currently lacks support from micro-level empirical data; and the two-agent setup has not yet directly incorporated the government and financial institutions into the endogenous game framework. [Conclusion] Policy interventions require a coordinated mix of measures: during the transition pain period, effective policies such as "training subsidies plus improvements in matching efficiency" may be implemented; during the human-machine collaboration period, instruments such as "robot taxes and automation taxes" can be introduced to guard against excessive automation and avoid a Malthusian trap; and an economic early-warning mechanism based on "distribution ratios" may be established.