Collaborative Supervision Mechanisms for Hallucination Issues in Generative Artificial Intelligence: An Evolutionary Game Perspective

The hallucination issues in generative artificial intelligence (GAI), stemming from training data bias and algorithmic misjudgment, constitute a significant source for information security problems. The resultant novel security threats, particularly the dissemination of false and biased information, pose a bottleneck constraining the development of GAI. Existing supervision mechanisms mostly focus on the binary game between the government and enterprises, neglecting the supervision value of users as direct risk-bearers. Based on the evolutionary game theory, this paper constructs a tripartite dynamic game model among enterprises, government and users, and analyzes the synergistic evolution mechanisms of responsible innovation in enterprises, user participation in supervision and the reward and punishment policy by government. Through the tripartite strategy matrix and replicator dynamics equation, we analyze the evolutionarily stable strategies and their influencing factors of each player, exploring the collaborative supervision mechanisms of GAI by numerical simulation. We find that: (1) There is an equilibrium point in the system, which is the evolutionarily stable strategies of responsible innovation in enterprises, user participation in supervision, and strict supervision by government. (2) The implementation of reward and punishment mechanisms by government for enterprises accelerates the realization of the evolutionarily stable strategies of the system. When the sum of government rewards obtained by enterprises and reputation premium brought by user feedback exceeds their additional costs, incentive measures such as subsidies or rewards adopted by the government are more effective in promoting enterprises to choose responsible innovation strategies, compared with simply intensifying punitive measures. (3) Designing an effective user evaluation feedback mechanism and increasing the rewards for user participation in supervision enhance the supervision willingness of users and accelerate the realization of evolutionarily stable states of collaborative supervision.