Media and responsible AI governance: a game-theoretic and LLM analysis

TL;DR

This work addresses how AI governance emerges from the strategic interactions of four actors—users, commentariat, developers, and regulators—under different regulatory regimes. It combines four-population evolutionary game theory with FAIRGAME-based LLM simulations to compare two roles for the media: investigating developers (soft regulation) or regulators. The finite- and infinite-population analyses reveal conditions under which safe AI development and trustworthy adoption arise, highlighting the pivotal role of objective media and the costs of high-quality commentary (e.g., $c_I$) in shaping outcomes. The LLM experiments largely corroborate the theoretical predictions, while also exposing model-specific differences, underscoring the media's potential to influence governance in regions lacking formal AI regulation. Overall, the paper argues that effective governance hinges on aligning incentives and reducing the cost of high-quality commentaries to harness media information as a regulatory substitute.

Abstract

This paper investigates the complex interplay between AI developers, regulators, users, and the media in fostering trustworthy AI systems. Using evolutionary game theory and large language models (LLMs), we model the strategic interactions among these actors under different regulatory regimes. The research explores two key mechanisms for achieving responsible governance, safe AI development and adoption of safe AI: incentivising effective regulation through media reporting, and conditioning user trust on commentariats' recommendation. The findings highlight the crucial role of the media in providing information to users, potentially acting as a form of "soft" regulation by investigating developers or regulators, as a substitute to institutional AI regulation (which is still absent in many regions). Both game-theoretic analysis and LLM-based simulations reveal conditions under which effective regulation and trustworthy AI development emerge, emphasising the importance of considering the influence of different regulatory regimes from an evolutionary game-theoretic perspective. The study concludes that effective governance requires managing incentives and costs for high quality commentaries.

Figures (14)

• Figure 1: Core features. The figure schematically illustrates the core features of the four-population model of AI governance. Users can either follow commentariat recommendations (Conditional Trust - CT) or not (N) (left). If they decide to follow the recommendations, their payoff depend on whether the commentariat invested in providing high quality information (investigate) or not (do not investigate) (lower middle). Commentariat can investigate either developers or regulators. Developers (upper middle) can either defect by creating unsafe AI products (D) or cooperate with regulations by creating safe ones (C), which entails additional costs. Regulators receive a benefit when users adopt AI systems. The regulator (right) can choose to cooperate by investing resources in regulating effectively (C) or not (D). If cooperating regulators catch unsafe developers, the latter are punished.
• Figure 2: Function of the Commentariat. The figure schematically illustrates what kind of information the two types of commentariat provides. The investigative (cooperating) will pay a cost to look for the hidden strategies of the developers, while the not investigative (defecting) will shun the cost and misclassify a developer with probability $p_w$.
• Figure 3: Numerical integration of the evolution equation for two models when the media investigation cost is low ($c_I=0.5$). The left column shows the results of Model I, and the right column shows the results of Model II. Parameters are set as $b_U = 4, b_P = 4, b_R = 4, c_P = 0.5, c_w = 1, u = 1.5, v = 0.5, b_{f_o} = 1, \epsilon = 0.2, p_w = 0.5$.
• Figure 4: Numerical integration of the evolution equation for two models when the media investigation cost is high ($c_I=5.0$). The left column shows the results of Model I, and the right column shows the results of Model II. Parameters are set as $b_U = 4, b_P = 4, b_R = 4, c_P = 0.5, c_w = 1, u = 1.5, v = 0.5, b_{f_o} = 1, \epsilon = 0.2, p_w = 0.5$.
• Figure 5: Hard regulation can be avoided in the presence of factual reporting about developers (Model I with a cooperative commentariat population). Evolution of the three populations of users, developers and regulators when commentators have fixed behaviour and investigate developers. Parameters set to: $b_U=b_R=b_P = 4$, $u=1.5$, $v=0.5$, $c_P=0.5$, $\beta=0.1$, $N_U=N_C=N_R=100$.

Theorems & Definitions (1)

• proof