Data Sharing with Large Language Models: Prisoner's Dilemma, Pareto Inefficiency, and Institutional Design

Abstract

This article analyzes organizational decision-making about data sharing in the context of large language models (LLMs) through the combined lenses of game theory, Pareto efficiency, and institutional design. It argues that the central problem is not the absence of Nash equilibrium, but the existence of a stable yet Pareto-inferior equilibrium in which widespread non-sharing remains individually rational while generating collective epistemic and innovative loss. To make this claim more precise, the article develops a formal model of the data-sharing dilemma using parameterized payoffs for private benefit, collective innovation gain, sharing cost, appropriative advantage, systemic loss from mutual withholding, and the effects of technical and institutional safeguards. On that basis, it shows how legal, technical, and organizational interventions may reparameterize the game so that limited, auditable, and protected cooperation becomes strategically attractive. The article further examines the paradoxes generated by both sharing and non-sharing paradigms, arguing that neither openness nor restriction can serve as a sufficient normative solution in isolation. A proof-of-concept evolutionary population simulation further illustrates that stronger governance regimes can increase cooperation, improve aggregate welfare, and reduce payoff dispersion, while also showing that cost-reducing safeguards may exert a stronger effect than trust alone. The article concludes by proposing a staged trust architecture for institutional cooperation and by outlining broader comparative and experimental paths for empirical validation. The broader claim is that the future of organizational data sharing around LLMs will depend less on moral appeals to openness than on the creation of governance structures that make cooperation rational, safe, and legitimate.

Keywords