OpenTox Virtual Conference 2026

Axiomatic Foundations for Adaptive Hybrid Human–AI Coalitions: An Information-Theoretic Framework for Dynamic Decision Authority

Salomé A. Sepúlveda-Fontaine and José M. Amigó *

Classical Game Theory and Expected Utility Theory rely on four core axioms: completeness, transitivity, independence, and continuity (the Archimedean axiom). Hybrid Human–AI coalitions challenge this axiomatic structure at the coalition level. Alternating execution authority can generate non-transitive revealed preferences and, most centrally, the Archimedean principle may fail because Human adjustment can be gradual while AI behavior may change through discrete optimization and hard constraints. Consequently, Human and AI may maintain substantially different state-conditioned policies, creating a mathematical gap for decision systems in which their policy distributions may diverge while execution authority changes over time. We present Neo-Game Theory, an extension of Classical Game Theory, as an axiomatic framework for hybrid Human–AI coalitions operating under Virtual Nature, the algorithmic counterpart of clas sical Nature that we define for digitally mediated environments. Virtual Nature captures algorithmic stochasticity, hard thresholds, and discontinuities generated by digital rules. The framework models the Human and the AI as distinct agents with state-conditioned policies and their own structural utility functions. At each decision step, both agents may propose an action, but only one action is executed. Policy discrepancy is quantified using the Jensen–Shannon divergence, a symmetric and bounded information-theoretic measure. Two thresholds partition the divergence space into agreement, contextual, and disagreement regions. These regions determine the delegation regime. The coalition utility is induced from the Human and AI utilities by the executed action after delegation rather than by blending their utilities. Regime selection therefore precedes utility evaluation, producing a lexicographic coalition structure. Unlike classical equilibrium concepts such as Nash, Stackelberg, or correlated equilibrium, which characterize fixed points of strategic best responses under static preferences, the equilibrium in Neo-Game Theory is dynamical. It is defined through the asymptotic behavior of the stochastic delegation-learning process generated by repeated Human–AI interaction under Virtual Nature. Specifically, the frequency-convergence equilibrium characterizes the limiting stabilization of the empirical occupation frequencies of the agreement, contextual, and disagreement regimes rather than an instantaneous best-response solution. This formal separation of agent policies, structural utilities, coalition utility, and execution authority provides an architecture for adaptive Human–AI decision-making in which authority responds dynamically to observed policy alignment.