Non-Additivity of Collective Free Energy: A Formal Foundation for Emergent Ethical Disposition in Multi-Agent Systems

How can an AGI system exhibit genuine concern for others as a structural feature of its objective—rather than a penalty bolted on from outside? The answer developed here runs through collective predictive coding (CPC), a variational framework for groups of agents whose generative models are coupled through a shared group-level variable w. The central result is a non-additivity theorem: when w is non-trivial, the collective free energy satisfies F_CPC ≠ Σ_k F_k. We prove this for the first time, giving a complete algebraic characterisation of when equality holds and when it fails. Non-additivity carries over directly to the action-selection objective G_CPC: the Collective Epistemic Value term is irreducible to any sum over individual objectives, so each agent's optimal policy depends structurally on the predicted states of all others. This constitutes a formal computational mechanism for what Sober and Wilson called psychological altruism—concern for others as a constitutive part of the objective, not an add-on. We assess, independently, the biological and computational conditions under which non-trivial w arises, grounding the theoretical result in realisable AGI architectures. The formal conditions established here—non-triviality of w as the key architectural lever—supply the mechanistic foundation that emerging research on co-creative machine ethics requires but has not yet formalised.