The Mathematical Logic of Binary Social Conflict

Active inference offers a powerful lens for understanding social cognition. When an agent's environment is populated by other agents running similar generative models, a shortcut becomes available: to predict another's behavior, simply ask what prior intentions would have produced that same action in oneself. This projection-based theory of mind is computationally cheap and socially potent — it seeds the convergence of beliefs, norms, and ontologies within a group as agents collectively minimize joint free energy.

This convergence can be formalized using Markov blankets drawn around social collectives. Within the blanket, shared inference drives ideological and linguistic Alignment; the group becomes an epistemic unit. Yet Markov blankets are never fully impermeable. A tension persists between ingroup coherence — which reduces Prediction error internally — and outgroup engagement, which supplies novel signals that prevent the system from collapsing into equilibrium. A fully sealed group is an echo chamber: it reaches stasis.

When this dynamic is modeled computationally, one result stands out. The only non-equilibrium stable state is a near-equal division between competing groups. A minority group below a critical threshold gets absorbed; a group that achieves total dominance loses the epistemic friction needed to sustain Active inference at the population level. This predicts — and explains — the persistent 50/50 pattern observed across Brexit, US presidential elections, and independence referenda. These are not statistical accidents; they are attractors of a system governed by collective belief-updating under social Markov blanket dynamics.