Why Polarized Societies Stabilize at Roughly 50-50 Splits

The insight begins with a recursive twist in the Free energy principle: when the thing being modeled is itself a modeler, qualitatively new dynamics emerge. If each agent minimizes surprise through self-evidencing, and the environment is populated by similar agents, the most efficient strategy is to assume others share your generative model. This grants a theory of mind essentially for free — no costly inference required. At the population level, this mutual assumption drives rapid convergence toward shared world models: common languages, ideologies, theologies, even conspiracy theories. One can draw a Markov blanket around the resulting Epistemic community.

But here lies the tension. If that blanket becomes impervious — if the community becomes informationally closed — the system reaches thermodynamic-style equilibrium. Beliefs ossify. Surprise is minimized to zero, and adaptive capacity vanishes. The blanket must remain porous to sustain non-equilibrium dynamics, meaning the community must remain open to information that challenges its shared model.

Simulations of multi-agent Active inference reveal a remarkable attractor in this landscape: the only dynamically stable non-equilibrium configuration is an approximately 50-50 split between in-group and out-group. Any significant asymmetry causes the smaller group to be absorbed, collapsing the system back to consensus equilibrium. The even split uniquely sustains both the free-energy-reducing comfort of in-group model-sharing and the Epistemic tension generated by an out-group whose models differ. This formal result maps strikingly onto empirical political phenomena — Brexit, Trump-Biden polarization, Scottish independence — suggesting these near-even divisions are not accidents of campaign strategy but expressions of a deep dynamical constraint on belief-forming populations.