# Social Science Applications of the Free Energy Principle ## Overview The Free Energy Principle extends naturally from individual cognition to social systems. Just as an individual brain minimizes variational free energy under a generative model, social groups, institutions, and cultures can be understood as **collective systems that minimize free energy through shared generative models**. This perspective -- developed primarily by Ramstead, Constant, Veissiere, and colleagues -- provides a unified framework for understanding: - **Cultural evolution** as the optimization of shared generative models - **Social norms** as collective priors that coordinate inference - **Communication** as free energy minimization through alignment of generative models - **Institutions** as formalized Markov blankets that structure collective inference - **Polarization and echo chambers** as precision dynamics in social networks - **Group formation** as the emergence of shared Markov blankets This framework -- sometimes called the **cultural active inference** or **variational niche construction** approach -- bridges cognitive science, anthropology, sociology, and political science under a common formal language. ## Theoretical Framework ### Shared Generative Models The core idea: social coordination requires that agents share (parts of) their generative models: ``` Individual generative model: m_i: Agent i's model of the world q_i(s): Agent i's beliefs about hidden states C_i: Agent i's prior preferences Shared generative model: m_shared: Common structure across agents q_shared(s): Coordinated beliefs about shared hidden states C_shared: Aligned prior preferences (social norms, values) Social coordination requires: D_KL[m_i || m_shared] < threshold for each agent i -> Agents' models must be sufficiently similar -> This similarity enables prediction of others' behavior -> Prediction enables coordination ``` ### Variational Niche Construction Ramstead et al. (2016) introduced **variational niche construction** -- the process by which agents actively structure their environment to make it easier to predict: ``` Standard niche construction (biology): Organisms modify their environment (e.g., beaver dams) Variational niche construction (FEP): Agents modify their environment to reduce free energy Two complementary processes: 1. Perceptual niche construction: Modify sensory access to environment -> Build tools that augment observation (microscopes, satellites) -> Create information systems (writing, libraries, internet) 2. Active niche construction: Modify the environment itself -> Build physical structures (houses, cities, infrastructure) -> Create social structures (laws, norms, institutions) -> Shape the cultural landscape (language, art, ritual) Result: Agents live in a world they have shaped to be predictable -> The "niche" is a structured environment that minimizes free energy -> Culture IS the variational niche of Homo sapiens ``` ### Cultural Affordances Affordances -- possibilities for action offered by the environment -- are formalized under the FEP as: ``` Affordance = policy pi such that G(pi) is low -> An action that the generative model predicts will reduce free energy Cultural affordance: = Affordance that exists because of cultural niche construction -> A door is a cultural affordance (built to afford passage) -> A word is a cultural affordance (created to afford communication) -> A social norm is a cultural affordance (established to afford coordination) Cultural affordances exist in the SHARED generative model: For agent i to perceive a cultural affordance: 1. m_i must include the cultural artifact in its generative model 2. m_i must represent the affordance (know what the artifact is for) 3. m_i must share this representation with other agents (shared model) ``` ## Social Norms as Collective Priors ### Norms as Shared Prior Preferences Social norms are formalized as **shared prior preferences** in the collective generative model: ``` Social norm: C_norm = p_shared(o) = shared preference for certain observations Examples: Turn-taking in conversation: C = p(speaking | other_speaking) is low Queuing: C = p(position_in_queue | arrival_time) = orderly Reciprocity: C = p(reciprocal_action | received_favor) is high Fairness: C = p(equal_distribution | resource_available) is high Norm compliance as free energy minimization: Violating a norm = generating observations inconsistent with C_norm -> High free energy for the individual (if they share the norm) -> High free energy for observers (norm violation is surprising) -> Social sanctions = actions by others to restore norm-consistent observations ``` ### Norm Internalization ``` Norm internalization = incorporating social priors into personal model: Stage 1: External norm C_social exists in others' models but not in agent's model Agent complies due to expected social sanctions -> Pragmatic value of compliance exceeds cost Stage 2: Partially internalized Agent begins to expect norm-consistent behavior from self C_agent partially aligns with C_social -> Norm violation generates internal prediction error (guilt, shame) Stage 3: Fully internalized C_agent = C_social for this norm Agent experiences norm-consistent behavior as natural/obvious -> Norm violations by others generate strong prediction errors (moral outrage) Formal: Internalization = D_KL[C_agent || C_social] -> 0 over time Through repeated social interaction and precision weighting ``` ### Norm Enforcement as Active Inference ``` When agent i observes agent j violating a norm: Prediction error: epsilon = o_j - g(mu_norm) [unexpected behavior] Three responses (all minimize free energy): 1. Update beliefs: "Maybe the norm doesn't apply here" -> Reduces error by changing the model 2. Perceptual inference: "I must have misperceived" -> Reduces error by reinterpreting the observation 3. Active inference: Punish/sanction agent j -> Reduces error by changing agent j's behavior -> "Making the world conform to my predictions" Which response depends on precision: High norm precision: Active inference (punish violator) Low norm precision: Perceptual inference (ignore/reinterpret) Model uncertainty: Belief update (revise the norm) ``` ## Communication as Free Energy Minimization ### Language as Shared Generative Model ``` Communication under FEP: Speaker: Selects utterance to minimize listener's free energy Listener: Updates beliefs to minimize free energy given utterance Successful communication: D_KL[q_speaker(s) || q_listener(s)] decreases after utterance -> Speaker and listener's beliefs converge -> Shared understanding = aligned generative models Failed communication: D_KL[q_speaker(s) || q_listener(s)] remains high or increases -> Models too different to align through available channel -> Misunderstanding = persistent divergence of generative models ``` ### Pragmatics as Precision Optimization ``` Pragmatic communication: Speaker selects utterance u to minimize: F_comm(u) = -Information_conveyed(u) + Cost(u) Where: Information conveyed = E[D_KL[q_listener(s|u) || q_listener(s)]] -> How much the listener's beliefs change Cost = length/complexity of utterance -> Grice's maxim of quantity: Be as informative as needed, no more Precision in communication: - Precise language: High information, high cost -> formal/technical contexts - Vague language: Low information, low cost -> casual/social contexts - Context: Shared generative model reduces need for explicit information -> "You know what I mean" = relying on shared prior ``` ### Narrative and Storytelling ``` Narratives as generative model transmission: Stories = compressed generative models about how the world works A narrative provides: 1. Generative model structure: Causal relationships, character types 2. Prior preferences: What outcomes are valued/feared 3. Policy templates: What actions lead to what outcomes 4. Precision calibration: How confident to be about these patterns Cultural transmission: Myths, legends, parables = time-tested generative models -> Transmitted across generations via narrative -> Each retelling = approximate inference over the model -> Culturally successful narratives = models that minimize free energy for many agents across many contexts ``` ## Group Formation and Social Identity ### Groups as Markov Blankets ``` Group formation under FEP: A group exists when agents form a SHARED Markov blanket: Individual blanket: (s_i, a_i) = agent i's sensory and active states Group blanket: (S_group, A_group) = collective sensory and active states Group internal states: Shared beliefs, coordinated activities Group external states: Other groups, environment Group blanket states: Sensory: Intelligence gathering, monitoring, communication with outside Active: Collective action, group representation, boundary maintenance Group persists when: F_group = collective free energy is lower than sum of individual free energies -> Being in the group is better than being alone -> The shared generative model provides better predictions ``` ### Social Identity as Generative Model Membership ``` Social identity = membership in a shared generative model: "I am an X" means: 1. My generative model m_i includes the group model m_X 2. My prior preferences C_i align with group preferences C_X 3. I predict my own behavior using the group's policy prior pi_X 4. I expect group members to share my model (mutual prediction) Identity strength = precision of group model components: Strong identity: Pi_group is high -> Group predictions heavily weight inference -> Group norms feel natural and obvious -> Out-group behavior feels surprising and wrong Weak identity: Pi_group is low -> Group model has little influence -> Flexible, context-dependent identification ``` ### Intergroup Conflict as Model Competition ``` Intergroup dynamics: Group A has generative model m_A Group B has generative model m_B Conflict arises when: 1. m_A and m_B make incompatible predictions about shared observations 2. Both groups have high precision on their models 3. Shared environment means both groups observe the same events Each group's observations of the other generate prediction errors: epsilon_A = observations_of_B - predictions_from_m_A epsilon_B = observations_of_A - predictions_from_m_B Resolution strategies: 1. Model updating: One group adopts the other's model (assimilation) 2. Active inference: One group forces the other to change (coercion) 3. Niche construction: Groups separate into different environments (segregation) 4. Model integration: Groups develop a shared higher-order model (pluralism) ``` ## Polarization and Echo Chambers ### Precision Dynamics in Social Networks ``` Polarization as precision amplification: Initial state: Agents have moderate beliefs with moderate precision Echo chamber formation: 1. Agent connects to like-minded others (homophily) 2. Shared beliefs are confirmed -> precision increases 3. Counter-evidence from out-group is low-precision (not trusted) 4. Precision on in-group model increases further 5. Out-group behavior becomes increasingly surprising (high prediction error) 6. High prediction error + high in-group precision -> active inference: -> Attempt to change out-group (persuasion, coercion) -> Or strengthen in-group boundary (us vs. them) Formal: Pi_in-group(t+1) = Pi_in-group(t) + alpha * agreement_rate Pi_out-group(t+1) = Pi_out-group(t) - beta * disagreement_rate When Pi_in-group >> Pi_out-group: -> Complete echo chamber -> Out-group evidence cannot update beliefs -> Only in-group evidence has precision to drive inference ``` ### Social Media as Precision Amplifier ``` Social media effects under FEP: 1. Algorithmic curation = precision optimization for engagement -> Show content with high precision-weighted prediction error -> Maximize surprise (emotional/outrage content) 2. Selective exposure = niche construction of information environment -> Users curate feeds to confirm generative model -> Reduce free energy by avoiding challenging content 3. Viral content = high-precision prediction errors that spread -> Content that generates strong prediction errors in many agents -> Spreads because it demands model updating or active inference (sharing) 4. Filter bubbles = precision-insulated generative models -> Within-bubble precision is very high -> Cross-bubble precision is very low -> Communication across bubbles fails (no shared model) ``` ### Depolarization Strategies ``` FEP-informed depolarization: 1. Increase precision of out-group signals: -> Facilitate direct contact (contact hypothesis) -> Humanize out-group members (reduce prediction error) -> Use trusted intermediaries (bridge precision) 2. Reduce precision of in-group model: -> Introduce within-group disagreement (model uncertainty) -> Highlight complexity and ambiguity (reduce certainty) -> Encourage epistemic humility (meta-cognitive precision reduction) 3. Create shared higher-order model: -> Identify superordinate goals (shared prior preferences) -> Construct common narrative (shared generative model) -> Build institutions that enforce cross-group precision ``` ## Institutional Design ### Institutions as Formalized Generative Models ``` Institution = formalized shared generative model with enforcement: Components: 1. Rules = explicit priors and preferences (what is expected) 2. Roles = positions in the generative model (who does what) 3. Procedures = policy templates (how to act in given situations) 4. Sanctions = precision enforcement (consequences for violations) 5. Communication channels = sensory/active states of institutional blanket Well-functioning institution: F_institution < sum_i F_individual -> Institution reduces collective free energy -> By providing shared model, reducing coordination costs -> By establishing predictable behavior, reducing uncertainty Dysfunctional institution: F_institution > sum_i F_individual -> Institution increases collective free energy -> Model no longer fits the environment -> Prediction errors accumulate faster than they can be resolved ``` ### Democratic Institutions as Collective Inference ``` Democracy under FEP: Democratic process = collective Bayesian inference Voting = aggregating beliefs across agents: q_collective(policy) = aggregate of q_i(policy) across citizens Deliberation = message passing between agents: Citizens share prediction errors and models -> Collective model is updated through argument -> Ideally converges to model that minimizes collective free energy Elections = model selection: Choose between competing generative models (political platforms) m* = argmax_m F(m | collective_observations) -> The model that best explains collective experience wins Failure modes: - Polarization: No shared model for aggregation - Manipulation: Distortion of precision (propaganda) - Capture: Institutional model serves subset, not collective ``` ### Legal Systems as Norm Formalization ``` Law = formalized social norms with explicit precision: Statute: C_law = specific prior preferences (what is acceptable) Penalty: Pi_law = precision of enforcement (how strongly enforced) Due process: Inference procedure for determining violations Precedent: Accumulated evidence for model interpretation Legal reasoning as Bayesian inference: p(guilty | evidence, law) = p(evidence | guilty, law) * p(guilty) / p(evidence) "Beyond reasonable doubt" = posterior precision threshold "Preponderance of evidence" = lower precision threshold (civil law) ``` ## Cultural Evolution as Model Optimization ### Cultural Transmission as Approximate Inference ``` Cultural transmission: Generation n has generative model m_n Generation n+1 learns m_{n+1} from observation of generation n m_{n+1} = argmin_m F[m | observations_of_generation_n] Cultural evolution = successive approximate inference: Each generation's model is an approximation of the previous -> Models gradually improve (reduce free energy over generations) -> Models also drift (each approximation introduces variation) -> Selection: Models that better predict the environment spread -> Variation: Approximation errors and innovation introduce novelty ``` ### Cultural Attractors ``` Cultural attractors (Sperber) under FEP: = Regions of model space where free energy is locally minimal Cultural practices converge toward attractors: - Music: Certain scales and rhythms are attractors (they minimize auditory prediction error across most listeners) - Language: Certain phonological patterns are attractors (easy to produce and distinguish) - Social norms: Certain cooperative arrangements are attractors (they minimize collective free energy) Cultural diversity arises from: - Multiple local minima (different attractors in different contexts) - Historical contingency (which attractor basin the culture entered) - Environmental variation (different environments -> different optimal models) ``` ### Religion and Ritual ``` Religion under FEP: Religious practice = technology for managing free energy 1. Cosmological models: Reduce existential uncertainty -> Generative model of the universe's purpose/meaning -> Reduces free energy from unpredictable, threatening events 2. Moral frameworks: Provide shared prior preferences -> Clear expectations for behavior -> Reduce social uncertainty 3. Rituals: Synchronize generative models across individuals -> Shared actions create shared prediction errors -> Coordinated precision dynamics (altered states) -> Strengthen group Markov blanket (sense of belonging) 4. Afterlife beliefs: Extend temporal horizon of generative model -> Death becomes a transition, not an end -> Reduces free energy associated with mortality salience ``` ## Current Research ### Computational Social Science ``` Agent-based models with active inference agents: - Simulate cultural evolution with principled agents - Model norm emergence, spread, and change - Study polarization dynamics quantitatively - Design interventions and test in silico ``` ### Cross-Cultural Computational Psychiatry ``` Connecting cultural models to mental health: - Cultural priors shape what counts as "aberrant" precision - Same computational profile may be pathological in one culture, adaptive in another - Cultural affordances shape available policies for free energy reduction See [[psychiatry]] for clinical details ``` ### Digital Societies ``` Applying FEP to online communities: - Platform design as niche construction - Algorithmic recommendation as precision manipulation - Online identity as virtual Markov blanket - Misinformation as adversarial generative model injection ``` ## Open Questions 1. **Measurement**: How do we empirically measure shared generative models? What are the observable signatures of collective free energy minimization? 2. **Scale**: How does the FEP apply at different social scales (dyad, group, institution, civilization)? Are the same principles at work, or do qualitative transitions occur? 3. **Power**: How do asymmetries of power map onto the FEP framework? Can we formalize exploitation, domination, and liberation in free energy terms? 4. **Change**: How do social systems undergo phase transitions (revolutions, reformations)? When does incremental model updating fail and radical model replacement become necessary? 5. **Normativity**: Does the FEP provide any basis for normative social theory? Can we say some social arrangements are "better" (lower free energy) than others? 6. **Methodology**: How should social scientists use the FEP? As metaphor, as formal model, or as empirical framework? Different disciplines may benefit from different levels of formalization. ## References 1. Ramstead, M. J. D., Veissiere, S. P. L., & Kirmayer, L. J. (2016). Cultural affordances: scaffolding local worlds through shared intentionality and regimes of attention. *Frontiers in Psychology*, 7, 1090. 2. Constant, A., Ramstead, M. J. D., Veissiere, S. P. L., Campbell, J. O., & Friston, K. J. (2018). A variational approach to niche construction. *Journal of the Royal Society Interface*, 15(141), 20170685. 3. Veissiere, S. P. L., Constant, A., Ramstead, M. J. D., Friston, K. J., & Kirmayer, L. J. (2020). Thinking through other minds: a variational approach to cognition and culture. *Behavioral and Brain Sciences*, 43, e90. 4. Kirchhoff, M. D., Parr, T., Palacios, E., Friston, K. J., & Kiverstein, J. (2018). The Markov blankets of life: autonomy, active inference and the free energy principle. *Journal of the Royal Society Interface*, 15(138), 20170792. 5. Ramstead, M. J. D., Badcock, P. B., & Friston, K. J. (2018). Answering Schrodinger's question: a free-energy formulation. *Physics of Life Reviews*, 24, 1-16. 6. Constant, A., Clark, A., Kirchhoff, M., & Friston, K. J. (2022). Extended active inference: constructing predictive cognition beyond skulls. *Mind & Language*, 37(3), 373-394. 7. Sperber, D. (1996). *Explaining Culture: A Naturalistic Approach*. Blackwell. 8. Henrich, J. (2016). *The Secret of Our Success: How Culture Is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter*. Princeton University Press. 9. Hesp, C., Ramstead, M., Constant, A., Badcock, P., Kirchhoff, M., & Friston, K. (2019). A multi-scale view of the emergent complexity of life: a free-energy proposal. In *Evolution, Development and Complexity* (pp. 195-227). Springer. 10. Albarracin, M., Constant, A., Friston, K. J., & Ramstead, M. J. D. (2022). A variational approach to scripts. *Frontiers in Psychology*, 13, 1003718. ## See Also - [[knowledge_base/free_energy_principle/cognitive/social_cognition|Social Cognition]] - [[economics|Economic Applications]] - [[education|Educational Applications]] - [[knowledge_base/free_energy_principle/systems/self_organization|Self-Organization]] - [[knowledge_base/free_energy_principle/systems/emergence|Emergence]] - [[knowledge_base/free_energy_principle/biology/ecology|Ecology]] - [[knowledge_base/free_energy_principle/biology/evolution|Evolution]]