Globalization and the rise and fall of cognitive control

The paper investigates how the unprecedented scale of modern interconnectedness influences the evolution of cognitive control versus automaticity. Integrating perspectives from cognitive psychology (automatic vs controlled processing) and evolutionary game theory, the authors pose the research question: how do local versus global interaction structures—via learning, social contact, and environmental externalities—shape the prevalence and dynamics of controlled (deliberative) versus automatic (habitual) cognition in populations? The study motivates this question by noting societal trends such as misuse of technological innovations and backlash against expertise, and by emphasizing the efficiency–flexibility trade-off in cognition where controlled processing yields better context-sensitive decisions but incurs costs. The purpose is to formalize and analyze how different forms of globalization alter selection pressures on cognitive strategies, including feedbacks between cognition and the environment.

The work builds on dual-process theories in psychology and neuroscience distinguishing automatic and controlled processing (e.g., Shiffrin & Schneider; Kahneman; Evans & Stanovich; Cohen et al.) and on the expected value of control framework (Shenhav, Botvinick & Cohen). It connects to evolutionary game theoretic models of cognition and cooperation showing evolution of intuitive cooperation versus deliberative self-interest, and prior models demonstrating cognition–environment feedback leading to cycles in automatic vs controlled processing (Tomlin et al.; Toupo et al.; Rand et al.). It also draws from extensive research on spatial and networked evolutionary dynamics demonstrating how population structure affects strategic evolution (Nowak & May; Szabó & Fath; Nowak, Tarnita & Antal; Watts & Strogatz), and from evidence linking automaticity to short-termism (e.g., temporal discounting, addiction, poor investment decisions, preference for simple information). The paper situates its contribution by adding population structure and multiple globalization dimensions to cognition–environment evolutionary models.

Agent-based evolutionary game-theoretic simulations were conducted on networks (primary: ring with N = 100 agents and degree k = 2; robustness checks on larger N and small-world networks). Each agent i has a strategy x_i ∈ [0,1], the probability of using controlled processing; automatic processing occurs with probability 1 − x_i. Per time step t, agent i’s fitness is: fi,t = x_i,t π^c_i,t + (1 − x_i,t) π^a_i,t, with π^c_i,t = 1 − c + w (1 − ⟨x⟩_t) and π^a_i,t = 1 − p_t in the basic formulation. Here c is the fixed cost of control; w parameterizes how contact with automatic agents affects controlled agents (w < 0 harms, w > 0 helps); ⟨x⟩ is the relevant average control level (local or global depending on scenario); and p_t is the environment-dependent advantage of control, with 0 < p_t < 1. Cognition–environment feedback: In the baseline (no lag) model, p_t = 1 − ⟨x⟩_{t−1} so greater population control reduces the relative advantage of control by making the environment more hospitable. A lagged feedback version updates p via relaxation dynamics: • Local environment: each agent i has p_i updated by p_{i,t} = p_{i,t−1} + ((1 − ⟨x⟩_{i,t−1}) − p_{i,t−1}) / τ_ρ, where ⟨x⟩_{i} is the local neighborhood average including i. • Global environment: a single p_t updated by p_t = p_{t−1} + ((1 − ⟨x⟩_{t−1}) − p_{t−1}) / τ_ρ. Evolutionary updating: Death–birth Moran process with exponential fitness. Each generation, a learner L is randomly selected. With probability 1 − μ, L copies a teacher T chosen from a specified interaction set with probability proportional to exp(s f_T). With probability μ, mutation assigns L a new x sampled uniformly from [0,1]. Parameters: selection intensity s = 10, mutation rate μ = 0.01. Initial condition: x = 0.01 for all agents. Population structure and globalization dimensions: The model orthogonally varies local vs global along three dimensions to yield 2^3 = 8 scenarios: 1) Learning: teacher set local (neighbors on ring) vs global (whole population). 2) Contact: the fraction of controlled agents used in π^c is local (neighbors) vs global (population-wide) when computing the term w(1 − ⟨x⟩). 3) Environment: p is local (agent-specific p_i based on local ⟨x⟩) vs global (single p based on population ⟨x⟩). Simulation protocol: For each c and w grid, simulations ran 8 × 10^4 generations; averages were computed over the final steady state and over 10 replicates. For time-lag analyses of oscillation magnitudes, runs lasted 1.2 × 10^6 generations across τ_ρ values. Assortment (mesoscale community structure) was quantified as the difference between (i) the fraction of an agent’s neighbors sharing its strategy and (ii) the fraction in the whole population (excluding the agent) sharing its strategy. Variance of x across agents was computed to characterize heterogeneity. Robustness checks included larger networks and small-world topologies (varying rewiring). All simulations were executed in parallel on MIT Sloan’s Engaging cluster, with code and data available at https://osf.io/fy94w/.

- Equilibrium control levels: • Global environment consistently reduces the average level of controlled processing by enabling automatic agents to benefit from environmental improvements produced by controlled agents. • Global contact has sign-dependent effects: when w < 0 (automatic agents harm controlled agents), globalization of contact decreases control; when w > 0 (automatic agents benefit controlled agents), global contact can increase control (though substantial automaticity remains). • Global learning homogenizes strategies and undermines clustering, diminishing the effects of local contact/environment. • Across scenarios (with w < 0), the fully local condition yields substantially higher average control than any globalization scenario. - Mesoscale communities and assortment: • Global learning eliminates mesoscale communities (assortment a = 0). • With local learning, substantial assortment emerges. When environment is local: a ≈ 0.65 (for both local and global contact). When environment is global: assortment is higher—approximately a ≈ 0.70 (local contact) and a ≈ 0.74 (global contact). - Strategy variance (heterogeneity): • Local environment yields low cross-sectional variance in x (0.000–0.008 depending on contact/learning). • Global environment increases variance: with global contact, variance ≈ 0.052 (local learning) and 0.080 (global learning); with local contact, variance is highest ≈ 0.171 (local learning) and 0.181 (global learning), reflecting coexistence of near-pure automatic and near-pure controlled agents. - Dynamics with time-lagged environment (τ_ρ > 1): • Increasing τ_ρ increases oscillation amplitudes in average control across globalization conditions. • Under global learning, populations exhibit strong synchronized cycles, swinging between near-all controlled and near-all automatic. • Local learning requires larger τ_ρ to produce oscillations of comparable magnitude; local mesoscale communities desynchronize phases and slow diffusion, suppressing aggregate oscillation amplitude. - Robustness: Results qualitatively hold for different network sizes and small-world topologies. - Initialization and convergence: Starting from x ≈ 0.01, systems reach steady states with minimal replicate variability in average x over long horizons (8 × 10^4 generations).

The findings demonstrate that the scale and pathway of interaction profoundly shape the evolution of cognitive control. Global learning homogenizes the population, dissolving mesoscale clusters and phase-locking agents, which promotes population-wide temporal fluctuations in control levels. Global contact’s effect hinges on the strategic externality parameter w: if automaticity harms control (w < 0), broader contact reduces control; if automaticity benefits control (w > 0), broader contact can bolster control. Global environment uniformly undermines control by creating global public goods from control-driven innovations that automatic agents can free-ride upon. Counterintuitively, global environment promotes spatial clustering and strategic diversity, while local environment induces homogenization due to local anti-coordination incentives created by cognition–environment feedback. With lagged feedback, cyclical dynamics in control are robust in both local and global settings; however, local learning dampens aggregate cycles due to desynchronization and slower spread of strategies. Together, these results address the research question by clarifying how distinct forms of globalization—learning, contact, and environmental coupling—differentially influence both equilibrium levels and spatiotemporal patterns of cognitive strategies, offering insight into modern societal dynamics such as the vulnerability of control-dependent systems to global free-riding and synchronized global downturns in control.

This work integrates cognitive psychology with evolutionary game theory and network structure to analyze how globalization affects the evolution of cognitive control. By introducing population structure and distinguishing globalization in learning, contact, and environment, the study shows: (i) global environmental coupling consistently depresses control and amplifies heterogeneity; (ii) global learning homogenizes and synchronizes populations, fostering large-scale temporal oscillations; (iii) global contact can either hinder or help control depending on whether automaticity harms or benefits controlled agents. It also extends prior cognition–environment models by demonstrating robust cyclical dynamics with feedback lags in both local and global contexts, and by identifying how local learning attenuates aggregate cycles. Policy-relevant implications include promoting interactions where control benefits from contact with automaticity (e.g., market structures that reward control-generated innovations), addressing global environmental free-riding (e.g., climate policy), and recognizing risks of global synchronization leading to worldwide collapses in control. Future research should test these conceptual predictions via laboratory experiments and empirical analyses, explore richer network structures, model additional cognitive dimensions, and assess intervention strategies that stabilize control.

The model is intentionally abstract and simplified, offering qualitative, conceptual predictions rather than quantitative forecasts. It reduces cognition to a single control–automaticity parameter, assumes specific payoff structures and simple environmental feedback dynamics, and focuses on stylized network topologies. Results may be sensitive to parameter choices (e.g., costs, interaction weights, mutation) and do not incorporate heterogeneous payoff structures, multi-layered networks, or empirical calibration. Empirical validation through experiments and observational data is needed to assess generalizability and to refine assumptions.