The Algorithmic Monoculture Paradox: Simulating the Collapse of Collective Intelligence in Foundation-Model-Mediated Social Networks

The “Wisdom of Crowds” principle demonstrates that distributed groups of imperfect decision-makers can achieve superior collective accuracy because independent, idiosyncratic errors cancel out upon aggregation. Today, the rapid integration of centralized foundation models (LLMs) into decision-making workflows threatens the epistemic independence that underlies this dynamic. When independent human agents query shared foundation models, individual error variance is suppressed, replacing independent noise with shared, correlated algorithmic bias. Here, we investigate this phenomenon by formalizing a discrete-event network simulation combining Scott Page’s Diversity Prediction Theorem with a hybrid DeGroot consensus framework across macro-scale social graphs (N = 1,000 agents). Across 2,970 Monte Carlo simulations spanning three canonical network topologies (Watts-Strogatz Small-World, Barabási-Albert Scale-Free, and Erdős-Rényi Random) and three correlation regimes, we show that high reliance on a centralized foundation model (ρ = 0.90) causes an 18- to 68-fold surge in Collective Squared Error (from 0.246 to 16.833), driven by the rapid collapse of cognitive diversity. Furthermore, we demonstrate that an oligopoly of competing foundation models (ρ = 0.50) fails to prevent this collapse (Collective Error surges to 10.314), whereas an unconstrained polyculture (ρ = 0.00) preserves collective wisdom (Collective Error bounded at 3.974). These findings prove that individual optimization under shared generative AI triggers systemic epistemic fragility, highlighting urgent implications for AI governance, platform auditing, and sociotechnical system design. Code and Replication Suite: https://github.com/Ashuraj-Herode22/algorithmic-monoculture-sim