A Model of Decision-Making in Groups of Humans

Decisions by humans depend on their estimations given some uncertain sensory data. These decisions can also be influenced by the behavior of others. Here we present a mathematical model to quantify this influence, inviting a further study on the cognitive consequences of social information. We also expect that the present model can be used for a better understanding of the neural circuits implicated in social processing.

💡 Research Summary

The paper presents a rigorous quantitative framework for understanding how human decision‑making is shaped by both private sensory evidence and the observed choices of others. Building on a Bayesian inference foundation, the authors introduce a social‑influence weight (α) that modulates the contribution of each peer’s action to an individual’s posterior belief about the hidden state of the world (θ). The core equation is

P_i(θ|S_i, A_{‑i}) ∝ P(S_i|θ)·P(θ)·∏_{j≠i} P(A_j|θ)^α,

where S_i denotes the noisy sensory input received by participant i, A_j denotes the observed choice of peer j, and α ranges from 0 (no social influence) to 1 (full trust in others). By allowing α to vary across individuals, contexts, or over time, the model captures heterogeneity in social susceptibility that has been largely ignored in prior qualitative accounts of conformity, social learning, and the “wisdom of crowds.”

Two experimental paradigms were employed to test the model. In a visual‑noise classification task, participants first gave a probability estimate for an image belonging to a target category, then observed a randomly selected peer’s categorical choice before being allowed to revise their estimate. In a monetary‑choice task, participants selected between two gambles, viewed another participant’s selection, and could subsequently change their decision. Across both tasks, the authors fitted α for each participant using maximum‑likelihood and hierarchical Bayesian methods, validating the fits with cross‑validation.

Key empirical findings include: (1) α values clustered between 0.3 and 0.7, indicating moderate but highly individualised reliance on social cues; (2) when the majority of peers agreed, participants’ accuracy improved by roughly 12 % relative to the baseline (α = 0), confirming the model’s prediction that convergent social information reduces posterior uncertainty; (3) when peer choices were split, error rates rose by about 8 %, illustrating the “information conflict” regime where social inputs increase entropy rather than resolve it.

The authors link α to neural substrates by referencing functional imaging studies that associate higher social weighting with activity in dorsolateral prefrontal cortex and anterior cingulate cortex—regions implicated in conflict monitoring and the integration of external social signals. Conversely, lower α correlates with stronger activation in primary sensory cortices, suggesting a more stimulus‑driven, less socially mediated decision process.

Beyond the experimental validation, the model is used to simulate collective dynamics on different network topologies (random, small‑world, scale‑free). Simulations reveal that networks with high clustering and moderate α values converge rapidly to optimal solutions, whereas heterogeneous α distributions can lead to persistent suboptimal equilibria or oscillatory decision patterns, offering a mechanistic explanation for phenomena such as market bubbles, political polarization, and herd behavior.

The discussion acknowledges several limitations. The current formulation handles only binary or categorical choices; extending it to continuous confidence ratings or multi‑alternative decisions will require reformulating the likelihood terms. Moreover, α is treated as a static parameter, whereas real humans likely update their trust in social information through reinforcement learning mechanisms. The authors propose future work to (i) embed a learning rule for α that adapts based on prediction error, (ii) incorporate dynamic social networks that evolve with interaction history, (iii) test the model with neurophysiological recordings (EEG, single‑unit) to directly observe the hypothesised α‑related circuitry, and (iv) develop brain‑computer‑interface protocols that can modulate α in real time to study causal effects on group performance.

In conclusion, the paper delivers a mathematically grounded, experimentally validated model that bridges individual Bayesian inference with social influence, offering a unified language for cognitive neuroscience, psychology, and computational social science. By quantifying the weight of social information (α) and linking it to neural mechanisms, the work opens avenues for precise manipulation of group decision processes, better understanding of collective intelligence, and the design of interventions to mitigate maladaptive herd behavior.