Belief polarization—where the same evidence causes people with opposing views to become even more extreme—is often cited as the "smoking gun" for human irrationality. When two people see the exact same data and move in opposite directions, we usually assume someone is being stubborn, biased, or cognitively limited. For example, if two managers see a mediocre test score for a job candidate, and the one who liked the candidate likes them more while the one who doubted them likes them less, it feels intuitively "wrong."
Psychological literature frequently labels this phenomenon as a cognitive failure. Previous attempts to explain it through a Bayesian lens usually relied on "bounded rationality"—the idea that humans have limited memory or processing power. However, this paper asks a deeper question: Can belief polarization emerge from "perfect" Bayesian reasoning? If so, under what structural conditions does it happen, and how common is it?
Modeling Belief with Bayesian Networks
To investigate this, the authors use Bayesian Networks (directed acyclic graphs) to model how people update their beliefs. They compare two hypothetical individuals, Alice and Bob, who agree on the "rules" of the world—meaning they agree on the structure of the causal graph and the conditional probabilities. However, they disagree on their initial "priors" (their starting beliefs about the root nodes).
The methodology hinges on several key steps:
- Defining "Contrary Updating": The authors distinguish between divergence (polarization, where beliefs move further apart) and convergence (where beliefs move closer together), contrasting both with standard parallel updating.
- Structural Analysis: They mathematically prove that a simple two-node model (Hypothesis $\rightarrow$ Data) can never produce polarization.
- Family Classification: The researchers analyzed all possible three-node network structures. They partitioned these into "Family 1" (which only allows parallel updating) and "Family 2" (which allows both parallel and contrary updating).
To determine if polarization is merely a mathematical fluke, the authors ran extensive synthetic simulations. They conducted 20,000 trials per model, randomly generating Conditional Probability Distributions (CPDs) from both biased (Beta) and uniform distributions to test the frequency of contrary updating. They also utilized qualitative case studies with hand-coded probabilities to illustrate how specific causal structures trigger rational divergence.
Key Insights: The Latent Variable "Lens"
The most critical insight from the paper is that polarization requires at least one additional variable ($V$) beyond the hypothesis ($H$) and the data ($D$). This variable acts as a "lens." If Alice and Bob have different priors about $V$, they will rationally interpret the exact same data $D$ as evidence for entirely different states of $H$.
Consider a "Promotion" scenario: if Alice believes a test is scored out of 60 and Bob believes it is a percentage out of 100, a score of "50" means two completely different things. Both individuals are updating their beliefs rationally given their background knowledge.
Furthermore, the simulations reveal that polarization is "generic" rather than a rare edge case.
- Frequency of Polarization: In Family 2 networks (the structures capable of polarization), contrary updating occurred in 9.6% to 18.2% of trials for binary variables.
- Complexity Increases Divergence: When the latent variable $V$ was increased from a 2-valued to a 4-valued node, the frequency of polarization jumped significantly, occurring in up to 23.3% of biased distribution simulations.
This suggests that as causal models become more complex, polarization becomes a natural feature of rational inference, not a bug.
Rethinking Rationality
The authors argue that we should not rely solely on "common sense" to judge rationality. Instead, Bayesian inference provides a precise normative standard demonstrating that many seemingly "irrational" behaviors are actually optimal given a person's subjective model of the world.
While the study formally proves that polarization can be rational, it does have limitations. The research focuses on small, three-node networks, whereas real-world belief systems involve hundreds of interconnected variables. It remains to be seen how the "density" of polarization scales in massive networks. Additionally, proving that polarization can be mathematically rational does not inherently prove that humans are acting rationally when they polarize in practice. Future work will need to fit these specific Bayesian models to experimental data from human subjects and map the exact ranges of probabilities that guarantee divergence.
Summary updated