Introduction
Psychologists have long observed a curious feature of human cognition: our remarkable ability to maintain beliefs even when faced with direct evidence to the contrary. Traditionally, this "stubbornness" has been labeled as an irrational bias or a cognitive flaw, often referred to as motivated reasoning. However, Samuel Gershman’s research offers a different perspective rooted in the philosophy of science. By applying the Duhem-Quine thesis to cognitive science, Gershman argues that what looks like irrationality may actually be a byproduct of rational Bayesian inference.
The Bayesian Framework of Auxiliary Hypotheses
The core of the argument lies in the "credit assignment problem." When a prediction fails, a reasoner must decide which part of their mental model is at fault. No theory exists in a vacuum; every central hypothesis ($h$) is supported by a "protective belt" of auxiliary hypotheses ($a$)—secondary assumptions about the environment, the reliability of sensors, or the validity of the data source.
In Gershman's model, a rational agent updates their belief in a central hypothesis by marginalizing over these auxiliary assumptions. If a reasoner holds a strong prior belief in their theory, the most "logical" move when faced with an anomaly is not to discard the theory, but to shift the blame to an auxiliary hypothesis. For example, if a scientific experiment fails to produce the expected result, a researcher might conclude the equipment was faulty rather than the underlying physics being wrong.
Key Results: The Mathematical "Protective Belt"
Gershman provides a mathematical formalization of this robustness. He demonstrates that the decrease in belief for a central hypothesis is inversely proportional to the strength of its prior and directly proportional to the likelihood of an available auxiliary hypothesis that can "explain away" the data.
As shown in the figure above, when the prior belief $P(h)$ is high, the ratio of the posterior to the prior remains closer to 1, indicating that the belief is resistant to change. This effect is amplified by two properties of human intuitive theories: sparsity (the assumption that only one or two things are likely to go wrong at once) and determinism (the expectation that theories should yield near-certain predictions).
Unifying Psychological Phenomena
This Bayesian framework provides a unified explanation for a wide array of seemingly disparate behaviors:
- Social Polarization: Individuals with strong views on topics like the death penalty often discount contradictory studies by invoking auxiliary hypotheses about the study's methodology or the researchers' biases.
- The Good News-Bad News Effect: People update their self-image more readily after positive feedback than negative. If you believe you are highly capable, negative feedback is a low-probability anomaly that is easily attributed to an "unfair test" or "bad luck."
- Stereotyping and Subtyping: When encountering an individual who contradicts a stereotype, the brain often "subtypes" them as an outlier. This is a rational move to preserve the sparse prior of the stereotype while acknowledging the specific data point.
- Conspiracy Theories and Religion: These systems often employ a "spiraling proliferation" of auxiliaries (e.g., "the evidence was planted" or "God's ways are inscrutable") to ensure the central belief remains unfalsifiable.
Limitations and Future Directions
While the model is powerful, it faces the critique of being "too flexible." If one can always invent a new auxiliary hypothesis, any belief can be maintained indefinitely. Gershman argues that the model remains falsifiable if priors and likelihoods are measured independently of the behavior being explained.
Furthermore, the model currently assumes the "hypothesis space" of possible auxiliaries is pre-defined. Future research is needed to understand how humans generate new auxiliary hypotheses during a crisis of belief. For those looking to change minds, the paper suggests a strategic shift: rather than attacking a core belief directly, one must systematically chip away at the "protective belt" of auxiliary assumptions that keep that belief afloat.
Summary updated