ECON30019 Chap.9 Confirmation Bias and Biased Belief Updating
Confirmation Bias and Biased Belief Updating
Bayesian updating carries an optimistic result: rational people fed the same evidence converge on one belief, whatever they held first, which the subject calls the washing out of the priors and describes as a hopeful picture of human nature. This chapter is about the mechanism that switches it off.
Confirmation bias is a tilt in how evidence is sought, handled, read and remembered, always toward what was already believed, and it operates through three separable channels that fail at different stages and call for different remedies.
The chapter covers falsification and the law of contraposition through the four-card selection task, the polarisation result in which shared evidence moved two groups further apart, and a formal model in which an otherwise ordinary Bayesian agent misreads disconfirming signals with some probability, with the consequence that error rates stop falling as evidence accumulates.
What this chapter covers
- 01
The washing out of priors, and why persistent disagreement then needs explaining
- 02
Confirmation bias defined, and the phenomena it is said to explain
- 03
Three channels: biased search, biased interpretation and biased memory
- 04
Why three channels means three different remedies
- 05
The law of contraposition, stated formally
- 06
The four-card selection task, and the two informative cards
- 07
Why the confirming card proves nothing
- 08
Attitude polarisation: shared evidence moving two groups apart
- 09
A formal model with a signal-accuracy parameter and a misreading probability
- 10
Error rates that stop falling, and belief held with near certainty
Which observations would refute the claim?
- 1Write the claim as an implication: over budget implies no scope document.
- 1Open the antecedent-true file, the over-budget project: a scope document inside it refutes the claim.
- 1Open the consequent-false file, the project with a scope document: by contraposition, having a document implies not over budget, so finding it over budget refutes the claim.
- 1Reject the other two and diagnose: the on-budget file is antecedent-false so the claim makes no assertion about it, and the file with no document is consequent-true and can only ever agree. Most people would open the over-budget file and that second one, which is the confirming pair.
Key terms
- Washing out of priors
- The result that rational people exposed to the same evidence converge on the same belief regardless of their starting points, which is the benchmark confirmation bias breaks.
- Confirmation bias
- A tilt in how evidence is sought, handled, read and remembered, always toward what was already believed, operating through three separable channels.
- Biased search
- Testing hypotheses in a one-sided way by looking for evidence consistent with the current hypothesis, so the evidence set is already selected before any reading happens.
- Biased interpretation
- Assigning different evidential weight to the same observation depending on which hypothesis it favours, which is the channel that produces polarisation.
- Law of contraposition
- The equivalence between a statement and its contrapositive, which identifies the second observation capable of refuting a conditional rule.
- Signal accuracy
- The probability that an informative signal is correct, lying strictly between one half and one. It is a property of the information source rather than of the agent.
Confirmation Bias and Biased Belief Updating FAQ
Why is polarisation more damaging than simply ignoring evidence?
Because both groups updated. Ignoring evidence would only be inertia. In the study, participants on both sides of a contested question read the same mixed material and each side moved further toward its prior position, so shared evidence increased disagreement. That is not something Bayes' rule permits, since it requires the same observation to carry the same likelihood for everyone.
What does the misreading parameter do in the model?
It sets how often a signal favouring the other hypothesis is read as supporting the existing belief. At zero the agent is an ordinary Bayesian and beliefs converge on the truth. At one every signal is read as confirming, so the belief is entirely determined by its starting point. In between, error rates rise and additional observations correct less and less.
What is the most alarming conclusion of that model?
That an agent can end up almost sure of something false no matter how much evidence keeps arriving. The washing out of priors is not merely slowed by the bias; for a sufficiently biased agent it stops working altogether, which is why more data is not always a remedy.
Exam move
Keep the three channels separate in your notes, because questions here are usually classification questions in disguise. Biased search is about the evidence set, biased interpretation about the likelihood attached to a shared observation, and biased memory about the record consulted later. Practise the selection task on rules from unrelated domains until the antecedent-true and consequent-false answer is automatic.
For the model, be able to say in one sentence what each of the two parameters is and what happens to error rates as each moves, since the in-lecture quiz questions on it are exactly that.
Working through Confirmation Bias and Biased Belief Updating in ECON30019? Sia is AskSia’s AI Economics tutor — ask any ECON30019 Confirmation Bias and Biased Belief Updating question and get a clear, step-by-step explanation grounded in how ECON30019 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
