Abstract
Decision-makers often process new evidence selectively, depending on their current beliefs about the world. We asked whether such confirmation biases result from biases in the encoding of sensory evidence in the brain, or alternatively in the utilization of encoded evidence for behavior. Human participants estimated the source of a sequence of visual-spatial evidence samples while we measured cortical population activity with magnetoencephalography. Halfway through the sequence, participants were prompted to judge the more likely source category. We find that processing of subsequent evidence depends on its consistency with the previously chosen category. Evidence encoded in parietal cortex contributes more to the estimation report when that evidence is consistent with the previous choice compared to when it contradicts that choice. Our results indicate that information contradicting pre-existing beliefs has little impact on subsequent behavior, despite being precisely encoded in the brain. This provides room for deliberative control to counteract confirmation biases. People often discard incoming information when it contradicts their pre-existing beliefs about the world. Here, the authors show that this discarded information is precisely encoded in the brain, but has little impact on subsequent behavior.
The Mechanism of Confirmation Bias
Confirmation bias—our tendency to favor information that fits our existing beliefs—is a hallmark of human cognition. However, the neural origins of this bias have remained a subject of debate. Does our brain simply fail to "see" contradictory evidence (selective encoding), or does it perceive the evidence accurately but choose to ignore it during decision-making (selective readout)?
Distinguishing between these two theories is vital for understanding human error. If the bias occurs during encoding, the contradictory information is essentially lost to the nervous system. If the bias occurs during readout, the information is still present in the brain, suggesting that we might be able to develop strategies or training to "listen" to this ignored data and make more objective decisions.
Experimental Design
To investigate this, researchers tasked 30 healthy human volunteers with estimating the average position of a sequence of 12 visual samples (checkered discs). The trial structure was designed to isolate the moment of commitment.
As shown in the figure above, participants viewed the first six samples and were then either asked to make a categorical "Choice" (is the source to the left or right?) or were given a "Cue." Following this, they viewed six more samples before providing a final estimate of the generative mean.
To track the flow of information, the team used Magnetoencephalography (MEG) and applied information-theoretic measures:
- Encoding ($I(S;R)$): Measuring how much information about the stimulus ($S$) is present in the neural response ($R$).
- Readout ($II(S;R;E)$): Measuring how much of that encoded stimulus information actually influences the final behavioral estimate ($E$).
Key Insights: The Parietal Filter
The study revealed that the brain actually "keeps the receipts" of contradictory evidence. The precision of sensory encoding ($I(S;R)$) was identical for evidence that supported the participant's choice and evidence that contradicted it. The brain's sensory systems recorded the truth regardless of the participant's bias.
However, a clear bias emerged in the behavioral reports. As illustrated in the figure above, participants consistently over-weighted samples that matched their mid-trial choice. The MEG data pinpointed this effect to a "selective readout" phase in the Dorsal Visual Cortex and Inferior Parietal Cortex. In these regions, the brain significantly reduced the weighting of contradictory samples when calculating the final estimate.
Interestingly, this bias was much stronger when participants made their own "Choice" compared to when they were simply given a "Cue." This suggests that the act of committing to a decision is a primary driver of the selective readout mechanism. The authors propose that confirmation bias happens because the "tuning" of neural weights shifts after a choice, creating a mismatch between the sensory representation and the decision-making readout.
Conclusion and Future Directions
The findings indicate that confirmation bias is a failure of utilization, not perception. Because the contradictory information is precisely encoded in the parietal cortex, it remains theoretically recoverable.
One limitation of the study is that the selective readout effect was most prominent for the samples immediately following the choice, suggesting the "filter" might be short-lived. Furthermore, the task involved value-neutral stimuli. Future research should investigate if these results hold for "hot" cognition—topics like politics or identity where confirmation bias is most entrenched. The preservation of encoded information offers hope that deliberative control or mindfulness could be used to force the brain to integrate the evidence it has already recorded but is currently choosing to ignore.
Summary updated