The Computational Underpinnings of Learning from False Information

Razi, Hamid (2025) The Computational Underpinnings of Learning from False Information. Doctoral thesis, University of East Anglia.

Preview

PDF Download (5MB) | Preview

Abstract

Misinformation poses a significant challenge to modern society, yet our understanding of how people process false information remains limited. This thesis investigates the computational and neural mechanisms underlying learning from false information, with a focus on whether well-established learning biases persist even when information is debunked. Across three studies combining behavioural testing, computational modelling, and neuroimaging, I demonstrate that people continue to learn from information explicitly marked as false, and that this learning is biased. In the first study (two experiments), using a reinforcement learning task, I show that confirmation bias persists for false information. Participants exhibited higher learning rates for confirmatory versus disconfirmatory feedback no matter the veracity. The second study, using a belief-updating paradigm, reveals that optimistic update bias similarly persists for false information. Participants updated their beliefs more strongly in response to false good news than false bad news about adverse future life events.

Computational modelling across both paradigms identified a consistent pattern: a model with four learning rates, separating information desirability (confirmatory/good news versus disconfirmatory/bad news) and accuracy (true versus false), best explained participants' behaviour. Further, the strength of both confirmation bias and optimistic update bias was similar for true and false information. Albeit effective in reducing false information integration, debunking was less effective for desirable vs undesirable false information.

The third study used functional MRI to examine the neural basis of biased false information processing. Results revealed that activity in the ventromedial prefrontal cortex (vmPFC) was modulated by an interaction between accuracy and confirmation, showing higher activation when participants learned that confirming (vs disconfirming) evidence was true, or that disconfirming (vs confirming) evidence was false. These findings identify mechanisms that support learning from false information despite debunking, with implications for understanding vulnerability to misinformation and developing effective interventions.

Item Type:	Thesis (Doctoral)
Faculty \ School:	Faculty of Social Sciences > School of Psychology
Depositing User:	Chris White
Date Deposited:	07 Jan 2026 13:07
Last Modified:	07 Jan 2026 13:07
URI:	https://ueaeprints.uea.ac.uk/id/eprint/101550
DOI:

Downloads

Actions (login required)

View Item