Truth Bias in Human Reinforcement Learning
Abstract
Humans display a truth bias, a default tendency to treat incoming information as true unless there is reason to doubt its veracity. Whether this bias is confined to the evaluation and recollection of declarative information, or extends to how we learn and make decisions, remains unknown. Here, we introduce a computational reinforcement-learning framework in which participants receive choice-outcome feedback from sources explicitly identified as truthful or untruthful. Because truthful and untruthful feedback are matched in their objective informational value, the task isolates asymmetries in how feedback truthfulness shapes learning. Across two behavioural studies (total n = 403; Study 2 pre-registered), truthful feedback exerted a greater influence on learning and subsequent choice than equally informative untruthful feedback. This learning asymmetry deviated from optimal statistical learning and persisted when we manipulated the prevalence of untruthful sources. The asymmetry also extended beyond reinforcement learning. Participants showed a corresponding truth bias in explicit inference, and individual differences in explicit and reinforcement-learning truth biases were positively correlated. Moreover, combining reinforcement learning with evidence-accumulation modelling revealed that merely anticipating feedback from an untruthful source reduced the efficiency with which learned values guided evidence accumulation, producing slower and less accurate choices before any feedback was received. Thus, truth asymmetries span multiple facets of cognitive processing, from learning and explicit inference through to value-based decision making. Together, these findings show that truthful and untruthful information are not cognitively equivalent even when they carry the same objective information.
Related
- Hippocampal-cingulate dynamics in the human brain link reinforcement-learning and memory
- How Past Experiences Shape New Beliefs: Generalisation and Adaptation in Human Reinforcement Learning
- Biased processing of multiple outcomes in human reinforcement learning: evidence from computational modeling and eye-tracking
- A novel methodology for assessing reinforcement choice in primates under human care
- Real-time reinforcement for human-machine interface control
BibTeX
@misc{vidalperez2026truth,
title = {{Truth Bias in Human Reinforcement Learning}},
author = {Juan Vidal-Perez and Raymond J Dolan and Rani Moran},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/76ewk_v1},
url = {https://osf.io/preprints/psyarxiv/76ewk_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Truth Bias in Human Reinforcement Learning
AU - Juan Vidal-Perez
AU - Raymond J Dolan
AU - Rani Moran
PY - 2026
DA - 2026/09/29
PB - PsyArXiv
DO - 10.31234/osf.io/76ewk_v1
UR - https://osf.io/preprints/psyarxiv/76ewk_v1/
AB - Humans display a truth bias, a default tendency to treat incoming information as true unless there is reason to doubt its veracity. Whether this bias is confined to the evaluation and recollection of declarative information, or extends to how we learn and make decisions, remains unknown. Here, we introduce a computational reinforcement-learning framework in which participants receive choice-outcome feedback from sources explicitly identified as truthful or untruthful. Because truthful and untruthful feedback are matched in their objective informational value, the task isolates asymmetries in how feedback truthfulness shapes learning. Across two behavioural studies (total n = 403; Study 2 pre-registered), truthful feedback exerted a greater influence on learning and subsequent choice than equally informative untruthful feedback. This learning asymmetry deviated from optimal statistical learning and persisted when we manipulated the prevalence of untruthful sources. The asymmetry also extended beyond reinforcement learning. Participants showed a corresponding truth bias in explicit inference, and individual differences in explicit and reinforcement-learning truth biases were positively correlated. Moreover, combining reinforcement learning with evidence-accumulation modelling revealed that merely anticipating feedback from an untruthful source reduced the efficiency with which learned values guided evidence accumulation, producing slower and less accurate choices before any feedback was received. Thus, truth asymmetries span multiple facets of cognitive processing, from learning and explicit inference through to value-based decision making. Together, these findings show that truthful and untruthful information are not cognitively equivalent even when they carry the same objective information.
ER -