Learning associations between actions and their outcomes is fundamental for adaptive behavior. Most studies examining reinforcement learning have focused on self-relevant learning, primarily in young individuals. However, self-relevant learning may be computationally distinct from prosocial learning—learning about actions that benefit others. Previous research suggests slower prosocial learning compared to self-relevant learning. While senescence is associated with cognitive decline, the preservation of affective processing and social cognitive abilities remains less understood, especially concerning its impact on social behavior and prosociality. Social isolation negatively affects physical health, highlighting the importance of prosocial behaviors in maintaining social bonds and overall well-being throughout the lifespan. Prosocial behavior is linked to improved life satisfaction, mental well-being, and physical health. A crucial aspect of prosocial behavior is learning the association between one's actions and their outcomes for others. This study uses computational models of reinforcement learning in young and older adults to examine the mechanisms underlying self-relevant and prosocial learning, and their associations with individual differences in socio-cognitive abilities. Reinforcement Learning Theory (RLT) provides a robust framework for understanding and modeling learning. In RLT, prediction errors signal unexpected outcomes and influence future choices. The learning rate quantifies the impact of past outcomes on behavior. The neurobiological plausibility of RLT is supported by the encoding of prediction errors by neurons in the ventral tegmental area. While crucial for adaptive behavior, reinforcement learning propensity may decline with age. Older adults show learning impairments, particularly with probabilistic or reversed action-outcome associations. Age-related learning decline is linked to changes in frontostriatal circuits and dopamine transmission, which significantly decreases with age and plays a key role in coding prediction errors. Administering L-DOPA, a dopamine precursor, to older adults increased their learning rate, supporting the link between dopamine and learning. Alternatively, prosocial learning might depend on motivation to help others, not just learning ability. Studies using economic games show older adults are often more generous and make more charitable donations. Older adults engage in more prosocial behaviors at work. Self-reported altruism also increases with age. However, a limitation of these studies is the conflict between self and other rewards; benefiting others often means less for oneself. Older adults also tend to have greater accumulated wealth, confounding studies of monetary exchange. Prosocial learning avoids this confound by separating outcomes for self and other. If older adults value outcomes for others more, we might expect preserved prosocial learning despite reduced self-relevant learning with aging. Comparing young and older adults on self-relevant and prosocial learning helps dissociate the effects of age-related changes in cognitive ability and social behavior on learning. Individual variability in concern for others, captured by trait measures of psychopathic traits, is also relevant. Reduced prosocial behavior and altered self/other reward processing are features of psychopathy, characterized by dysfunctional affective-interpersonal and antisocial behavioral dimensions. Psychopathy is considered dimensional, ranging from clinical psychopathy to subclinical levels in the general population. Subclinical self-report measures, like the Self-Report Psychopathy Scale (SRP) used here, capture this range. The SRP assesses 'affective-interpersonal' (lack of empathy and guilt) and 'lifestyle-antisocial' (impulsive and antisocial behaviors) dimensions. Ageing may be associated with changes in self-reported psychopathic traits, potentially impacting our understanding of an aging population. In community samples, ageing is associated with a decrease in both dimensions of the SRP. This highlights the importance of assessing the relationship between self-reported psychopathic traits and prosocial behaviors.

Existing research offers conflicting hypotheses regarding the association between aging and self-relevant versus prosocial reinforcement learning. One hypothesis posits that older adults will be impaired in learning regardless of the recipient, consistent with age-related declines in learning ability and dopamine transmission. Conversely, a potential increase in valuing outcomes for others in older adults might predict preserved prosocial learning but reduced self-relevant learning. Finally, it was hypothesized that variation in self-reported psychopathic traits would be associated with learning for others but not self in both age groups. The current study aimed to test these competing hypotheses.

This study involved 75 young (18–36 years old, mean = 23.07, 44 females) and 77 older (60–80 years old, mean = 69.84, 40 females) adults matched on gender, education, and IQ. Participants completed a probabilistic reinforcement-learning task designed to separate self-relevant (rewards for self) from prosocial learning (rewards for another person), and a control condition (rewards for neither). The task involved learning the probability of reward associated with abstract symbols. Points earned in the 'self' condition translated into additional payment for the participant; points in the 'other' condition were converted into money for another participant (a confederate); and points in the 'no one' condition were displayed but not converted into money. To ensure comparability, older adults with dementia (as determined by the Addenbrooke’s Cognitive Examination, ACE) were excluded. The two age groups were matched on gender, years of education, and IQ (Wechsler Test of Adult Reading, WTAR). Additional analyses controlled for IQ, memory, and attention (ACE subscales). Neuropsychological tests and a self-report measure of psychopathic traits (SRP-IV-SF) were also administered. The SRP assesses affective-interpersonal (lack of empathy and guilt) and lifestyle-antisocial (impulsivity and antisocial behavior) traits. Computational models of reinforcement learning were used to estimate learning rates (α) and temperature parameters (β). Four candidate models were compared: 1α1β (one learning rate and one temperature parameter for all recipients), 3α1β (separate learning rates for each recipient, one shared temperature parameter), 2α1β (separate learning rates for self vs. other/no one, one shared temperature parameter), and 3α3β (separate learning rates and temperature parameters for each recipient). Model identifiability and parameter recovery were assessed using simulations. Bayesian model selection (based on exceedance probability and integrated Bayesian Information Criterion, BIC) was used to determine the best-fitting model. A robust linear mixed-effects model (RLMM) was used to analyze learning rates, considering age group, recipient, and their interaction. Generalized linear mixed-effects models (GLMMs) analyzed trial-by-trial choices. Correlations between learning rates and self-reported psychopathic traits were calculated using Spearman's rho. Indirect effects of age group on relative prosocial learning rate were analyzed using mediation models.

Model comparison revealed that a model with separate learning rates for each recipient (3α1β) best explained participant choices. Young adults exhibited faster learning when their actions benefitted themselves compared to others. Older adults showed reduced self-relevant learning rates but preserved prosocial learning rates. Bayesian analyses supported no significant difference in prosocial learning rates between young and older adults. Older adults had significantly lower levels of self-reported psychopathic traits compared to young adults. In older adults, lower self-reported affective-interpersonal psychopathic traits correlated negatively with prosocial learning rates. This correlation was significantly more negative in older adults than in young adults. These effects were not explained by differences in IQ, memory, or attention. The RLMM analysis showed that across age groups, participants had a higher learning rate when rewards were for themselves compared to others. However, this difference was significantly reduced in older adults compared to younger adults. Older adults learned more slowly for themselves than younger adults, but their prosocial learning rates did not differ significantly from those of young adults. Young adults learned faster for themselves than for others, whereas older adults showed no significant difference between these learning rates. Both age groups learned preferentially for themselves compared to the 'no one' condition. Older adults, but not young adults, differentiated between learning for another person and the 'no one' condition. Older adults showed slower overall learning and higher exploration (higher β). Trial-by-trial analysis showed older adults chose the high-reward option less frequently and improved less during the task. Overall, older adults showed preserved prosocial learning despite age-related declines in self-relevant learning rates. Analysis of self-reported psychopathic traits revealed significantly lower scores in older adults on both affective-interpersonal and lifestyle-antisocial subscales. In older adults, prosocial learning rates were negatively correlated with affective-interpersonal psychopathic traits. This correlation was significantly more negative than in young adults. This negative correlation remained significant after controlling for choice exploration (β) and after FDR correction. There were no significant correlations between self-reported psychopathic traits and self-relevant or control condition learning rates. Moderated mediation analysis indicated a significant indirect effect of age group on relative prosocial learning rate (αother - αself) via self-reported psychopathic traits for older adults, but not for young adults.

This study demonstrates that despite age-related declines in self-relevant reinforcement learning, prosocial learning is preserved in older adults. Using computational models, we precisely examined the influence of reward history on learning by isolating the learning rate. We replicated previous findings of computationally separable self-relevant and prosocial learning, with separate learning rates for different recipients providing the best model fit. The increased learning rate for self-relevant rewards compared to prosocial rewards was reduced in older adults, who exhibited relatively higher prosocial learning rates than young adults. The preservation of prosocial learning rates, despite declines in self-relevant learning rates, is unlikely due to changes in executive function or general intelligence, as our results remained consistent after controlling for these factors. The preserved prosocial learning in older adults could be attributed to increased prosocial motivation. This is supported by the observed link between learning rates and psychopathic traits. Self-reported psychopathic traits were lower in older adults, and these traits negatively correlated with prosocial learning rates specifically in older adults. This suggests that age-related differences in prosocial learning could be linked to changes in individual traits and motivations rather than domain-general cognitive decline. The control condition (no one benefits) ruled out the possibility that the lack of difference between self and other learning rates for older adults was due to a general reduction in the dynamic range of learning rates. Older adults exhibited a relative increase in learning rates specifically in the prosocial condition, demonstrating their sensitivity to the recipient condition. The similar magnitude of decrease in self-relevant learning rates associated with aging and the decrease shown by young participants when learning for someone else supports distinct neural mechanisms for learning from outcomes for self and other. The findings suggest that age-related increases in prosocial motivation may contribute to the preservation of prosocial learning. This could be driven by factors like vicarious reinforcement, as reputational concerns were minimized in the experimental design. Our task separated self-relevant learning, prosocial learning, and a control condition, allowing us to differentiate between increases in the value of prosocial outcomes and decreases in the value of self-outcomes. The results align with older adults exhibiting both decreased self-relevant learning rates and increased prosocial learning rates compared to the control condition.

This study demonstrates that despite age-related declines in self-relevant reinforcement learning, the ability to learn actions that benefit others is preserved in older adults. The bias towards self-relevant outcomes is reduced, with older adults showing relatively preserved prosocial learning and lower levels of psychopathic traits, which are linked to prosocial learning. These findings have important implications for understanding reinforcement learning and healthy aging. Future research should investigate the age at which these changes occur, use community-based samples, examine the sustainability of motivation in older adults, and explore the role of empathy in prosocial learning. Further studies could also manipulate recipient identity to assess potential effects of perceived social distance.

The cross-sectional design limits conclusions about the age at which these changes occur. While education and IQ were controlled for, the sampling method (university databases) might not fully represent the general population. Future research should address this by using community-based samples spanning the whole adult lifespan. Further investigation is also needed to determine whether older adults reach the same performance ceiling as younger adults over time or whether any differences in performance persist. The study's limited time window prevents conclusions about the sustainability of higher motivation levels in older adults. Finally, the study did not explicitly measure empathy, a factor that could influence prosocial learning.