Ageing is associated with disrupted reinforcement learning whilst learning to help others is preserved

The study investigates how ageing impacts reinforcement learning (RL) when outcomes affect oneself versus another person (prosocial learning). Prior work indicates age-related declines in self-relevant RL, potentially due to changes in frontostriatal circuitry and dopamine function. Conversely, older adults often show increased prosocial behaviours and generosity, suggesting prosocial motivation might be preserved or enhanced with age. The authors contrast two hypotheses: (1) general learning decline predicts reduced learning rates for both self and other in older adults; (2) preserved or enhanced valuation of others’ outcomes predicts reduced self-relevant learning but preserved prosocial learning in older adults. They also examine whether individual differences in subclinical psychopathic traits—particularly affective-interpersonal traits—relate to prosocial learning and whether these traits differ with age.

Reinforcement Learning Theory posits that prediction errors guide future choices, with dopamine systems encoding these errors. Ageing is associated with declines in probabilistic learning and reversal learning, linked to frontostriatal and dopaminergic changes; pharmacological evidence (e.g., L-DOPA) suggests dopamine modulates learning rates in older adults. In contrast, economic game studies suggest prosocial behaviour, charitable giving, and workplace helping increase with age, though many paradigms confound self-other trade-offs and wealth effects. Prosocial learning avoids these confounds by separating outcomes for self and other. Psychopathic traits, conceptualized dimensionally, encompass affective-interpersonal and lifestyle-antisocial components; higher traits are associated with reduced empathy and prosociality. Community studies show age-related decreases in self-reported psychopathic traits. These lines of evidence motivate testing whether ageing differentially affects self versus prosocial learning and how traits relate to prosocial learning.

Design: Cross-sectional comparison of young (18–36) and older (60–80) adults, matched on gender, years of education, and IQ test performance. Participants: Recruited 80 young and 80 older adults; exclusions (psychiatric diagnosis, prior psychology study, incomplete/low-quality data) yielded final N=152 (75 young: 44 females, mean age 23.07; 77 older: 40 females, mean age 69.84). Older adults were screened with ACE-III; those below dementia cut-off were excluded. Two older participants with extreme learning rate estimates (>3 SDs in two conditions) were excluded from learning-rate analyses; one participant per group lacked SRP data for trait analyses. Task: Probabilistic reinforcement-learning task with three recipient conditions per block: Self (points convert to participant bonus), Other (points convert to anonymous partner bonus via role-assignment procedure minimizing reputational concerns), and No one (control; points not converted). Each recipient condition comprised three blocks of 16 trials (48 trials per recipient; 144 total). On each trial, participants chose between two symbols with fixed reward probabilities (high: 75%, low: 25%) and received feedback. Role-assignment: Participants and a confederate (identity concealed) underwent a standardized procedure to enhance belief that prosocial points benefited another person while preventing reputational effects and controlling identity cues. Measures: IQ via WTAR (age-standardized). Older adults: ACE-III (attention and memory subscales used as executive function proxies). Psychopathic traits via SRP-IV-SF, analyzed as affective-interpersonal and lifestyle-antisocial factors. Computational modelling: RL models with recipient-specific parameters compared. Four models: (i) 1α1β (single learning rate and temperature across recipients), (ii) 3α1β (separate α for self/other/no one; single β), (iii) 2α1β (αself and a shared αnot-self; single β), (iv) 3α3β (separate α and β for each recipient). Fitting via hierarchical maximum a posteriori (MAP) across combined sample; bounded parameters with appropriate link functions. Model comparison used Laplace-approximated log evidence with random-effects Bayesian model selection (exceedance probability), integrated BIC, and model R^2. Model identifiability and parameter recovery assessed via simulations; additional simulations established optimal α (~0.55) and relation to accuracy. Statistical analyses: Robust linear mixed-effects models (RLMM) for learning rates with fixed effects of age group, recipient, and interaction; random intercepts. Generalized mixed-effects models for trial-by-trial accuracy including trial number and interactions. Nonparametric tests for planned comparisons; Bayes factors (BF01) for null support in non-significant comparisons. Spearman correlations for associations with psychopathic traits; partial correlations controlling for β, IQ, attention, and memory. Mediation and moderated mediation analyses tested indirect effects of age group on relative prosocial learning (Δ = αother − αself) via affective-interpersonal traits.

- Model comparison: The 3α1β model (separate learning rates for self, other, no one; single β) best fit behaviour (exceedance probability ≈ 97%; integrated BIC favored 3α1β; model R^2 ≈ 51%). Simulations established model identifiability, parameter recovery (all r > 0.7), and optimal α ≈ 0.55; participants’ α values were below this, so higher α associated with better accuracy (overall Spearman ρ(150) = 0.58 [0.46, 0.68], P < 0.001). - Learning overall: Participants learned above chance for all recipients and age groups (all t > 15.46, P < 0.001; trial number effects Z > 4.48, P < 0.001). - Age and recipient effects on learning rates (α): Across groups, αself > αother (recipient self vs. other: b = −0.02 [−0.03, −0.01], Z = −4.79, P < 0.001). Critically, the self–other difference was reduced in older vs. younger adults (interaction b = 0.02 [0.002, 0.03], Z = 2.29, P = 0.02). Older adults learned more slowly for self than young (W = 3512, Z = −2.63, rs(150) = 0.22 [0.06, 0.36], P = 0.009). Prosocial learning was preserved: no significant group difference in αother (W = 3042, Z = −0.86, rs(150) = 0.07 [0.003, 0.24], P = 0.39; BF01 = 4.26 supporting null). Within-group: young αself > αother (V = 659, Z = −4.04, rs(75) = 0.47 [0.26, 0.63], P < 0.001); older showed no self–other difference (V = 1150, Z = −1.45, rs(75) = 0.17 [0.01, 0.38], P = 0.15; BF01 = 1.08). - Self vs. no one: αself > αno one overall (b = −0.02 [−0.03, −0.01], Z = −4.57, P < 0.001); no interaction with age (b = 0.008 [−0.006, 0.02], Z = 1.15, P = 0.25). Within groups, both preferred self over no one (young V = 928, Z = −2.62, P = 0.009; older V = 901, Z = −2.76, P = 0.006). No significant age-group difference in αno one (W = 3241, Z = −1.61, P = 0.11; BF01 = 2.04 inconclusive). - Other vs. no one: Young did not differentiate other vs. no one (V = 1533, Z = −0.57, P = 0.57; BF01 = 5.08 supporting null). Older showed αother > αno one (V = 976, Z = −2.37, P = 0.02), indicating sensitivity to recipient. - Temperature (β) and overall learning: Older had higher exploration (median β young 0.05 vs. older 0.19; W = 1511, Z = −4.89, rs(150) = 0.40, P < 0.001) and slower learning overall (b = −0.02 [−0.03, −0.01], Z = −3.73, P < 0.001). Trial-by-trial accuracy: older chose high-reward option less (mean 64% vs. young 80%; b = −1.20 [−1.65, −0.70], Z = −4.84, P < 0.001) and improved less over trials (interaction b = −0.81 [−1.34, −0.27], Z = −2.95, P = 0.003). Across groups, self > no one in accuracy (b = −0.36 [−0.68, −0.05], P = 0.02); self vs. other accuracy difference not significant (b = −0.22 [−0.49, 0.05], P = 0.10). - Psychopathic traits: Older adults had lower self-reported affective-interpersonal (young mean = 24.36, older = 21.09; W = 3558, Z = −3.15, r(148) = 0.26, P = 0.002) and lifestyle-antisocial scores (young = 22.89, older = 20.27; W = 3471, Z = −2.82, r(148) = 0.23, P = 0.005). In older adults, affective-interpersonal traits negatively correlated with prosocial learning (αother) (rs(74) = −0.33 [−0.52, −0.11], P = 0.005; FDR-corrected P = 0.03); no significant relation in young (rs(74) = 0.21 [−0.02, 0.42], P = 0.07); correlation more negative in older vs. young (Z = 3.28, P = 0.001). Relative prosocial learning (αother − αself) also negatively correlated with traits in older (rs(74) = −0.25, P = 0.03) but not young (rs(74) = 0.11, P = 0.34; difference Z = 2.18, P = 0.03). Effects robust to excluding outliers and controlling for β, IQ, attention, and memory. - Moderated mediation: No indirect effect in standard mediation. Moderated mediation showed an indirect effect of age group on relative prosocial learning via affective-interpersonal traits for older adults only (unstandardised indirect effect = 0.006 [0.001, 0.01], P = 0.006; proportion mediated = 0.30), not for young (−0.001 [−0.007, 0.006], P = 0.67).

The findings resolve competing hypotheses by showing that ageing is associated with reduced self-relevant learning rates but preserved prosocial learning rates. This suggests that while neurocognitive changes with age (e.g., dopaminergic decline, altered frontostriatal function) may impair self-relevant RL, prosocial learning remains intact, potentially due to increased prosocial motivation with age. Model-based analyses revealed distinct learning rates by recipient, indicating separable computational processes for self vs. other outcomes. Older adults’ preserved prosocial learning, coupled with reduced psychopathic traits and their specific negative association with prosocial learning in older adults, indicates that trait-level socio-affective differences contribute to prosocial learning in ageing beyond general cognitive ability. The inclusion of a no-one control condition demonstrated that older adults specifically distinguished outcomes for others over no one, refuting a general insensitivity to recipient. These results align with evidence that neural systems for learning from others’ outcomes may differ from self-related learning and suggest leveraging preserved social motivations could support healthy ageing.

Despite age-related declines in self-relevant reinforcement learning, the ability to learn which actions benefit others is preserved in older adults. Young adults showed a self-bias in learning rates, whereas older adults did not differentiate between self and other to the same extent and uniquely distinguished other from no one. Older adults reported lower affective-interpersonal and lifestyle-antisocial psychopathic traits; lower affective-interpersonal traits related to higher prosocial learning specifically in older adults, mediating age effects in a moderated fashion. This work advances understanding of RL across the lifespan and highlights the importance of distinguishing recipient-specific learning. Future research should employ lifespan samples to pinpoint trajectories, extend task duration to assess ceiling performance and sustained motivation, incorporate measures of empathy and other socio-affective traits, and manipulate recipient identity and social distance to test boundary conditions of preserved prosocial learning.

- Cross-sectional two-group design prevents identifying the precise age and rate at which changes emerge. - Sample, though matched on education and IQ, may not be fully representative due to recruitment sources and self-selection. - Limited task duration and conditions restrict assessment of whether older adults eventually reach similar performance ceilings or sustain motivation over longer periods. - Empathy and related socio-affective constructs were not directly measured alongside psychopathic traits; their roles in prosocial learning remain to be tested. - Recipient identity was kept minimal to control reputation; effects may vary with social distance or recipient characteristics not examined here.