- The paper demonstrates that LLM-generated counterarguments reversed moral judgments in over 30% of cases in classic trolley dilemmas.
- It reveals that lower cognitive functioning and increased metacognitive uncertainty in older adults heighten susceptibility to algorithmic persuasion.
- The study finds that explicit trust in AI did not predict reversal, while lower initial confidence and task difficulty were significant drivers.
Persuasive Effects of LLM-Generated Counterarguments on Moral Judgment in Aging Populations
Introduction
The investigation titled "LLM Counterarguments in Older Adults: Cognitive Offloading or Vulnerability to Moral Persuasion?" (2604.22356) provides a comprehensive experimental analysis of how LLM-generated counterarguments, specifically utilizing ChatGPT-4o, impact moral decision-making in both younger and older adults. By employing classic trolley dilemmas (switch and footbridge), the study isolates the persuasive properties of LLMs across age groups and cognitive strata, examining associations with cognitive functioning, trust in AI, prior LLM experience, and individual metacognitive states.
Experimental Design and Analytical Framework
The experimental protocol involved 130 healthy participants (56 younger, mean age ~21, and 74 older adults, mean age ~74), who were presented with moral dilemmas and subsequently exposed to standardized ChatGPT-4o-generated counterarguments directly opposing their initial stance. Changes in judgment were operationalized as binary reversals pre- versus post-counterargument exposure. The study leveraged GLMMs and logistic regression to dissect the influence of age, dilemma type, cognitive metrics (WAIS-IV), AI trust (Katase, 2021), prior LLM use, initial confidence, and subjective difficulty ratings. The power analyses confirmed sensitivity to medium effect sizes.
Key Findings
Magnitude of LLM-induced Judgment Reversal
- Approximately one-third of participants reversed their moral judgments after exposure to LLM counterarguments (32.31% in switch, 36.92% in footbridge), evidencing substantial persuasive efficacy, notably high compared to typical egocentric advice discounting rates observed in human-human advice paradigms.
- Contrary to the dual-process morality hypothesis, no statistically significant difference in reversal rates was observed between cognitive-control-dominant (switch) and emotionally-aversive (footbridge) dilemmas.
Age Effects and Cognitive Offloading
- Older adults were either as likely or more likely than younger adults to reverse their moral judgments in response to LLM counterarguments, with the effect being prominent in the switch dilemma.
- Cognitive offloading mediated LLM influence: Among older adults, lower cognitive functioning (across all WAIS-IV domains except coding)—especially in the emotionally-aversive footbridge dilemma—was significantly associated with judgment reversals. This suggests that reduced executive resources heighten susceptibility to external, algorithmic persuasion.
- These patterns support a heuristic "seems good enough" evaluation strategy: Participants across both dilemmas deferred to LLM outputs when metacognitive uncertainty or resource limitations prevailed, overriding both utilitarian and deontological prior stances.
- General trust in AI and prior LLM experience—despite strong age stratification—were not significant predictors of judgment reversal, empirically decoupling persuasion from explicit AI trust metrics.
- Lower pre-exposure confidence and higher perceived task difficulty robustly predicted reversals, emphasizing the primacy of metacognitive factors in mediating LLM persuasive power.
- Algorithm aversion effects, commonly hypothesized to restrict AI influence in subjective judgment domains, were not protective in this context; content-based plausibility of LLM output, not its source, drove user adoption.
Theoretical and Practical Implications
The findings underscore that LLMs act as persuasive agents capable of meaningfully altering complex, value-laden judgments—particularly among older adults and those with diminished cognitive resources. The LLM can function both as a compensatory cognitive prosthesis and a vector for undue moral persuasion, implicating LLMs as dual-use tools in environments where users face high decisional uncertainty.
For AI developers and policymakers, these results strongly indicate the necessity for age- and cognition-sensitive interaction paradigms. LLM-mediated decision aids must balance assistive benefits against the risk of undermining autonomous value-based reasoning, particularly for populations exhibiting cognitive decline. The observed disconnect between AI trust and susceptibility to persuasion highlights the limited utility of trust calibrations as safeguards. Instead, interface designs should prioritize fostering metacognitive engagement and critical evaluation over wholesale cognitive offloading.
Additionally, the high reversibility rate in classic philosophical dilemmas generalizes to broader societal decisions—raising questions regarding LLM deployment in healthcare, law, and high-stakes public communications. The demonstrated impact on populations less familiar with LLMs suggests that digital literacy interventions alone may be insufficient to mitigate vulnerability.
Prospects for Future Research
Further studies should expand to longitudinal designs to determine the duration of LLM-induced belief change and test effects in ecologically valid, everyday moral or practical decisions. Real-time, dialogic LLM-human interactions may further augment persuasive impact—a direction unaddressed due to experimental constraints. Identification and reinforcement of cognitive and interface-level "friction points" may be critical to ensure end-user agency in an AI-pervasive decision landscape.
Conclusion
This study provides robust evidence that LLM-generated counterarguments can effectuate substantial, sometimes complete, reversals in moral judgment, with effects amplified in older adults and those with lower cognitive capacity. The work delineates the dual nature of LLMs as both cognitive aids and sources of over-persuasion, emphasizing the importance of user-centered, ethically-aligned AI system design. Individuated risk of undue persuasion emerges primarily from cognitive and metacognitive vulnerabilities, independent of explicit AI trust, underscoring a need for future work on decision literacy and robust interface interventions as AI continues to mediate critical human judgments.
