Large Language Model Counterarguments in Older Adults: Cognitive Offloading or Vulnerability to Moral Persuasion?

TL;DR

This study investigates how large language model-generated counterarguments influence moral judgments in younger and older adults, revealing significant susceptibility and highlighting potential cognitive and ethical implications.

Contribution

It demonstrates that LLMs can substantially persuade moral judgments, especially among older adults with lower cognitive functioning, and explores factors influencing this effect.

Findings

Over 30% of participants reversed moral judgments after LLM arguments.

Older adults showed greater judgment change, especially in the switch dilemma.

Lower cognitive functioning in older adults correlated with increased susceptibility in emotionally charged dilemmas.

Abstract

This study examined whether counterarguments generated by large language models (LLMs) influence the moral judgments of younger and older adults and whether these effects vary as a function of dilemma type, cognitive functioning, trust in AI, and prior experience using LLMs. Using the switch and footbridge trolley dilemmas, 130 participants (56 younger adults and 74 older adults) were presented with ChatGPT arguments that opposed their initial judgments. Results revealed that more than 30% of participants reversed their moral judgments in both dilemmas (32.31% in the switch dilemma and 36.92% in the footbridge dilemma), suggesting that LLMs possess substantial persuasive power. Older adults tended to be more likely than younger adults to reverse their judgments, and they showed a significantly greater degree of judgment change in the switch dilemma. Notably, in the emotionally aversive footbridge dilemma, older adults with lower cognitive functioning were significantly more likely to align with the LLM-generated counterargument. General trust in AI and prior experience with LLMs did not predict judgment reversal, supporting a disconnect between trust and persuasion. Instead, individual factors such as lower initial confidence and higher perceived task difficulty were associated with greater susceptibility to AI influence. These findings suggest that, although LLMs may serve as tools for cognitive offloading that compensate for age-related cognitive decline, they may also pose a risk of undue persuasion for cognitively vulnerable individuals.