AI ageism and its consequence: Benevolent ageist attitudes expressed by LLMs could reinforce ageism in humans

Large language models (LLMs) can generate biased content, potentially reinforcing negative social attitudes. The present study investigates ageist attitudes embedded in LLMs and their influence on human attitudes following Human-LLM interactions, in the context of rapid global aging. Three studies were conducted. Study 1 employed the ValueBench paradigm (Ren et al., 2024) to assess ageism in seven LLMs, comparing their responses to human participants ( N = 150, M age = 31.61 years). Study 2 ( N = 526, M age = 30.89 years) and 3 ( N = 320, M age = 31.64 years) examined whether exposure to ageist social media comments attributed to either LLM or human agents could shift ageism in human participants. Study 1 revealed that while LLMs generally exhibited lower levels of ageism than humans, they expressed significantly more benevolent ageism, at levels comparable to human participants. Study 2 and 3 demonstrated that participants reported stronger ageist attitudes after exposure to pro-ageism comments and weaker ageist attitudes after exposure to anti-ageism comments. Importantly, this persuasive effect occurred only when the comments were attributed to LLM agents, not to human agents (Study 3). These findings identify a significant prevalence of benevolent ageism in generative AI, a bias that may be overlooked in value alignment processes. Moreover, the results demonstrate that LLMs can exert greater influence than human in shaping individuals’ ageist attitudes. Future research should investigate the psychological mechanisms underlying this effect and explore how LLMs can be designed to promote, rather than undermine, intergenerational solidarity. • LLMs are generally less ageist than humans but still express certain ageist biases. • Similar to humans, LLMs exhibit more benevolent than hostile ageism. • LLMs could influence participant's ageist attitudes more effectively than humans.