Generative AI harbors subtle bias against older adults, KAIST study finds

Even when AI does not explicitly portray older adults negatively, it tends to implicitly assign them lower levels of agency and competence. Such depictions may appear positive on the surface, but repeated exposure could narrow social expectations of older people and limit how they perceive themselves.

KAIST announced Sunday that a research team led by Choi Moon-jung, a professor at the Graduate School of Science and Technology Policy, had quantitatively analyzed subtle age-related stereotypes embedded in text generated by OpenAI's ChatGPT-4o.

Generative AI is now widely used in everyday information searches and decision-making, but concerns have long been raised that it may reproduce social biases present in its training data. While previous studies have largely focused on gender and racial bias, this research is notable for examining ageism — discrimination against or negative perception of individuals based on age — from an AI perspective, a dimension gaining importance amid global population aging.

The research team collected 900 text samples generated by GPT-4o using neutral prompts asking the model to describe characteristics of age groups in 10-year increments from age 10 to 90. The team then applied the Stereotype Content Model, a leading social psychology framework that explains perceptions of people and groups along two dimensions — warmth and competence.

The analysis found that older adults (aged 60 and above) scored relatively high on warmth but tended to score lower on competence compared with younger age groups.

The generated responses also showed a tendency to divide the human life cycle into three broad clusters: youth (teens and 20s), middle age (30s to 50s), and old age (60s and above). Descriptions of adults in their 70s and older were particularly repetitive and uniform.

The research team also examined assertiveness — expressions reflecting confidence and initiative. The frequency of assertiveness-related language declined as age increased, suggesting that ChatGPT-4o tends to portray older adults as wise and benevolent while depicting them as relatively lacking in autonomy and initiative.

The study is significant for quantitatively identifying subtle biases in generative AI by combining social science theory with computational analysis. The findings show that generative AI tends to depict older adults as "warm but less competent" — a pattern that mirrors stereotypical portrayals of the elderly recurring across mainstream media.

The research team said repeated exposure to such depictions through conversational AI services could reinforce social prejudice against older adults. The team also raised the possibility that the phenomenon could lead to "digital ageism," in which older people are discouraged from participating in digital life.

"AI bias is not a technological problem — it is a social one," Choi said. "For AI to be truly inclusive, people of diverse generations must be involved in the development process."

The findings were published in a special issue of the February edition of The Gerontologist, an international academic journal.