A KAIST research team provides quantitative evidence that large language models reproduce ageist stereotypes, highlighting the need for inclusive AI training frameworks.
Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have identified significant age-related stereotypes embedded within the output of OpenAI’s GPT-4o model. The study, published in the February 2026 issue of The Gerontologist, provides a quantitative assessment of how large language models replicate human social biases during text generation.
Professor Moon Choi, lead researcher at the Graduate School of Science and Technology Policy, directed the team in analyzing 900 distinct text samples produced by the model. The team utilized neutral prompts designed to elicit descriptions of individuals across age cohorts ranging from 10 to 90 years in decade-long intervals.
To evaluate these outputs, the researchers applied the Stereotype Content Model, a framework in social psychology that maps perceptions along the dimensions of warmth and competence. The model consistently assigned higher warmth scores to individuals aged 60 and above, characterizing them as kind and trustworthy.
Conversely, the analysis revealed a marked decline in competence scores for the same older demographic compared to younger age groups. The model’s outputs frequently associated competence—defined as expertise, efficiency, and capability—with younger and middle-aged subjects while diminishing these traits for those in their 60s and beyond.
The research also examined the linguistic frequency of terms related to assertiveness, which the team defined as the capacity to act with confidence and initiative. The data showed a statistically significant decrease in assertive language as the age of the subject increased, suggesting the model internalizes a narrative of declining agency in older populations.
Descriptions of individuals aged 70 and older exhibited a high degree of uniformity, indicating that the model relies on simplified, stereotypical representations rather than nuanced characterizations. This pattern suggests that the training data likely contains historical human biases that the model reproduces through its probabilistic text generation processes.
The research team further categorized the human life course into three distinct segments based on the model’s output: youth, middle age, and older adulthood. This segmentation reveals how the model maps social roles to specific age brackets, often reinforcing traditional societal expectations that may not align with modern demographic realities.
Statistical analysis performed by the team confirmed that these findings were not random fluctuations but consistent patterns across the 900-sample dataset. By applying rigorous quantitative methods to semantic outputs, the researchers demonstrated that the model exhibits a systematic bias that mirrors historical media portrayals of aging.
The findings indicate that generative AI models may inadvertently reinforce digital ageism by codifying these stereotypes into the information retrieval and decision-making processes that users rely on daily. By framing older adults as warm but fundamentally less capable, the model risks influencing user perceptions and potentially discouraging the digital participation of older generations.
The research highlights that bias in large language models is not merely a technical artifact but a reflection of deep-seated social dynamics. Professor Choi emphasizes that mitigating these issues requires the inclusion of diverse generational perspectives throughout the development and fine-tuning phases of model architecture.
Addressing these biases necessitates a shift from purely algorithmic optimization to a more sociotechnical approach in model training. Future development must prioritize the representation of diverse age groups in training datasets to ensure that generative systems do not perpetuate discriminatory social structures.
The study serves as a critical benchmark for evaluating the ethical implications of deploying generative AI in social contexts. As these models become increasingly integrated into professional and personal decision-making, the need for rigorous auditing of model outputs for demographic bias remains a primary concern for the machine learning community.
The team suggests that future research should investigate whether these biases persist across different languages and cultural contexts. Understanding the universality of these stereotypes will be essential for developers aiming to build inclusive AI systems that operate effectively across global populations.
Ultimately, the KAIST study underscores that the path toward more equitable AI requires moving beyond simple performance metrics. Engineers must now account for the sociological impact of their training data to prevent the automation of historical prejudices.
