The Aging and AI Harms Project: An Empirical Analysis of Risks to Older Adults

Background and Purpose: Studies on AI harms to older adults have been minimal and narrowly focused, making it difficult to systematically understand those harms while AI becomes more pervasive in everyday life. Researchers have lacked a comprehensive approach to categorize actual harms and identify their underlying patterns. The purpose of this study is to learn the frequency and nature of documented harms to older adults. To achieve this, we developed a structured coding approach tailored to older adults by integrating existing AI harms taxonomies with inductive qualitative coding.

Methods: Our dataset was drawn from three international curated public news repositories: the AI Incident Database, the OECD AI Hazards and Incidents Monitor, and the AI, Algorithmic, and Automation Incidents and Controversies (AIAAIC) Repository. To identify those incident reports in which at least one older adult was harmed, we applied 18 search terms (e.g., elder, aging/ageing). As an analytical tool for coding the incidents, we categorized each by two taxonomies widely used in research on algorithmic harms, the Sociotechnical Harms Taxonomy (Shelby et al., 2023) and the MIT Domain Taxonomy of AI Risks. Each incident was independently coded by two of six gerontology researchers working in three pairs, with discrepancies resolved through pair-level consensus meetings and full-team discussions. We conducted frequency counts complemented by qualitative interpretation to understand the meaning and context of the coded harms.

Results: Our search yielded 159 incidents. Omitting duplicates and those that did not impact an older adult reduced the sample to 80 incidents. Multiple incidents reported harms that were experienced by “hundreds” or “thousands” of older adults, while others reported harm to a single person. Our final sample of incidents occurred in 25 countries. The largest share occurred in the United States (35.3%), China (10.6%) and India (9.4%), with 21.2% in Europe. This approach organized diverse AI harms into a total of 13 domains and 38 subdomains across the two taxonomies, enabling a clearer understanding of complex harm patterns. We found that certain AI related harms were not well covered by the two taxonomies. Those that extend beyond the taxonomy categories include death, murder, suicide, and sexual assault, as well as additional harms such as those related to extreme dehydration and being made to leave a nursing home residence for stroke rehabilitations.

Conclusions and Implications: This research drew from three AI harms news repositories to develop the first insights into the frequency and nature of publicly reported incidents that have directly harmed older adults. Despite only capturing curated news incidents, this approach shows that older adults experience a wide range of AI harms and that these can be systematically structured and examined, providing a more integrated understanding of complex harm patterns. It also reveals important gaps in existing taxonomies by identifying multi-dimensional harms that have not been adequately captured, particularly in the lived experiences of older adults. Finally, the analytical approach developed in this study can be extended to support the systematic analysis of AI harms affecting other specific populations.