Korean AI personas leave disability and sexual identities out
An investigation into Korean synthetic personas finds disability, sexual orientation and gender identity missing from the categories used to represent adults. Researchers describe models completing unspecified traits with majority identities. Nvidia points to gaps in public gender-identity statistics, while official disability data already exist. The findings concern generated representations used in research and public-policy work.
Artificial Intelligence··Morning
Missing categories affect whose lives the models represent
A synthetic population intended to stand in for Korean adults leaves disability, sexual orientation and gender identity outside its demographic categories. An investigation by Hankyoreh draws on research at the Graduate School of Public Health of Seoul National University, by Kim Seungseop, Kim Sumin and Pang Eunji. The researchers describe language models supplying dominant-majority identities when characteristics are unspecified. Their findings concern answers generated for artificial personas, rather than responses collected from people.[1]
Nvidia cites a statistics gap, but disability data exist
Nvidia, the computing company behind the dataset, says comprehensive public statistics on Korean gender identity were unavailable. Disability presents a different situation: official figures exist, yet the classification omits it. Hankyoreh cites census data putting registered disabled adults aged 20 or older at 5.9 per cent of that age group in 2025. Gender identity and biological sex are distinct characteristics in the company’s explanation.[1]
The dataset was built to improve Korean representation
Released in April, Nvidia’s dataset aimed to represent a million adults in Korea, starting at age 19, addressing gaps in models trained predominantly on English and other mainstream languages. Its fields cover age, marital status, military service, education, occupation, residence and family and housing types. The investigation also describes similar directional findings with Meta’s Llama 3.3 alongside systems made by Google, Anthropic and OpenAI. Kim Seungseop warns that excluding minority lives can deepen inequality as AI enters policy research and public programmes.[1]
