Global investment bank Goldman Sachs estimates that global spending on artificial intelligence (AI) infrastructure this year will reach $1 trillion (approx. 1300 trillion KRW). This massive investment is accelerating AI technological advancements and greatly expanding its application scope. However, a "brave new world" is not the only thing waiting for us. In a two-part series, The Hankyoreh explores how AI algorithms exclude minorities and social suffering.

In an artificial intelligence (AI)-built virtual world, a phenomenon where minorities disappear has emerged, according to a recent research study. Critics point out that large language model (LLM) algorithms—such as ChatGPT and Gemini—are erasing individuals such as people with disabilities and sexual minorities from their computational processes. Concerns are rising that if AI models carrying minority-excluding biases are utilized in government policies and other areas, social discrimination and inequality could deepen.

On the 18th, The Hankyoreh obtained a preprint paper registered on arXiv, the open-access repository operated by Cornell University, by a research team led by Kim Seung-sup, a professor at the Department of Environmental Health at Seoul National University Graduate School of Public Health (along with researchers Kim Soo-min, Pang Eun-ji, and Kim Seung-sup). The paper is titled, "Architects of Silence: Persona-Injected Large Language Models Render Minorities and Social Suffering Invisible."

In this paper, after the research team inputted a 1-million-scale synthetic persona (virtual subject) dataset of South Koreans—created by AI semiconductor giant Nvidia—into an LLM and randomly extracted 10,000 virtual subjects to question them, it was revealed that not a single case responded identifying themselves as a minority (such as persons with disabilities, sexual minorities, or individuals born abroad). This confirmed a phenomenon where minorities are completely erased within the AI-generated virtual world.

Last April, Nvidia officially released and began providing a free 1-million-scale "persona dataset," stating that it "precisely reflects and reproduces South Korea's actual demographic distribution as well as geographical and cultural characteristics." This data can be utilized across various fields, including AI service development and learning, customer marketing, and research.

The research team stated in the paper, "A thorough additional review is urgent before AI models are more widely applied to human-targeted research or policy establishment that could amplify social inequality."

What is Persona Data? — Virtual Subject Datasets Based on National Population Statistics
Since last year, Nvidia has been releasing large-scale virtual subject datasets named "NeMoTron Persona Data," which meticulously embody genders, occupations, and other traits based on national demographic statistics. Adaptable to individual national realities, data has currently been built for 10 countries, including the United States, India, Japan, and France. The scale of virtual subjects reaches 3 million in India, and 1 million each in South Korea, the United States, Japan, France, and Brazil.

Each virtual subject possesses a unique personality. For example, detailed attributes are assigned, such as: "Baek Jung-sook is a diligent homemaker who has witnessed the urbanization of Pyeongtaek, cultivating a vibrant old age through historical exploration and badminton as a meticulous and realistic woman in her 60s." Furthermore, family relationships, hobbies, interests, and cultural backgrounds are precisely configured, including statements like: "She values maintaining a peaceful relationship sharing minor daily routines without major conflicts with her husband."

The application scope of such persona data is exceptionally broad. Beyond machine learning for other artificial intelligence (AI) systems, it can be applied to countless fields such as product preference surveys and public service satisfaction surveys.

Crucially, it boasts the advantage of drastically reducing the time, cost, and risk of personal information leaks associated with surveying real people. To ask 1,000 people about their preference for a specific product requires several times that number of phone connections or face-to-face surveys, whereas utilizing persona data yields immediate responses. Evaluations of various public policies can also be conducted with ease. Moreover, risks like personal information leaks that could occur during human surveys do not exist because these are fictional figures.

This trend of building virtual subject data is expanding across the entire tech sector. Earlier, in June 2024, Chinese tech giant Tencent unveiled a 1-billion-scale virtual subject dataset through its "Persona Hub" project, preceding Nvidia.

[Copyright (c) Global Economic Times. All Rights Reserved.]