[단독]AI가 만든 세계에서 ‘소수자’는 지워진다

The wire material reports that Goldman Sachs estimated the global investment in artificial intelligence (AI) infrastructure reached $1 trillion (1.3 quadrillion won) this year, leading to rapid advancements in AI technology and its wider applications. However, the wire also highlights a concerning phenomenon where AI algorithms are excluding minority groups and causing social distress in virtual worlds.

A study by researchers from Seoul National University found that large language model algorithms, such as ChatGPT and Jemiina, are removing disabled individuals and marginalized groups like sexual minorities from the data used to train them. This exclusion could potentially exacerbate social discrimination and inequality if AI models with biased tendencies are utilized by governments.

The researchers obtained a paper titled "Silent Designers: Large Language Models Excluding Minority and Social Distress" from arXiv, a preprint repository run by Cornell University. The study involved inputting a database of 1 million virtual personas created by Nvidia, a computer chip manufacturer, into an LLM. Out of these 1 million virtual personas, none of them identified as minorities such as disabled individuals, sexual minorities, or those born abroad.

The researchers suggest that more scrutiny is urgently needed before AI models are widely applied in studying human issues or informing government policies.