When artificial intelligence is asked about historical atrocities such as the wartime forced mobilization of "comfort women," the answer you receive can depend heavily on the language you use.

According to a recent analytical report titled "How Language Models Answer Korea's History, Territory, and Culture"—compiled by Koo Rum, co-CEO of data tech firm Big Value—AI models overwhelmingly affirm historical facts when queried in Korean, English, or Chinese. However, when the exact same questions are posed in Japanese, the rate of affirmation plummets to roughly 50%.

The study tested 27 AI models from 12 global developers by inputting 50 questions regarding South Korean history, territory, and culture across four languages. The responses were then evaluated using Anthropic’s Claude Opus 5.5 to measure alignment with historical consensus and the South Korean standpoint.

A Dramatic Drop in Japanese

When asked in Korean, 94.9% of the AI models concurred that Japanese military "comfort women" were forcibly mobilized. Similar high rates were recorded in English (92.2%) and Chinese (95.2%).

In sharp contrast, when prompted in Japanese, only 56.3% of the responses agreed. The remaining 40.7% of responses avoided a direct answer, instead citing "diverse views and ongoing debates" or arguing that "opinions remain divided over the definition of coercion."

A similar trend emerged regarding forced labor during the Japanese colonial period. While Korean, English, and Chinese queries yielded affirmation rates between 93.7% and 97.0%, a Japanese-language query dropped the acknowledgment rate down to 54.8%, with 41.1% of responses remaining ambiguous.

Dataset Disparities and Global Risks

Analysts point to the composition of training datasets as the primary culprit. Because AI developers do not fully disclose their training sources, it is widely suspected that Japanese-language training data heavily reflect conservative or revisionist perspectives held within Japan regarding historical disputes.

While AI-generated responses are not official historical records, experts warn that as global reliance on AI grows, skewed responses in foreign languages could propagate misinformation on a global scale, distorting public perception.

Calls for National-Level Data Initiatives

South Korea, a relative latecomer in the foundational Large Language Model (LLM) race compared to the United States and China, has been ramping up government support for domestic AI development. However, industry insiders note that securing high-quality foreign-language datasets remains a challenge.

With foreign AI models potentially absorbing biased historical data, researchers emphasize that the South Korean government should take proactive measures. Experts argue that building open-source, high-quality foreign-language datasets containing accurate historical and cultural accounts is essential—not only to train domestic models, but to ensure that global LLMs learn objective facts rather than localized historical distortions.

[ⓒ Global Economic Times. 무단전재-재배포 금지]