Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hermit Kingdom Through the Lens of Multiple Perspectives: A Case Study of LLM Hallucination on North Korea

Eunjung Cho, Won Ik Cho, Soomin Seo

TL;DR

This paper addresses the challenge of LLM hallucination and misinformation by evaluating multilingual LLMs on North Korea across English, Korean, and Mandarin Chinese. It constructs a ground-truth dataset with two topic categories—false rumours and lesser-known information—tested on five models using five prompts per topic, totaling $n_T=5$ trials per topic. The evaluation combines quantitative metrics ($R_{RtA}$, $Acc.$, $Con.$) with qualitative analysis, revealing substantial variation in performance across languages and models and highlighting the influence of language choice on perceived truth. The findings underscore the need for targeted data curation, cross-language benchmarks, and nuanced mitigation strategies in high-stakes geopolitical domains where misinformation can have wide-reaching consequences.

Abstract

Hallucination in large language models (LLMs) remains a significant challenge for their safe deployment, particularly due to its potential to spread misinformation. Most existing solutions address this challenge by focusing on aligning the models with credible sources or by improving how models communicate their confidence (or lack thereof) in their outputs. While these measures may be effective in most contexts, they may fall short in scenarios requiring more nuanced approaches, especially in situations where access to accurate data is limited or determining credible sources is challenging. In this study, we take North Korea - a country characterised by an extreme lack of reliable sources and the prevalence of sensationalist falsehoods - as a case study. We explore and evaluate how some of the best-performing multilingual LLMs and specific language-based models generate information about North Korea in three languages spoken in countries with significant geo-political interests: English (United States, United Kingdom), Korean (South Korea), and Mandarin Chinese (China). Our findings reveal significant differences, suggesting that the choice of model and language can lead to vastly different understandings of North Korea, which has important implications given the global security challenges the country poses.

Hermit Kingdom Through the Lens of Multiple Perspectives: A Case Study of LLM Hallucination on North Korea

TL;DR

This paper addresses the challenge of LLM hallucination and misinformation by evaluating multilingual LLMs on North Korea across English, Korean, and Mandarin Chinese. It constructs a ground-truth dataset with two topic categories—false rumours and lesser-known information—tested on five models using five prompts per topic, totaling trials per topic. The evaluation combines quantitative metrics (, , ) with qualitative analysis, revealing substantial variation in performance across languages and models and highlighting the influence of language choice on perceived truth. The findings underscore the need for targeted data curation, cross-language benchmarks, and nuanced mitigation strategies in high-stakes geopolitical domains where misinformation can have wide-reaching consequences.

Abstract

Hallucination in large language models (LLMs) remains a significant challenge for their safe deployment, particularly due to its potential to spread misinformation. Most existing solutions address this challenge by focusing on aligning the models with credible sources or by improving how models communicate their confidence (or lack thereof) in their outputs. While these measures may be effective in most contexts, they may fall short in scenarios requiring more nuanced approaches, especially in situations where access to accurate data is limited or determining credible sources is challenging. In this study, we take North Korea - a country characterised by an extreme lack of reliable sources and the prevalence of sensationalist falsehoods - as a case study. We explore and evaluate how some of the best-performing multilingual LLMs and specific language-based models generate information about North Korea in three languages spoken in countries with significant geo-political interests: English (United States, United Kingdom), Korean (South Korea), and Mandarin Chinese (China). Our findings reveal significant differences, suggesting that the choice of model and language can lead to vastly different understandings of North Korea, which has important implications given the global security challenges the country poses.
Paper Structure (24 sections, 1 figure, 1 table)

This paper contains 24 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. North Korea and misinformation
  4. Large language models, hallucination, and misinformation
  5. Two challenges combined: North Korea misinformation and LLM hallucination
  6. Dataset design
  7. Languages
  8. Topic categories
  9. Experiment setup
  10. Data construction
  11. Topic candidates for each category
  12. Prompt formatting
  13. LLMs
  14. Evaluation
  15. Quantitative
  16. ...and 9 more sections

Figures (1)

  • Figure 1: Workflow Diagram. We first instruct GPT-4 to generate five topic candidates for each of two categories: false rumours (FR) and lesser-known information (LKI). The model is prompted in three languages - English (EN), Chinese (ZH), and Korean (KO) – with ten iterations per language. The generated candidates are then aggregated, de-duplicated, and filtered to remove topics mentioned only once. This refined set is further reviewed by a domain expert, resulting in seven topics for FR and six for LKI. Next, five LLMs are tested on these 13 topics. GPT-3.5, Claude, and Gemini are prompted in all three languages, while Qwen is prompted only in Chinese, and Solar only in Korean. Each model is prompted five times per topic. Finally, the responses are reviewed by two annotators, with verification by a domain expert. The answers are evaluated using both quantitative metrics (accuracy, consistency, refusal to answer (RtA) rate) and qualitative analysis. Further details are available in Sections 3-5.