Using AI to Decipher North Korean messages: an LLM-based text analysis of KCNA news

Introduction Empirical research on North Korea has often been limited by scarce data, but advances in computational methods now allow systematic analysis of state-produced texts. This study applies large language model (LLM)-based sentiment analysis to English-language articles published by the Korean Central News Agency (KCNA) between 1997 and 2014, treating KCNA as both a propaganda instrument and a strategic communication channel. Methods A corpus of 99,757 KCNA articles was analyzed using the Bidirectional and Auto-Regressive Transformers (BART) model to classify sentiment at the article level. Hypotheses were tested using ordinary least squares regression. Results The findings reveal clear differences in sentiment toward allies and adversaries. Articles referencing the United States, South Korea, and Japan are significantly more negative, while those mentioning China and Russia are more positive. Joint military exercises intensify hostility toward weaker adversaries, particularly South Korea, but constrain overt criticism of stronger adversaries such as the United States. Broader political contexts, including negotiations and cooperative visits, exert weaker influence, while domestic provocations such as nuclear and missile tests are framed positively to reinforce regime legitimacy. Discussion These results underscore North Korea's rational signaling behavior and highlight the dual role of KCNA as both a propaganda instrument and a strategic communication channel. Methodologically, the study demonstrates the utility of LLM-based sentiment analysis in capturing nuanced rhetorical patterns in authoritarian media. Future research should extend the dataset to more recent years and incorporate reciprocal communication from external actors to better understand the dynamics of strategic signaling in Northeast Asia.