Why “cheap” threats are meaningful: Threat perception and resolve in North Korean propaganda

Abstract

Threatening propaganda—particularly when extreme and frequent—is often considered “cheap talk.” However, this article argues systematic and comprehensive analysis of such threats can still lend valuable insights. In particular, the aggregate content of threats reveals information about the threat perceptions of the messenger, while the frequency of threats provides information about the messenger’s resolve. To test this theory, I analyze a comprehensive dataset of North Korean propaganda between 1996 and 2018, showing that North Korea systematically issues threats to its adversaries when they engage in joint military exercises or when they take steps, such as the development of missile defenses, that challenge the survivability or deterrent capability of the North Korean nuclear arsenal. Additionally, North Korea’s rhetoric signals its resolve. As the volume of North Korean threats increases, so too does the likelihood that North Korea will engage in military provocations, including nuclear and missile tests.

La propaganda amenazante, especialmente cuando es extrema y frecuente, generalmente se considera "discurso barato". Sin embargo, este artículo sostiene que un análisis sistemático e integral de dichas amenazas puede aportar observaciones valiosas. En concreto, el contenido global de las amenazas revela información sobre las percepciones de la amenaza del mensajero, mientras que la frecuencia de las amenazas proporciona información sobre la resolución del mensajero. Para probar esta teoría, analizo un conjunto detallado de datos de propaganda de Corea del Norte de entre 1996 y 2018, que muestra que Corea del Norte envía amenazas sistemáticamenta a sus adversaries cuando participant en ejercicios militares conjuntos o cuando toman medidas que desafían la capacidad de supervivencia o de dissuasion del arsenal nuclear norcoreano, como el desarollo de defensa con misiles. Además, la retórica de Corea del Norte indica resolución. A medida que las amenazas de Corea del Norte aumentan, también lo hacen sus probabilidades de involucrarse en provocaciones militares, como pruebas nucleares y de misiles.

La propagande menaçante – particulièrement lorsqu’elle est extrême et fréquente – est souvent considérée comme étant de la « conversation libre ». Cependant, cet article soutient qu’une analyse systématique et exhaustive de telles menaces peut tout de même apporter de précieux renseignements. En particulier, le contenu agrégé des menaces révèle des informations concernant la manière dont le messager perçoit ces menaces, alors que la fréquence des menaces offre des informations sur la détermination du messager. Pour mettre cette théorie à l’épreuve, j’analyse un jeu de données complet portant sur la propagande nord-coréenne entre 1996 et 2018 et je montre que la Corée du Nord émet systématiquement des menaces envers ses adversaires lorsqu’ils s’engagent dans des exercices militaires conjoints ou lorsqu’ils entreprennent des démarches, telles que le développement de défenses antimissiles, qui remettent en cause la capacité de survie ou de dissuasion de l’arsenal nucléaire nord-coréen. De plus, la rhétorique de la Corée du Nord signale sa détermination. Plus le volume des menaces nord-coréennes augmente, plus il en va de même pour la probabilité que la Corée du Nord se livre à des provocations militaires, notamment à des essais nucléaires et de missiles.

Keywords:

• North Korea
• nuclear weapons
• cheap talk
• signaling
• international security
• crisis politics
• propaganda

Acknowledgments

The author would like to thank the Stanford Center on International Conflict and Negotiation, the Freeman Spogli Institute, and the Stanford Center for East Asian Studies for their funding and support, as well as Jordan Bernhardt, Daina Chiba, Charles Crabtree, Steven Denney, Justin Grimmer, Scott Sagan, Kenneth Schultz, Michael Tomz, Minh Trinh, Taehee Whang, and the anonymous reviewers for their feedback.

Notes

1 I use “propaganda” to refer to messaging from North Korean state-run media.

2 For example, debates on the deployment of THAAD have often focused on Chinese (and sometimes Russian) objections to the system, dismissing North Korean concerns and noting the probable ineffectiveness of THAAD against a North Korean attack.

3 North Korean leaders may, however, have to remain accountable to a small selectorate.

4 Provocations include missile tests, nuclear tests, exchanges of fire, territorial incursions, and other significant actions that heightened tensions between North Korea and its adversaries (“Database: North Korean Provocations” 2019).

5 For example, Mahdavi and Ishiyama (2020) use reports from KCNA to identify the presence of elites at “inspection tours.”

6 Scholars have studied other topics in North Korean media (Mahdavi and Ishiyama 2020; Merkel 2014; Myers 2010).

7 On open-source intelligence, see Lin-Greenberg and Milonopoulos (2021), Zegart (2010), and Benes (2013).

8 I refer to resolve as time- and issue-dependent.

9 North Korea controls the media, art, literature, and so on, produced and distributed in North Korea through the KWP (Korean Workers’ Party) Propaganda and Agitation Department (Lim 2015).

10 The First and Second Battles of Yonpyong in 1999 and 2002, the 2009 Battle of Daechong, the 2010 sinking of the Cheonan, and the 2010 shelling of Yongypong.

11 Future work could investigate 2018–2020, during which North Korea declared a moratorium on nuclear and long-range missile testing (although several provocations, including short-range missile tests, occurred during this period.) Examining more recent patterns of behavior, particularly in light of increased provocations since 2020, could also be valuable.

12 KCNA publishes on a site hosted in North Korea and a mirror hosted in Japan. They publish essentially the same content. The North Korean-hosted website is available from the United States, but the Japanese site is designed to be accessible only from Japanese IP addresses. The Japanese-hosted site has no search function, while the North Korean-hosted site has very limited searching. The Japanese-hosted site retains archives back to the late 1990s, while the North Korean-hosted site retains minimal archives, known to be edited.

13 Rarely, articles are available in English and not Korean.

14 KCNA is accessible online, in print, on television, and by radio. While elites are able to access KCNA online or on television, radio is the primary news source for much of the public. KCNA is one of only two sources of North Korean radio news.

15 They, at least, de facto communicate to foreign audiences.

16 Moreover, many forms of North Korean media are banned in South Korea under the National Security Law

17 See Online Appendix 1.1 for an illustration of how topic modelling works in this context. I utilize the LDA from the topicmodels package in R. This package requires the user to set a parameter k for the topic number. I present the model with k = 10; this is the highest number of topics for which a) all topics made sense under clear labels, and b) no topics seem to be repeated. For k < 10, most models still create a foreign policy topic of some sort; however, articles without clear threats are included in the topic at these lower values. For k > 10, the LDA performs well at keeping threats within a single topic. However, other topics begin to be repeated, such as those discussing leadership. I test the LDA with all values of k between 2 and 20. For each LDA, I qualitatively examine articles with the highest and lowest gamma values for each topic to assess appropriateness. For k = 10, I perform additional robustness checks.

18 I use stop words from R’s tidytext package. I don’t stem, in order to assess differences between word forms and tenses. This is consistent with Denny and Spirling (2018) I do not add stop words to the generic set, although I remove “states.” Some words are frequent but not meaningful. For example, because articles about unrelated topics will often conclude with praise for the North Korean leader, “Kim Jong Il” and “Kim Jong Un” are frequent and not always meaningful. In addition, KCNA often references other publications, like Minju Joson and Rodong Sinmun. KCNA re-publishes articles from these outlets, so analysis of KCNA should provide sufficient insight into the broader set of outlets. Rodong Sinmun is the official publication of the Workers’ Party of Korea; along with KCNA, it is a main publication. However, there is only limited access to Rodong Sinmun archives, with select articles available since 2002. Minju Chosun is associated with the cabinet and is less frequently used. North Korean officials also make rare public statements, including in international forums, which are usually republished in KCNA. Very rarely, North Korean journalists, officials, or researchers may publish outside of North Korea.

19 I use LDA instead of other models, such as Structural Topic Modelling (STM), Hierarchical Dirilecht Process (HDP), Correlated Topic Models (CTM) and Dynamic Topic Models (DTM). STM operates similarly to LDA but uses meta-data to guide assignment of words to latent topics. Since I do not predict significant variation in the type of language used to convey threats over time, and since there is no author information available for most articles, STM is not appropriate. DTM takes into account the order of documents; this would be most relevant if over-time evolution were critical. HDP does not require a topic number selection and can be appropriate with large numbers of topics. Since the topics of interest here are bounded, LDA is preferable. CTM is most useful with large numbers of topics or heavily correlated topics. Since, here, topics are limited and discrete, CTM is not appropriate.

20 Topic labels are provided by the author.

21 See Online Appendix 1.1 for details, including beta values.

22 Because Richey (2019) excludes the word “imperialist,” many threats are excluded from his data collection.

23 This assumption may not work with other propaganda, but the very high hostility level here makes this assumption plausible.

24 Articles are sorted into the topic for which they have the highest gamma value, which reflects the probability a document matches a topic. Assigning each article to its closest topic does create complications for articles discussing multiple topics. However, this is uncommon given the short average article length. I sample 67 threatening articles.

25 See Online Appendices 1.3 and 1.4 for more on trigger and responses types. For example, North Korea’s threatened responses are often severe, with the majority of responses referencing total victory over adversaries, full-scale war, or nuclear strikes.

26 See Online Appendix 1.2 for topic trends over time.

27 “Puppet” or “puppet army” commonly references the South Korean military’s relationship with the United States. “Imperialists” is an insult leveraged primarily against the United States (as well as Japan). It references the “aggression” and intentions for an expansionist “war” North Korea argues are behind US-South Korean “joint military exercises.” “Moves” or “military moves” is often used interchangeably with “(joint) military exercises.” North Korea accuses the United States of making “moves” toward “war” or using “exercises” to gear up for an “attack.” It frequently references US and South Korean “military forces” and “arm[ies]” as well as its own military.

28 Many of these words can be used in a variety of contexts, such as “force[s]” and “policy” but often appear as modifiers for “nuclear.” The remaining unigrams are understandably common; they refer to major actors, places, or concepts relevant to North Korea’s security.

29 On patterns in the use of preemption as a legal justification, see Sukin and Weiner (2022).

30 See Online Appendix 2.3 for provocation frequency.

31 See Online Appendix 2.3 on threat and provocation rates over time.

32 See Online Appendix 2 on the co-occurrence of threats and provocations.

33 The weekly mean is 11 threats.

34 See Online Appendix 2 for robustness tests.

35 These increase significance in most models, so normal standard errors are presented here, since they pose a harder test.

36 Because of the inclusion of the year fixed effect, Newey-West standard errors are not applied to Model 7.

37 See Online Appendix 3.

38 For work on rhetoric in Russia state-run media, see Jonsson and Seely (2015) and Becker (2004).

39 While China does sometimes use this aggressive style of rhetoric in its state media, scholars have noted that foreign-facing Chinese propaganda has also aimed to promote Chinese soft power and highlight Chinese technological advancements. Even in the context of Taiwan, Chinese propaganda has occasionally emphasized unification with a framing that centers historical ties between China and Taiwan. See, for example, Brady (2015a), Brady (2015b), and Edney (2012).