On 1 June 2021, Xinhua published a summary of comments made by Xi Jinping 习近平 to a collective study session of the Chinese Communist Party (CCP) Politburo on China's “international communications work.”Footnote 1 Among the comments was a call for communicating a more “trustworthy, lovable and respectable” image of China.Footnote 2 This led to global speculation about possible changes in China's sometimes pugilistic public diplomacy.Footnote 3 The People's Republic of China's (PRC) image in Western countries had been increasingly negative,Footnote 4 and this seemed to be definitive recognition that the style of “wolf warrior diplomacy” (WWD), which had become prevalent in recent years, was backfiring. WWD, so named after a nationalistic movie franchise and a label which PRC diplomats dislike,Footnote 5 is characterized by robustly defending PRC policies when criticized abroad, emphasizing the hypocrisy of foreign critics and standing up to “the West” (particularly the United States), often using insulting language.Footnote 6 Perhaps Xi wanted to tame the wolves.
The “lovable” comment grabbed the headlines, but the full report about the study session was more complex. It was not immediately clear that it was meant to change the PRC's communication in a straightforward or uniform way, or indeed at all. Some contemporaneous reporting of Xi's comments indicated scepticism that the tone of China's public diplomacy would change.Footnote 7 Analysts noted that the report came with language about a “public opinion struggle” that suggested a delineation of friends and enemies, with the former praised and the latter to be made to “understand” China better.Footnote 8 Observers of PRC foreign policy communication often note the tendency for the message to change with the audience.Footnote 9 The full statement's emphasis on increasing PRC “discourse power” indicated a continued focus on influencing the global conversation commensurate with China's underlying material power.Footnote 10 The expert chosen to address the study session was Professor Zhang Weiwei 张维为 of Fudan University 复旦大学, who has a track record of calling for China to be more outwardly confident about the PRC's governing system and to amplify the faults of “the West.”Footnote 11 Xi himself has long called for China's diplomats to take a more assertive tone.Footnote 12 Perhaps this announcement was not about taming the wolves and was instead about reinvigorating the current approach of being lovable for friends and assertive with enemies.Footnote 13
Clearly, the study session was an important event for PRC external communications work.Footnote 14 However, as noted thus far, the summaries issued by Xinhua contained mixed messages, meaning that an empirical approach may help us to understand the meeting's outcome more accurately. With this research report, we take advantage of the timing of the meeting's summary report to lend new empirical evidence to discussions about PRC public diplomacy and WWD. Specifically, based on a text-scaling technique applied to over 200,000 English tweets, we measure the WWD rhetoric of China's official and affiliated Twitter accounts before and after the meeting was first made public on 1 June by Xinhua. This approach allows us to see whether and how the collective study session advice was implemented. Since Twitter is blocked in China itself, the messages that PRC officials and entities post there will be viewed primarily by foreign audiences.
For this research report, we do not develop a detailed theory to explain China's external communications or public diplomacy. Rather, we approach one aspect of it with an empirical lens. Specifically, we investigate heterogeneous effects between diplomatic Twitter accounts based in OECD and non-OECD countries in the aftermath of the 1 June announcement. This distinction is made because OECD membership broadly comprises wealthy democratic states, often a key target of WWD tactics. In the months following the announcement, PRC diplomatic accounts in the OECD did moderate their tweets to become slightly friendlier compared to their counterparts in non-OECD countries, but we do not detect from this data a major substantive break in PRC communications strategies as a result of the session.
Our findings have relevance for important theoretical debates. They speak to scholarship about principal-agent relationships in China's foreign policy apparatus and the role of domestic political incentives,Footnote 15 nationalism and PRC foreign policy,Footnote 16 and the country's external image management and propaganda.Footnote 17
Context: China's Image Abroad and WWD
The PRC pays sustained attention to how China is perceived abroad.Footnote 18 Even prior to taking power, the CCP was keen to manage its image to persuade foreign sympathizers.Footnote 19 PRC image crafting has proactive and reactive dimensions.Footnote 20 Proactive efforts include showing a peaceful and non-threatening China while highlighting its achievements.Footnote 21 Reactive efforts include responding to foreign criticism of PRC policies.Footnote 22 Messages are communicated via party-state media like Xinhua or CGTN, statements from diplomats, PRC-friendly think tanks or institutes, and/or foreign public relations firms. At times, the “message” is coercive, as when the PRC attempts to sanction, intimidate, deny access to or otherwise silence those who criticize its policies.Footnote 23 In recent years, Western social media has become important in China's external communication strategies as the PRC fears ceding the public opinion battle in platforms controlled from the West.Footnote 24
WWD emerged in this context. In 2019, Zhao Lijian 赵立坚, a PRC diplomat in Pakistan who had previously served in the US, began to be noticed for his Twitter account.Footnote 25 His tweets played up his love of Pakistan, defended PRC policy and challenged American hypocrisy. Zhao and others who took similar rhetorical approaches were promoted, and “diplomats across the [foreign] ministry noticed that the shift in tone was being rewarded – just as they too began opening their own Twitter accounts” around 2019.Footnote 26 The confluence of Chinese diplomats and state media bolstering their global social media profiles, declining US–China relations, increased foreign attention to China's repression of Uyghurs, the PRC's focus on “discourse power” to shape international narratives, domestic nationalism, Xi's stated preference for assertiveness, visible career progression for diplomats like Zhao, and China's growing material power all combined to set the context for the birth of WWD.Footnote 27 Did the 1 June report on the collective study session change the approach in noticeable ways? It is that question we try to answer in the next section.
Data, Methods and Results
The messages that diplomats and embassies post on Twitter are indicative of the PRC's communication strategy. Relying on a list of Chinese diplomatic accounts (ambassadors, embassies, consulates, Ministry of Foreign Affairs accounts, or high-level staffers) and state-backed media accounts, we retrieved the tweets of all accounts that posted at least once between 1 January 2021 and 7 October 2021.Footnote 28 We limited the analysis to English language tweets and excluded retweets. Next, we extracted a random sample of 1,000 tweets. We manually coded each tweet for whether it is “friendly,” “WWD” or “neutral.” Based on this hand-coding, we identified “seed words” that represent two ends of a unidimensional scale ranging from “friendly” to “WWD.” Afterwards, we used Latent Semantic Scaling (LSS) to score the text of all 200,000 tweets. Generally, LSS assigns a score to each term in the text corpus based on the semantic similarity with the seed words and then allows a predicted text score based on frequencies of these terms in each document.Footnote 29
Table 1 reports the 100 terms with the lowest and 100 terms with the highest scores. Many of these terms, identified through the small set of 23–25 seed words (see Table A1 in the online Appendix), make intuitive sense and speak to the method's validity. Terms like “accusations,” “fabricated,” “disinformation,” “smear” and “oppression” appear in the category of the most unfriendly terms, highlighting the WWD tone. Words such as “gratitude,” “brotherly,” “donation,” “thank” and names of countries with amicable relations with the PRC received very high scores.
Moving from keywords to texts, Tables A2–A3 list the “friendliest” and most “WWD” tweets based on the LSS scores. Again, higher values imply a “friendlier” tone. For instance, the following tweet, portraying foreign critics as liars, has one of the lowest LSS scores:
@ChinaConSydney (Chinese Consulate General in Sydney), 7 March 2021: “The claim that there is genocide in #Xinjiang couldn't be more preposterous. It is just a rumor fabricated with ulterior motives, and a lie through and through.”
Conversely, one of the highest-scoring tweets portrays China as a friend and partner:
@PDChina (People's Daily, China), 6 May 2021: “President Xi Jinping on Thursday exchanged congratulatory messages with Brunei's Sultan Haji Hassanal Bolkiah on the 30th anniversary of the establishment of diplomatic ties between the two countries.”
Using this data, we exploit the timing of the Xinhua announcement to identify whether the study session report led to meaningful changes in the tone of PRC public diplomacy. We considered several methodological options. We opted not to use a regression discontinuity in time (RDiT) approachFootnote 30 or a simple two-period difference-in-difference design using the date of the study session summary as a cut-off. Since there are no comparison (untreated) units, we cannot recover an estimate of an average treatment effect (ATE) or an average treatment effect of the treated (ATT). In other words, a simple pre-post comparison is very likely to be spurious to any other temporal factors that might influence tweeting.
Rather, given the aforementioned mixed and complex signals in the May 2021 collective study session announcement, and because there is no true “non-treated” comparison group (such as a group of accounts that would not have seen the announcement but continued to tweet), we compare two groups of accounts that we think should have been differentially influenced by the announcement. Consistent with the notion that the PRC's external communication advances different messages for different audiences, we evaluate if there is a differential effect of the announcement on tweets from accounts based in the OECD (mostly wealthy democracies) versus those from non-OECD countries (broadly, states in the Global South and state-affiliated media accounts based in China). PRC diplomacy has a long history of building friendly relations with countries in the developing worldFootnote 31 and at times it portrays itself as a leader of this group vis-à-vis the West.Footnote 32 In an August 2013 national propaganda work conference, Xi himself stressed the need to spread PRC viewpoints more effectively among “developing states.”Footnote 33 Underlying the WWD approach is an anti-imperialist disposition that is more critical of Western states and allies and more sympathetic to states perceived to not be in this category. We use the OECD/non-OECD distinction as a proxy to capture these categories to see if diplomats in these countries reacted differently to the collective study session.
Our approach is therefore based on an expectation of heterogeneous treatment effects of two “treated” groups, rather than on a “treated” group and a control group. We expect this heterogeneity based on the underlying assumption that PRC Twitter accounts will react to stimuli with the understanding that the PRC crafts its external diplomacy with different messages for different audiences. This comparison necessitates a parallel trends assumption between tweets from OECD and non-OECD accounts, which we substantiate below and in the online Appendix.
Using tweets as the unit of analysis, we assign an indicator status based on the account's location. We generate a binary variable and assign the value of “1” to tweets from accounts in OECD countries. We then create a binary temporal variable where we assign a value of “1” to the time period after the collective study session report. Our data contain a total of 10,066 tweets from accounts in OECD countries, 5,904 prior to the 1 June announcement and 4,162 after. Likewise, we identify 190,542 tweets from accounts in non-OECD countries and state-backed media accounts, with 79,997 of these prior to the 1 June announcement and 110,545 after. These data allow us to employ the classic difference-in-difference equation to evaluate the differential effect before and after the summary report comparing the OECD and non-OECD accounts.Footnote 34 Note that we train the LSS models on all tweets, while the difference-in-difference design is limited to a symmetric time window ranging from ±1 day to ±115 days.
Starting with a descriptive overview of temporal developments, Figure 1 shows the daily average scores for all accounts by diplomats, ambassadors, institutions and staff. Vertical bars show 95 per cent confidence intervals. Overall, we do not observe a consistent change in WWD scores around the announcement in June. Yet the plot reveals considerable and meaningful variation over time. Late March 2021 is a clear outlier, which directly corresponds with a highly contentions meeting on 19 March in Alaska between top US and Chinese diplomats. During the meeting, PRC diplomats accused the US of encouraging countries “to attack China,” while US representatives listed criticisms of several PRC policies.Footnote 35 Many PRC accounts tweeted about the meeting, with the WWD scores during this time lending credence to the coding and validation.
Moving forward, in the main models we do not include any covariates owing to the potential of post-treatment bias. We standardize our measure by account, which should absorb much of the unobserved heterogeneity at the account level in our main models and ease comparative interpretation. In other words, if some accounts have a high amount of variation in their tone, and others have low variation, the same absolute difference would equate to a larger proportion of the standard deviation for the low variance account vis-à-vis the high variance account. Standardizing the measure ensures we are comparing the degree of variance relative to the account. When comparing OECD and non-OECD diplomat accounts, we see that the standard deviation of the score is almost identical, 1.403 to 1.396. As such, it is no surprise that we find results in the Appendix that are substantively consistent when evaluating models that use the non-standardized scores (see Table A7).
The main results are presented in Figure 2 and in Table A4 (for ±100 days) in the Appendix. Each point estimate and error bar shows the DiD estimate for varying specific time windows. Running models on a window increasing from ±1 to ±115 days allows for a visual inspection of the robustness of our findings. Values above 0 mean that OECD accounts are friendlier than non-OECD accounts. The vertical lines show 90 per cent and 95 per cent confidence intervals.
The results show that OECD accounts seemed to become more positive immediately after the announcement (±1–±4 days). Owing to the small sample size of tweets, these large estimates are not immediately statistically significant. However, a clear pattern emerges and when the window is increased, where after about 1.5 months we observe consistently positive DiD estimates. The coefficients usually range between 0.1 and 0.15 standard deviations. This small but relatively consistent change is also visible when considering distinct classes of accounts. We make comparisons for tweets from all accounts (diplomatic and state-backed media, where state-backed media accounts are considered in the non-OECD group), tweets only from diplomatic accounts, tweets from institutional accounts (embassies or consulates), and tweets only from ambassadorial and staffer accounts (Figure A1). Based on these results, we conclude that accounts based in the OECD became somewhat more positive after the study session, relative to accounts in non-OECD countries. The time lag may be owing to the amount of tweets needing to become large enough to reliably recognize shifts, infrequent tweeters taking time to build up a measurable corpus after the announcement, or perhaps more detailed internal guidance that followed in subsequent weeks after the session. However, the data also do not reveal a dramatic strategy change resulting from the session.
What does a difference of 0.1–0.15 standard deviations mean in practice? A comparison of two tweets that are 0.12 standard deviations apart in their WWD score, the estimated difference for ambassadors in OECD versus non-OECD countries, illustrates this difference. A tweet at the mean is not aggressive but hints at some hypocrisy by foreigners:
@YXiusheng, ambassador in Barbados, 27 July 2021: “Study of origin of COVID-19 calls for international collaboration, not blame.”
A tweet 0.12 standard deviations more friendly (less WWD) retains the positive tone about China but sheds the implicit criticism:
@AmbLiuQuan, ambassador of Suriname, 1 February 2021: “More and more countries start to approve the use of Chinese vaccines.”
The shift is subtle but real. In our first set of robustness checks (online Appendix Section B), we introduce both country and account fixed effects to deal with variation not already absorbed by standardization of the score by account. Our second set of robustness checks adds tweet-level and account-level control variables, including the number of words in the tweet, the number of followers of the account on the day of the announcement, the number of retweets and the number of “likes” received. Finally, we test models with a non-standardized WDD score. The results do not depend on the inclusion or exclusion of controls and fixed effects, and they are substantively consistent when we use unstandardized measures of our textual measure.
Returning to the parallel trends assumption, recall that this is a potential threat to our inferential strategy, as post-statement differences between OECD and non-OECD accounts could be driven by other, external factors which have a differential effect. We examine the parallel trends assumption by considering only tweets between 30 days before and after the announcement. For each comparison, we plot the raw data (with LOESS pre- and post-treatment trends) in Figure A2. For some comparator groups, it is conceivable (but potentially unknowable) that the assumption of parallel pre-treatment trends is violated. For example, collective study sessions feature expert input in advance, so perhaps there was awareness of the announcement and a corresponding change of behaviour among some tweeters before it happened; however, we cannot say with certainty. If anticipation effects are in the same direction as the treatment effect, this pre-trend could lead to unadjusted difference-in-difference results understating the magnitude of the treatment.Footnote 36 This would mean that the true treatment effect of the announcement is actually larger than what we report. On the other hand, if the pre-trend was endogenous, it would mean that the study session made its announcement in response to tweets that were becoming less WWD; however, we find this unlikely given that the Party leads and the ministry and media follow.Footnote 37 Finally, any pre-trend could also suggest that our treatment effect is simply picking up some spurious correlation. We cannot completely discount that any potential pre-trend we see for OECD accounts is partially driven by some unobserved event but, overall, we think the parallel trends assumption holds reasonably well.
Conclusions
In sum, the May 2021 collective study session of the CCP Politburo and subsequent June 2021 announcement did have some effect on PRC Twitter diplomacy. Over time in OECD countries, some of which are often the target of ire for WWD, diplomats apparently softened their tone in comparison to the tone of accounts in non-OECD countries. However, we do not interpret from this a dramatic shift in PRC communications strategy as a result of the study session and, indeed, the session appears to have reaffirmed some pre-existing elements as prominent WWD purveyors continue to tweet and are promoted. This speaks to the mixed messages in the summary statement as well as the tendency in PRC external communications to cater the message for the audience.
We recognize that this brief research report has limitations. The time period is right-censored, meaning that it cannot capture ongoing changes in the PRC's public diplomacy. The analysis is limited to Twitter and does not discuss other transnational social media platforms.Footnote 38 Nor does it capture the effectiveness of the messages on viewers’ attitudes.Footnote 39 Finally, we cannot entirely discount that some pre-trend influenced the findings.
Nevertheless, we maintain that these findings are useful to scholars in at least three main ways. First, they illustrate nearly in real time the responsiveness of the PRC foreign policy apparatus, thus lending a data point to scholarship about principal-agent dynamics and bureaucratic cohesiveness in Chinese foreign policy.Footnote 40 Second, they illustrate empirically an effort by the Party leadership to shape the style of nationalism portrayed abroad. Our findings show a responsiveness by diplomats to changing guidance, which contributes to debates about how much the Party can control, direct or tamp down nationalistic narratives in its foreign policy, or how much it is hostage to a nationalistic approach.Footnote 41 Third, the report sheds light on China's external communication strategies, further reinforcing that different audiences are exposed to different messages as the PRC aims to accustom the world to its increasing power.Footnote 42
Supplementary material
To view supplementary material for this article, please visit https://doi.org/10.1017/S0305741022001722.
Conflicts of interest
None
Data availability statement
Replication materials are available on Harvard Dataverse at https://doi.org/10.7910/DVN/FR2ZPL. The authors declare the collection of tweets for this paper was reviewed and approved by the Office of Research Ethics at University College Dublin.
Samuel Brazys is associate professor in the School of Politics and International Relations at University College Dublin in Dublin, Ireland. His research and teaching focus on the nexus between international political economy and development. His recent work has been published in Review of International Organizations, International Studies Quarterly and the Journal of Development Studies. He is editor of the Journal of International Development.
Alexander Dukalskis is associate professor in the School of Politics and International Relations at University College Dublin in Dublin, Ireland. His research and teaching focus on authoritarianism, human rights and Asian politics. His most recent book, Making the World Safe for Dictatorship, was published by Oxford University Press in 2021. In 2020–2021, he was a Wilson Center China fellow.
Stefan Müller is assistant professor and Ad Astra fellow in the School of Politics and International Relations at University College Dublin in Dublin, Ireland. His research and teaching focus on political representation, political communication and public opinion. His recent work has been published in the American Political Science Review, The Journal of Politics and Political Communication. He is co-author of quanteda, an R package for managing and analysing textual data.