COVID-19 Increased Censorship Circumvention And Access To Sensitive Topics In China

Keng-Chi Changa, William R. Hobbsb, Margaret E. Robertsa,1, and Zachary C. Steinert-Threlkeldc

aDepartment of Political Science, University of California San Diego, La Jolla, CA 92093; bDepartment of Human Development and Department of Government, Cornell University, Ithaca, NY 14850; cDepartment of Public Policy, Luskin School of Public Affairs, University of California Los Angeles, Los Angeles, CA 90095

This manuscript was compiled on July 1, 2021

1 Crisis motivates people to track news closely, and this increased en- pare information seeking during crisis in China to other countries that 34 2 gagement can expose individuals to politically sensitive information had similar COVID-19 outbreaks. To draw a comparison, we investi- 35 3 unrelated to the initial crisis. We use the case of the COVID-19 gate the same patterns in countries with no censorship or in authoritar- 36 4 outbreak in China to examine how crisis affects information seeking ian regimes where the platforms we study are not censored that also 37 5 in countries that normally exert significant control over access to experienced large outbreaks of COVID-19 cases soon after China. 38 6 media. The crisis spurred censorship circumvention and access to in- Consistent with other work on information seeking during lockdown 39 7 ternational news and political content on websites blocked in China. in democracies (8), we find higher levels of engagement with online 40 8 Once individuals circumvented censorship, they not only received news media and online entertainment generally, though importantly 41 9 more information about the crisis itself but also accessed unrelated not a similar pattern of users seeking information about sensitive po- 42 10 information that the regime has long censored. Using comparisons litical topics unrelated to the crisis. Together, these findings suggest 43 11 to democratic and other authoritarian countries also affected by early that crisis information seeking provides a gateway to sensitive content 44 12 outbreaks, the findings suggest that information seeking spurred by in authoritarian contexts. This gateway can be especially pronounced 45 13 crisis can undermine censorship and provide abrupt access to previ- when a regime has previously censored a large amount of political in- 46 14 ously hidden content. formation and circumvention tools provide access to a wide variety of 47 current and historical censored content. Previously hidden, accumu- 48 COVID-19 | CHINA | CENSORSHIP | TWITTER lated, and potentially regime-damaging content is then more likely to 49 be discovered and consumed during crisis. 50

This analysis contributes to the literature on the impacts of crisis 51 1 cholars have long predicted that during crises or uncertain time on authoritarian resilience (9–11). Although authoritarian regimes 52 2 Speriods, people will rely more on mass media for information may have logistical advantages of responding to crisis (12), crisis 53 3 relevant to their own safety and spend more time seeking out infor- can still undermine the censorship apparatus, exposing individuals to 54 4 mation (1). Increased attention to media during crisis has been shown long-censored information. While undermining of censorship may 55 5 empirically in democracies, such as during democratization in East- not have immediate ramifications, increased exposure to unfiltered 56 6 ern Europe (2), during the eruption of Mt. St. Helens (3), and im- information and capabilities of censorship evasion could have a long 57 7 mediately after the September 11 terrorist attacks (4–6). Increased term impact on the information environment and could undermine 58 8 attention to the media present opportunities for large changes in opin- trust in government information (13, 14). 59 9 ion or political socialization (2, 7), and crisis disruptions can also shift 10 attention toward entertainment due to lack of mobility and boredom Crisis is a Gateway to Censored Information 60 11 (8). In many authoritarian countries, traditional and online media 61

12 This paper identifies another effect of crisis: abrupt exposure to limit access to information (15–18). While this control is im- 62 13 prior sensitive information blocked by governments. We examine perfect, studies have shown that media control in autocracies 63 14 the effect of crisis on information seeking in highlyDRAFT censored environ- 15 ments by studying the impact of the COVID-19 public health crisis on 16 censorship circumvention in China. In January and February of 2020, Significance Statement 17 COVID-19 cases in China were spiking, official news sources were 18 slow to acknowledge the crisis, and many regions of China locked We study the impact of crisis on information seeking in au- 19 down. Using a variety of measures of Twitter and Wikipedia data, thoritarian regimes. Using digital trace data from China 20 both of which are inaccessible within China, we show large and sus- during the COVID-19 crisis, we show that crisis motivates 21 tained impacts of the crisis on circumvention of censorship in China. citizens to seek out crisis-related information, which subse- 22 For example, the number of daily, geo-locating users of Twitter in quently exposes them to unrelated and potentially regime- 23 China increases by up to 40% during the crisis and is 10% higher damaging information. This gateway to both current and 24 long-term, while politically sensitive accounts gain tens of thousands historically sensitive content is not found for individuals in 25 of excess followers, up to 3.8 times more than under normal circum- countries without extensive online censorship. While infor- 26 stances, and these followers persist one year after the crisis’ end. mation seeking increases during crisis under all forms of gov- ernance, the added gateway to previously unknown and sen- 27 Moreover, beyond information seeking about the crisis itself, we sitive content is disproportionate in authoritarian contexts. 28 find that information seeking across the Great Firewall extended to 29 information the Chinese Communist Party (CCP) has long censored, K.C., W.H., M.R., and Z.S-T. designed research, performed research, analyzed data, and 30 including information about sensitive historical political events and wrote the paper. 31 leaders. The authors declare no competing interest. 32 Although just one of many crises, the global nature of the COVID- 1 33 19 crisis makes this case a unique and important opportunity to com- To whom correspondence should be addressed. E-mail: [email protected]

PNAS | July 1, 2021 | vol. XXX | no. XX | 1–30 64 has large effects on the opinions of the general public and the trustworthiness of their government. While the impact on the 115 65 resilience of authoritarian regimes (19–25), even though there autocrat may not be immediate and could be outweighed by 116 66 are moments when it can backfire (26–32). Evidence from a successful, rapid government response to the crisis, gradual 117 67 China suggests that media control may be effective in part be- accumulation of censored information over time could have 118 68 cause individuals generally do not expend significant energy negative impacts on authoritarian resilience. 119 69 to find censored or alternative sources of information.*

70 While many have studied the impact of information control The COVID-19 Crisis in China 120 71 in normal times in authoritarian regimes, less is known about On December 31, 2019, officials in Wuhan, China confirmed 121 72 information seeking during crisis. In democracies, informa- that a pneumonia-like illness had infected dozens of people. 122 73 tion seeking intensifies during crisis, increasing consumption By January 7, 2020, Chinese health officials had identified the 123 74 of mass media. Ball-Rokeach and Defleur (1) describe a model disease – a new type of coronavirus called novel coronavirus, 124 75 of dependency on the media where audiences are more reliant later renamed COVID-19. By January 10, the first death from 125 76 on mass media during certain time periods, especially when COVID-19 was reported in China, and soon the first case of 126 77 there are high levels of conflict and change in society. These COVID-19 was reported outside of China, in Thailand. As 127 78 findings are largely consistent with research on emotion in of December 2020, COVID-19 has infected over 91,000 peo- 128 79 politics, which concludes that political situations that produce ple in China with over 4,500 deaths, and at least 73.5 million 129 80 anxiety motivate people to seek out information (34). While † people worldwide with over 1.6 million deaths. 130 81 in normal times information seeking is strongly influenced by While initial reports of COVID-19 were delayed by offi- 131 82 pre-existing beliefs, several studies have suggested that crisis cials in Wuhan (39), Chinese officials took quick steps to con- 132 83 can cause people to seek out information that might contradict tain the virus after it was officially identified and the first 133 84 their partisanship or worldview (7, 35), although they may pay deaths reported. On January 23, 2020, the entire city was 134 85 disproportionate attention to threatening information (36). placed under quarantine – the government disallowed trans- 135 86 Similar patterns may exist in authoritarian environments. portation to and from the city and placed residents of the city 136 87 Because the government controls mass media, citizens aware on lockdown (40). The next day, similar restrictions were 137 88 of censorship may not only consume more mass media that placed on 9 other cities in Hubei province (41). While Hubei 138 89 is readily available during crises, but also seek to circumvent province and Wuhan were most affected by the outbreak, cities 139 90 censorship or seek out alternative sources of information that all over China were subject to similar lockdowns. By mid- 140 91 they may normally not access. For example, during the SARS February, about half of China – 780 million people – were liv- 141 92 crisis in China in 2003, Tai and Sun (37) find that people in ing under some sort of travel restrictions (42). Between Jan- 142 93 China turned to SMS and the Internet to gather and corrob- uary 10 and February 29, 2020, 2,169 people in Wuhan died 143 94 orate information they received from mass media. Cao (38) of the virus (43). 144 95 shows an increase in censorship evasion and use of Twitter 96 from China during “regime-worsening” events, such as wors- The Effect of Crisis on Information Seeking and Cen- 145 97 ening of trade relations between the U.S. and China and the sorship Circumvention 146 98 removal of Presidential term limits in the constitution in 2018. 147 99 Outside of facilitating access to information about the cri- We use digital trace data to understand the effect of the 148 100 sis, evasion of censorship during crisis could undermine the COVID-19 crisis on information seeking. Table 1 summa- ‡ 149 101 censorship regime by giving information to consumers not just rizes the empirical tests conducted in this paper. First, we show that the crisis increased the popularity of virtual private 150 102 about the crisis itself, but also about informationDRAFT that has long 151 103 been censored. This phenomenon, a “gateway effect,” has network (VPN) applications, which are necessary to jump the 152 104 been shown in the context of censorship of entertainment (30), Great Firewall, downloaded on iPhones in China. We also 153 105 where consumers are motivated to circumvent censorship for show that the crisis expanded the number of Twitter users in 154 106 one reason, but in doing so are exposed to unrelated sensitive China, which has been blocked by the Great Firewall since 155 107 political information. Crisis might lead to a similar, or even 2009. The crisis further increased the number of page views 156 108 more pronounced, “gateway effect” because people may be of Chinese language Wikipedia, which has been blocked by the 157 109 more likely to seek out political information during crisis than Great Firewall since 2015. We also show that the areas more 158 110 those seeking to circumvent censorship for entertainment pur- affected by the crisis – such as Wuhan and Hubei Province – 159 111 poses. Anxiety about the epidemic, perhaps especially when were more likely to see increases in circumvention. 160 112 accompanied by boredom during quarantine and lockdown, Next, we show that the increase in circumvention caused by 161 113 could lead consumers of information to explicitly seek out in- the crisis not only expanded access to information about the 162 114 formation that has long been censored to better understand the crisis, but also expanded access to information that the Chi- nese government censors. On Twitter, blocked Chinese lan- 163 *Stockmann (2012) (23) provides evidence that consumers of newspapers in China are unlikely guage news organizations and exiled dissidents disproportion- 164 to go out of their way to seek out alternative information sources. Chen and Yang (2019) (33) provided censorship circumvention software to college students in China, but found that ately increased their followings from mainland China users. 165 students chose not to evade the Firewall unless they were incentivized monetarily. Roberts † (2018) (25) provides survey evidence that very few people choose to circumvent the Great Source: New York Times, December 15, 2020. https://www.nytimes.com/interactive/2020/ Firewall because they are unaware that the Firewall exists or find evading it difficult and world/coronavirus-maps.html bothersome. ‡Replication materials will be posted on Dataverse.

2 | Chang et al. 166 On Wikipedia, sensitive pages such as those pertaining to

167 Chinese officials, sensitive historical events, and dissidents Fig. 1. Download Rank of iPhone Application in China: Facebook, Twitter, and 168 showed large increases in page views due to the crisis. Last, Wikipedia. Data from AppAnnie. 169 the fourth subsection shows that these dynamics do not oc- 170 cur on Wikipedia in countries with similar crisis but where VPN Ranking 171 Wikipedia is uncensored.

Wuhan Table 1. Empirical Tests Lockdown

Question Test ● ●●●●●●●●●● 1. Do individuals circumvent cen- VPN ranking; increased use of ● ●● ●● ● ●● ● ●●●● Rank ●●●●●●● ●●●●● ● sorship more during crisis? blocked services; new Twitter ●●● ●● ● ● ●● ●●●● ●●●●●● ●●● ● ●●●●●●● ● ●●● ●● ●● users. ●●●●●●●●● 2. Do individuals access crisis in- Wikipedia traffic about current formation? leaders; new mainland China fol- lowers for certain account types. Jan 01 Feb 01 Mar 01 3. Do individuals access non- Wikipedia traffic to blocked pages; Date (2020) crisis sensitive information? new mainland China followers for activists and foreign political fig- ures. Facebook, 4. Do these same dynamics occur Wikipedia page views in German, App Ranking in democracies and less censored Italian, Persian, and Russian. 0 environments? Wuhan 200 Lockdown

●● ●●● ●●● ●●● ●●●● ● ●● ●● ● ●●●●●●●● ●●●●●●●●●● ●● 400 ●● ●● ● Rank ● ● ● ● 172 ●●● ●● ● ●● Crisis Increased Censorship Circumvention. We show ● ● ●● ●● ●● ● ●● ●●● ● ●● ●●●●●●●●● 173 that censorship circumvention increased in China as a result 600 ● ● 174 of the crisis using data from application analytics firm AppAn- 175 nie, which tracks the ranking of iPhone applications in China. Jan 01 Feb 01 Mar 01 176 While most VPN applications are blocked from the iPhone Ap- Date (2020) 177 ple Store, we identified one still available on it. Around the 178 time of the Hubei lockdown, its rank popularity increased sig- § Twitter, 179 nificantly and maintained that ranking (top panel of Figure 1). App Ranking 180 Concurrent with the increase in popularity of the VPN ap- 0 181 plication is a sudden increase in popularity of Facebook, Twit- Wuhan Lockdown ¶ ● 182 ter, and Wikipedia applications, as Figure 1 shows. These 200 ● ● ● ●●● ● ● ● ●●● 183 ● ●●● ● increases indicate that those jumping the Firewall as a result ● ● ● ●●●● ●● ●● ● ●● ● ● ●●●●●●●● ●● 400 ● ● ● ● 184 of the crisis were engaging in part with long blocked websites Rank ● ● ● ● ● ● ●●●● ●●●● ●●●● ● ● ● ● ● ●● 185 in China – Twitter and Facebook have been blocked since 2009 ●● ●● ● 600 ●●● ● 186 and Chinese language Wikipedia since 2015.

187 This finding is consistent with data we collected directly DRAFTJan 01 Feb 01 Mar 01 188 from Twitter and Wikipedia. The top panel of Figure 2 shows 189 the number of geolocating users in China posting in Chinese Date (2020) 190 in the time period of interest. Immediately following the lock- 191 down, Chinese language accounts geolocating to China in- Wikipedia, Reference App Ranking 192 creased 1.4 fold, and post-lockdown, 10% more accounts were 0 193 active from China than before. The bottom panel of Figure Wuhan 20 Lockdown 194 2 shows that the crisis also coincided with increases of new ● 195 40 users, indicating that increases are due to new users and not ● ￿ ● ●● ● ●● 196 60 ● ● ●● ●● dormant ones reactivating. We provide a rough, back-of-the- ●● ●●●●●● ● ●●●●● ●●●●●●●● Rank ● ● ● ● ● ● ● ●●●●●●●●●●●● ●●●● ● ● ● ●●●● 80 ● ● ●● ● ● ● 197 envelope calculation for the absolute size of these effects. If ●●●●●● ●● ●●●● 198 there were 3.2 million Twitter users in China (44) prior to the 100 199 COVID-19 pandemic and the 10% increase in usage applies 200 generally to Twitter users (i.e. not just those geotagging), then Jan 01 Feb 01 Mar 01 Date (2020) §To protect the application and its users, we are not disclosing its name or the exact ranking, though results are available for review upon request. ¶Note that increase in popularity is not comparable across applications because popularity is Note: The top panel of this figure intentionally omits the name of the VPN app and its measured in terms of ranks. More highly ranked applications (like Facebook and Twitter) precise ranking. may need many more downloads to achieve a more popular ranking. ￿SI Appendix S2 provides more detail, and Figure A1 shows trends per province.

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 3 Fig. 2. (Top) Number of Unique Geo-Locating Users in China Posting in Chinese. Fig. 3. Views of Wikipedia Pages in Chinese (Bottom) The Fraction of Active Users Who Joined Twitter in the Last 30 Days.

ZH: Change in Page Views per Day

Lockdown Lockdown 1.4x 1.4 1500 ● 1.3 1.2

1250 1.1x 1.1 Total Views Total 1 1000 0.9

750 Level) (Compared to Pre−Covid Jan Mar May Jul Jan Mar May Jul Date in China posting Chinese

Number of geo−locating users Date Note: This figure shows the ratio of total daily views of Wikipedia pages in Chinese compared to December 2019 views (12.7 million views per day in December 2019). Lockdown The beginning of the Hubei lockdown and the first relaxation of lockdown in Hubei are +30 days indicated in gray. 0.15

0.10 throughout China as a result of the Wuhan lockdown; Hubei, 227 1.5x the most impacted province, experienced the most sustained 228 0.05 increase in geolocated users. 229 Figure 4 measures the initial increase of Twitter volume on 230 0.00 January 24, 2020, the day after Wuhan’s lockdown and the 231 Jan Mar May Jul who joined in last 30 days start of lockdown in twelve other cities in Hubei, in compari- 232

Fraction of geo−locating users Fraction Date son to the average from December 1, 2020 to January 22, 2020 233 Note: These figures display unique users and unique users who signed up within the last 30 days. The decline in ‘new’ users after the end of lockdown in the right panel is in each province in China (the x-axis). The y-axis measures 234 driven by a decline in new sign-ups after lockdown easing, rather than lockdown users how sustained the increase was – the ratio of Twitter volume 235 leaving the site (they are no longer considered new after 30 days). 30 days after the quarantine to the baseline before the outbreak. 236 Hubei is in the top-right corner of the plot: Twitter volume 237 201 320,000 new users joined Twitter because of the crisis, includ- there doubled in comparison to the previous baseline, and the 238 †† 202 ing users who do not post or post publicly. We assess this esti- doubling persisted 30 days after the crisis. These estimates 239 203 mate in SI Appendix S4 using the estimated fraction of posts are drawn from polynomial models fit to the daily number of 240 204 in Chinese that are geotagged (1.95%) and the total number users per province – Figure A1 in the appendix displays the 241 205 of unique Twitter users in our sample (47,389 users posting in modeled lines over the raw data for each province. 242 206 Chinese and in China). To further validate that this increase in Twitter usage in 243 207 Data from Wikipedia on the number of views of Wikipedia China is related to the Wuhan lockdown, we collected real- 244 208 pages by language matches the App Annie and Twitter pat- time human mobility data from Baidu, one of the most pop- 245 209 terns.** We measure the total number of views for Chinese ular map service providers in China. The decrease in mobil- 246 210 language Wikipedia by day from before the coronavirus crisis ity in 2020 is correlated with the increase in Twitter users 247 211 to the time of writing. Figure 3 reveals large and sustained in- across provinces in China, net of a New Year’s effect (Fig- 248 212 creases in views of Chinese language Wikipedia,DRAFT beginning at ure A3). However, as the crisis spreads, the demobilization 249 213 the Wuhan lockdown and continuing above pre-COVID levels effect disappears, while Twitter usage remains elevated. The 250 214 through May 2020. Views of all Wikipedia pages in Chinese overall increase in Twitter users across China two weeks after 251 215 increased by around 10% during lockdown and by around 15% the lockdown and beyond cannot be explained by further de- 252 216 after the first month of lockdown. This increase persisted long creases in mobility or New Year seasonality (Figure A4). SI 253 217 after the crisis subsided. In absolute terms, the total number Appendix S3 presents more detail. 254 218 of page views increases from around 12.8 million views per

219 day in December 2019, to 13.9 million during the lockdown Crisis Provided a Gateway to Censored Political In- 255 220 period (January 24 through March 13), and up to 14.7 million formation. This subsection examines how the crisis impacted 256 221 views per day from mid-February through the end of April. what content Twitter users from mainland China and users of 257 Chinese language Wikipedia were consuming. Both Twitter 258

222 Increases in Circumvention Occurred Throughout and Wikipedia facilitate access to a wide range of content, not 259 223 China. Whereas the data from AppAnnie and Wikipedia can- just information sensitive to the Chinese government. New 260 224 not distinguish between circumvention patterns within China, users of Twitter from China might follow Twitter accounts pro- 261 225 the geo-location in the Twitter data enables the examination of ducing entertainment or even Twitter accounts of Chinese state 262

226 subnational variation. Circumvention occurred in provinces ††While almost all provinces experience a sustained increase in Twitter volume, Beijing and Shanghai have an overall decrease in Twitter volume after the outbreak. We suspect many **This page view data is publicly available: https://dumps.wikimedia.org/other/pagecounts-ez/ Twitter users in Beijing and Shanghai left those cities during the outbreak, which is corrobo- merged/ rated by the Baidu mobility data we detail in SI Appendix S3.

4 | Chang et al. Fig. 4. Increases in GeoLocated Twitter Activity by Province (modeled) Chinese embassies in Pakistan and Japan. Category 6 is also 300

301 2.1 not sensitive, as these accounts mostly do not tweet about pol- 2 1.9 302 Hubei itics, but instead are entertainment or commercial accounts or 1.8 Gansu 1.7 Anhui Shanxi accounts of non-political individuals. 303 1.6 Shandong Inner Mongolia Guangxi 1.5 Heilongjiang Hebei Guizhou Hunan We want to understand how the coronavirus crisis affected 304 1.4 Henan Average Number of Jiangxi 1.3 Qinghai Users Geo−Locating ShaanxiLiaoningHainanJilin trends in follower counts of each of the six categories, and in 305 1.2 Sichuan Each Day After Fujian Xinjiang Wuhan Lockdown 1.1 Yunnan particular, compare how the crisis affected the followings of 306 GuangdongJiangsuNingxia Zhejiang a 50 1 307 Tianjin a 100 categories 1-3 to those in categories 5 and 6. We therefore 0.9 ShanghaiBeijing Jump after 30 days Jump a 150 downloaded the profile information of all accounts that began 308 0.8 following these popular accounts after November 1, 2019. We 309 0.7 Tibet then use the location field to identify which of the 38,050,454 310 0.6 followers are from mainland China or Hong Kong (see SI Ap- 311 0.9 1.3 1.7 2.1 2.5 2.9 Jump − January 24, 2020 pendix S2 for more details). 312

Note: This figure shows the increase in geo-located Twitter users compared to the Because Twitter returns follower lists in reverse chronolog- 313 average number of geo-located Twitter users in a province before the Hubei lockdown. Estimates for 30 days after and day of lockdown are drawn from a five term polynomial ical order, we can infer when an account started following 314 regression on the number of unique geo-located Twitter users per day after the another account (46). For the accounts in the six categories, 315 lockdown. These province-by-province polynomials are displayed over the raw data in Figure S1 in SI Appendix. we compare the increase in followers from mainland China to 316 the increase in followers from Hong Kong accounts relative to 317 their December 2019 baselines; we chose Hong Kong because 318 263 media and officials, who have become increasingly vocal on it is part of the PRC but is not affected by the Firewall. The ul- 319 264 the banned platform (45). New users of Wikipedia might only timate quantity of interest is the ratio of these two increases. If 320 265 seek out information about the virus and not about politics. If the ratio is greater than one, then the increase in following re- 321 266 the crisis produced a gateway effect, we should see increases lationships is more pronounced among mainland Twitter users 322 267 in consumption of sensitive political information unrelated to as compared to those from Hong Kong. 323 268 the crisis. Figure 5 shows this ratio by category-day. Relative to Hong 324

325 269 Types of Twitter accounts mainland Twitter users started to follow Kong, the crisis in mainland China inspired disproportionate 326 270 as a result of the crisis. increases in the number of followers of international news 327 271 We use data from Twitter to examine what types of accounts agencies, Chinese citizen journalists, and activists (some of 328 272 received the largest increases in followers from China due to whom might otherwise, without exposure on Twitter, be ob- 329 273 the crisis. For this purpose, we identify 5,000 accounts that scure within China, especially ones who have been banned 330 274 are commonly followed by Chinese Twitter users. The Materi- from public discourse for a long time) – all information con- 331 275 als and Methods and SI Appendix S2 detail how we identified sidered sensitive that has long been censored. In comparison, 332 276 these accounts. there is only a small increase in mainland followers of Chi- 333 277 We assigned each of the 5,000 popular accounts into one nese state media and political figures during the lockdown pe- 334 278 of six categories: 1) international sources of political informa- riod and a slight decrease for non-political bloggers and enter- 335 279 tion, including international news agencies; 2) Chinese citizen tainers. Figure 6 reports the regression estimate for the rela- 336 280 journalists or political commentators, which include non-state tive ratio of number of new followers (akin to a difference-in- 337 281 media discussions of politics within China; 3) activists, or ac- differences design with Hong Kong as control group and De- DRAFTcember 2019 as pre-treatment period). The result is the same. 338 282 counts disseminating information about politics in the U.S., 283 Taiwan, or Hong Kong; 4) accounts disseminating pornogra- We then demonstrate that the result does not depend on 339 284 phy; 5) state media and political figures; and 6) entertainment the choice of comparison group and that the relative increase 340 285 or commercial influencers. Categories 1, 2, and 3 are accounts starts no earlier than Wuhan lockdown. Figure A6 in SI Ap- 341 286 that might distribute information sensitive to the Chinese gov- pendix S2 conducts a placebo test by running weekly regres- 342 287 ernment, such as international media blocked by the Great Fire- sions, showing that the relative increase in followers in China 343 288 wall (e.g. New York Times Chinese and Wall Street Journal starts precisely during the week of lockdown. Figures A7, A8, 344 289 Chinese); Chinese citizen journalists and political commenta- and A9 show that the same pattern holds with alternative com- 345 290 tors such as exiled political cartoonist Badiucao and currently parison groups such as overseas Chinese in Taiwan and the 346 291 detained blogger Yang Hengjun; and political activists such United States. 347 292 as free speech advocate Wen Yunchao and Wu’er Kaixi, for- SI Appendix S4 provides effect size estimates. There, we 348 293 mer student leader of the 1989 Tiananmen Square Protests. roughly estimate that around 320,000 new users came from 349 294 Accounts in Category 4 are pornography, which we consider China. Further, based on December 2019 follower growth 350 295 sensitive because it is generally censored by the Chinese gov- rates, 53,860 excess accounts follow citizen journalists and po- 351 296 ernment, but not politically sensitive like Categories 1-3. Ac- litical bloggers; 52,144 for international news agencies. By the 352 297 counts in Category 5 include accounts linked to the Chinese end of the lockdown, citizen journalists and political bloggers 353 298 government, including the government’s news mouthpieces benefit from 3.63 times the number of followers they other- 354 299 Xinhua and People’s Daily, as well as the Twitter accounts of wise would have had; activists, 2.97. Importantly, 88-90% of 355

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 5 the followers from China follow accounts in these categories 356 Fig. 5. Increases in Twitter Followers from China vs Hong Kong By Category one year later, and these rates are higher than for accounts 357 which start following in the weeks after the end of the Hubei 358 New Followers Compared to Baseline, China / Hong Kong lockdown. In addition, Figure A10 in SI Appendix S5 shows 359 International News Agencies Citizen Journalists / Political Bloggers 360 3x 1.31x 3x 1.42x that new users from China persist in tweeting at the same rates 2x 2x as those from Hong Kong and Taiwan. 361

1x 1x Types of Chinese language Wikipedia pages that received the most 362

0.7x 0.7x attention. 363

0.4x 0.4x To better understand patterns of political views in the 364 Wikipedia data, we leverage existing lists (see Materials and 365 Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 Removal Removal Methods for additional details) to categorize the Chinese lan- 366 Activists or US / Taiwan / Hong Kong Politics Pornography Accounts guage Wikipedia views into three different categories: 1) 367 3x 1.23x 3x 1.08x Wikipedia pages that were selectively blocked by the Great 368 2x 2x ‡‡ Firewall prior to Wikipedia’s move to https (after which all 369

1x 1x of Wikipedia was blocked), 2) pages that describe high level 370 §§ 0.7x 0.7x Chinese officials , and 3) historical leaders of China since 371 Mao Zedong. Whereas we would expect that a crisis in any 372 0.4x 0.4x country should inspire more information seeking about cur- 373 Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 374 Removal Removal rent leaders in Category 2, only if crisis created a gateway

State Media or Chinese Officials Non-Political Bloggers or Entertainment Accounts to historically sensitive information would we expect propor- 375 3x 3x 1.06x 0.85x tional increases in information seeking about historical leaders 376 2x 2x in Category 3 or information about sensitive events that were 377 selectively blocked by the Great Firewall on Wikipedia prior 378 1x 1x

0.7x 0.7x to 2015 in Category 1. 379 Figure 7 shows the increase in page views for each of these 380 0.4x 0.4x categories on Chinese Wikipedia relative to the rest of Chi- 381

Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 nese language Wikipedia. We find that the lockdown not only 382 Removal Removal increased views of current leaders (purple), but also views 383

Note: Gain in followers from mainland China compared to Hong Kong across six types of historical leaders (yellow) and views of pages selectively 384 of popular accounts, relative to December 2019 trends. Ratios here approximate the blocked by the Great Firewall (red). Tables A2 and A3 in SI 385 incidence rate ratios estimated in the models for Figure 6. A value greater than 1 means more followers than expected from mainland China than from Hong Kong. Accounts Appendix show specific pages disproportionately affected by 386 creating sensitive, censored information receive more followers than expected once the the increase in views of Wikipedia. While pages related to 387 Wuhan lockdown starts. Accounts that are not sensitive or censored, such as state media or entertainment, do not see greater than expected increases. coronavirus experienced a jump in popularity, other unrelated 388 sensitive pages including the “June 4 Incident,” “Ai Weiwei,” 389 and “New Tang Dynasty Television” (a television broadcaster 390 affiliated with Falun Gong) also experienced an increase in 391 Fig. 6. Increases in Twitter Followers China vs Hong Kong By Category (Re- ¶¶ gression Estimate) page views. 392 DRAFTFor more detail on this analysis as well as the Wikipedia 393 Relative Size of New Followers, China / Hong Kong pages that received the largest absolute and relative increases 394 in traffic, see SI Appendix S6. 395 International News Agencies Comparison with Other Countries Affected by the Cri- 396 Citizen Journalists / Political Bloggers sis. Since information seeking during crisis is common (1), we 397

Activists or US / Taiwan / investigate Wikipedia data from other countries affected by the 398 Hong Kong Politics crisis. We show that the gateway effect of crisis on historically 399 Pornography sensitive information is unique to the currently censored web- 400 Accounts pages in China. For comparison, we focus on Iran, another 401 State Media or Chinese Officials authoritarian country affected by COVID-19 that previously 402 censored Wikipedia (but does not any longer), and Russia, an 403 Non-Political Bloggers or Entertainment Accounts ‡‡Using data from https://www.greatfire.org/. 0.75x 1x 1.25x 1.5x 1.75x §§These lists are based on offices in the CIA World Facebook. We use this list for easeof Mean and 95% Confidence Interval comparisons with other countries and remove the Ambassador to the United States from each Note: Incidence rate ratios shown above are from negative binomial regressions of list. China’s list is available here (and there are links to leaders of other countries on the same number of new followers on the interaction between indicator variables for ‘in page): https://www.cia.gov/library/publications/resources/world-leaders-1/CH.html, exclud- ing Hong Kong and Macau. lockdown period’ and ‘in mainland China’, with December 2019 as control period and ¶¶ Hong Kong as control group. The June 2020 increase in China is due to the anniversary Tiananmen Square protests. Our claim is not that only the COVID-19 crisis causes increases in views of sensitive content. That the same behavior is observed around another crisis event supports this paper’s argument.

6 | Chang et al. are always available to the public. 422 ZH: Page Views by Category Table 2 shows these results. While overall Wikipedia views 423 Lockdown and page views of current leaders increase in three out of 424 Current Leader Pages Historical Leader Pages 425

4 four comparison countries, only in China do historical lead- Blocked Pages (pre−https) Rest of Wikipedia (=1) Ratio of Mobile to Desktop Views (7−day MA) ers increase disproportionately and consistently throughout 426 2 the whole time period. That is, we see an overall effect on 427

428

1 information-seeking throughout the world, including for his-

Wikipedia Page Views Wikipedia Page torical leaders; in China, we see larger increases for histor- 429 Ratio to Rest of Wikipedia

0.5 ical leaders compared to Wikipedia page views in general. 430 Jan Mar May Jul Date The small increases in historical political leader page views 431 in German and Italian did not correspond with the start of the 432

DE: Page Views by Category COVID-19 crisis or their respective lockdowns (Figure 7). 433

434 Lockdown in Italy Lockdown Further, we do not see increased attention to pages previ- Current Leader Pages 435 Historical Leader Pages ously blocked in Iran (47) during the crisis – Wikipedia pages 4 Rest of Wikipedia (=1) Ratio of Mobile to Desktop Views (7−day MA) that can now be accessed without restriction in Iran. 436

2 In SI Appendix S6.1, we replicate these results for much 437 larger sets of 1) historical leaders and 2) ‘politically sensitive’ 438 1 pages (pages related to the pre-https blocked pages in Iran and 439 Wikipedia Page Views Wikipedia Page Ratio to Rest of Wikipedia

0.5 China, and political opposition pages in Russia). We expand 440 Jan Mar May Jul these sets of pages using Wikipedia2vec (48), and find that 441 Date very broad information-seeking about historical leaders and 442 politically sensitive topics occurred only in China. 443 IT: Page Views by Category

Lockdown Table 2. During the lockdown period, Wikipedia views in Chinese Current Leader Pages Historical Leader Pages increased relative to overall views for politically sensitive Wikipedia Rest of Wikipedia (=1)

4 Ratio of Mobile to Desktop Views (7−day MA) pages and political leader pages, as well as for historical political leaders. 2

Change: Overall Blocked, pre-https Leaders Historical Leaders

1 Language relative to overall:

Wikipedia Page Views Wikipedia Page Chinese 1.09 1.15 1.86 1.42 Ratio to Rest of Wikipedia

0.5 (1.05 - 1.12) (1.09 - 1.22) (1.67 - 2.07) (1.32 - 1.52) Jan Mar May Jul <0.001 <0.001 <0.001 <0.001 Date Persian 1.42 0.84 0.91 0.82 (1.37 - 1.46) (0.79 - 0.89) (0.80 - 1.05) (0.75 - 0.90) Fig. 7. Views of Blocked, Current Leader, and Historical Leader Wikipedia Pages <0.001 <0.001 0.20 <0.001 in Chinese, German, and Italian. Russian 1.23 1.73 0.90 Note: Vertical lines indicate the starts and ends of lockdown periods – see Table A4 for (1.18 - 1.28) (1.48 - 2.02) (0.82 - 0.99) specific dates. <0.001 <0.001 0.03

German 1.16 2.36 1.21 (1.12 - 1.20) (2.02 - 2.76) (1.05 - 1.40) 404 authoritarian country that does not censor Wikipedia – for Iran, <0.001 <0.001 0.01 405 like China, we know which Wikipedia pages were previously Italian 1.47 3.29 1.17 406 censored (47). We also show data from democraciesDRAFT without (1.40 - 1.53) (2.72 - 4.00) (1.02 - 1.34) <0.001 <0.001 0.03 407 censorship affected early on by the COVID-19 crisis, Italy and 408 Germany.*** Note: Incidence rate ratios shown above are from a negative binomial regression estimating the daily number of views within a category in the lockdown period 409 To make the comparison, we use lists of current leaders compared to December 2019 relative to the number of views across the rest of 410 from these countries (based on office lists in the CIA World Wikipedia compared to December 2019 (using the same difference-in-difference specification as the Twitter follower analysis). Observations are the total views per 411 Factbook, see Materials and Methods), and create lists of his- category by day. 95% confidence intervals are shown in parentheses, and p-values are 412 torical leaders using de facto country leaders since World War shown in the third row for each language. See the SI for over-time ratios by day for all comparison languages (Figure A11), and for the dates of the lockdowns used (Table 413 II (see Table A4 in the appendix for a list of these titles and A4). German and Italian pages of historical leaders (shown in orange in the figures 414 offices). All of these countries were affected by the crisis in above) saw several large and short-lived spikes in views not clearly related to those countries’ lockdowns. Figures A12, A13, A14, and A15 in the SI replicate these results 415 late February or early March and Italy imposed relatively strin- for much larger sets of Wikipedia pages, including Russian language pages related to 416 gent lockdowns. Therefore, we expect increases in informa- opposition leaders and movements (which did not see broad increases in views). 417 tion seeking for current leaders, as citizens begin to pay more 418 attention to current politics as the crisis hits. However, none

419 of these countries block Wikipedia. Information seeking about Discussion 444 420 the current crisis therefore should not act as a gateway to infor- Crisis in highly censored environments creates a gateway to 445 421 mation about historical events or controversies, as these pages sensitive, censored information unrelated to the crisis. Like 446 ***Citizens in each of these countries speak languages relatively specific to their country, and in democracies, consumers of information in autocracies seek 447 therefore we expect most of the page views of Italian, German, Persian, and Russia Wikipedia to originate in Italy, Germany, Iran, and Russia respectively. out information and depend on the media during crisis. How- 448

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 7 449 ever, in highly censored environments, increased information phy). This sampling allows us to estimate the impact of the coronavirus on 512 pornography while decreasing our requests to the Twitter API. 513 450 seeking also incentivizes censorship circumvention. This new Mobility data. Human mobility data is publicly available from Baidu 514 451 ability to evade censorship allows users to discover a wider Qianxi (https://qianxi.baidu.com/2020/), which tracks real-time move- 515 452 variety of information than they may have initially sought. ment of mobile devices and is used in studies of human mobility and COVID- 516 453 While evaluations of responses to an ongoing crisis and 19 containment measures (50). Our robustness checks use data across China 517 during the Lunar New Year period in both 2020 and 2019. We scraped the 518 454 comparisons to other governments’ responses to the same cri- daily within city movement index (an indexed measure of commute popula- 519 455 sis may have benefited government officials in China in this tion relative to the population of the city), as well as daily moving out in- 520 456 particular circumstance (12), beyond these evaluations, in- dex (an indexed measure based on the volume of population moving out of 521 the province relative to the total volume of migrating population on that day 522 457 creased access to historical and long-censored information, as across all provinces in China). See Section S3 for more details. 523 458 documented here, has the potential to dampen positive changes Wikipedia data. Data on the number of Wikipedia page views is publicly 524 459 in trust or compound negative changes in trust, and may also available here: https://dumps.wikimedia.org/other/pagecounts-ez/ 525 merged/. To better understand patterns of political views in the Wikipedia 526 460 contribute to easier access to uncensored information about data, we use existing lists to categorize the Chinese language Wikipedia 527 461 a government in the future. Natural disasters, including epi- views into three different categories: 1) Wikipedia pages that were selectively 528 462 demics, tend to alter trust in government officials. When a pol- blocked by the Great Firewall (https://www.greatfire.org/ maintains a list 529 of websites censored by the Great Firewall) prior to Wikipedia’s move to https, 530 463 icy response is perceived as efficacious, support for the level of after which all of Wikipedia was blocked, 2) pages about high level Chinese 531 464 government perceived to have directed the response increases officials (using offices listed in the CIA World Factbook https://www.cia. 532 465 (13, 49). On the other hand, neglectful responses can induce gov/library/publications/resources/world-leaders-1/CH.html, exclud- 533 ing Hong Kong and Macau as well as the Ambassador to the United States), 534 466 subsequent protest participation (10). In China, the average and 3) historical ‘paramount’ leaders of China since Mao Zedong. 535 467 effect of natural disasters from 2007-2011 was to decrease po- In comparing multiple languages and countries, we use the same offices 536 468 litical trust, and internet users have decreased baseline levels listed in the CIA World Factbook to create lists of current leaders from Iran, 537 Russia, Italy, and Germany (for office holders as of February 2020), and create 538 469 of political trust (13, 14). lists of historical leaders using de facto country leaders since World War II. 539 470 While the results do not link the COVID-19 crisis gateway See Table A4 in the appendix for a list of these titles and offices, as well as 540 471 effect to the political fortunes of the Chinese government, they the lockdown start and end dates used for the language by language Wikipedia 541 page view models displayed in Table 2. The list of pages of Wikipedia pages 542 472 do suggest that crisis has the ability to undermine censorship blocked in Iran was published by (47). 543 473 in highly censored environments. While in normal times cen- In SI Appendix S6.1, we replicate the Wikipedia page view results for 544 474 sorship can be highly effective, it may create unintended side much larger sets of 1) historical leaders and 2) ‘politically sensitive’ pages 545 475 effects of facilitating access to sensitive information during (pages related to the pre-https blocked pages in Iran and China, and po- 546 litical opposition pages in Russia). We expand these sets of pages using 547 476 crisis. Wikipedia2vec (48). 548 Models. Incidence rate ratios for the follower analyses and the Wikipedia 549 page view analyses are from negative binomial regressions. In the follower 550 477 Materials and Methods analysis, this models the number of new followers per day, with a separate 551 model for each account category. Independent variables are ‘in lockdown pe- 552 478 Application download rank data. Download rank data for Facebook, Twit- riod’ and ‘in mainland China’, and the effect of interest is the interaction be- 553 479 ter, Wikipedia, and the VPN app come from application analytics firm AppAn- tween these indicator variables (i.e. a difference-in-difference), with Decem- 554 480 nie (https://www.appannie.com), which tracks the popularity of iPhone ap- ber 2019 as control period and Hong Kong as control group. The Wikipedia 555 481 plication downloads in China. While most VPN applications are blocked from page view analyses use the same specification, reporting the coefficient for 556 482 the iPhone Apple Store (and there are other means of obtaining VPNs), we ‘in lockdown period’ and ‘in page set’ (current leader, historical leader, pre- 557 483 identified one still available on it. VPN download rank shown in the text is viously blocked) relative to December 2019 and relative to page views for 558 484 for that VPN application. This data contains the ranking of an application – the rest of Wikipedia. Observations are the total views per category by day. 559 485 for Wikipedia, its rank within the Reference App category – rather than the Figures displaying (log scale) ratios of followers/Wikipedia page views ap- 560 486 number of downloads. To protect the VPN application and its users, we do not proximate coefficients from these negative binomial regressions. Negative 561 487 disclose its name or the exact ranking, though results are available for review binomial regressions were estimated using the MASS library in R. 562 488 upon request. DRAFT 563 489 Twitter data. For the Twitter analyses, we collected 1,448,850 tweets Increases in geolocated Twitter activity (unique users) by day and by 564 490 (101,553 accounts) from mainland China from December 1, 2019 until June province were modeled using a five-term polynomial regression (by day) for 565 491 30, 2020. These tweets were identified using Twitter’s POST statuses/filter time trends after the Hubei lockdown and a mean without any time trend prior 566 492 endpoint. Our analyses are limited to the 367,875 that were posted in Chinese to lockdown (see Figure A1 for a province by province visualization of this 567 493 (47,389 accounts that posted in Chinese, 43,114 that had names or descriptions model). The points in Figure 2 are predicted values by province for the first 568 494 in Chinese). day of lockdown and day 30 of lockdown. 495 The Twitter follower analysis examines accounts that Twitter users from 496 China commonly follow. To find those accounts, we randomly sampled 5,000 1. SJ Ball-Rokeach, ML DeFleur, A dependency model of mass-media effects. Commun. 569 497 users geolocated to China. For each of these users, we gathered the entire list Res. 3, 3–21 (1976). 570 2. M Loveless, Media dependency: Mass media as sources of information in the democra- 571 498 of whom they follow, their Twitter “friends.” From these 1,818,159 friends, tizing countries of central and eastern europe. Democratization 15, 162–183 (2008). 572 499 we extracted the 5,000 most common accounts. We also selected only ac- 3. PL Hirschburg, DA Dillman, SJ Ball-Rokeach, Media system dependency theory: Re- 573 500 counts that were Chinese language accounts or had Chinese characters in their sponses to the eruption of mount st. helens. Media, audience, social structure, 117–126 574 501 name or description field to ensure that we were studying relevant accounts: (1986). 575 502 those disseminating information easily accessible to most Chinese users. SI 4. SL Althaus, American news consumption during times of national crisis. PS: Polit. Sci. 576 503 Appendix S2 provides more detail. Polit. 35, 517–521 (2002). 577 504 We downloaded the profile information of all accounts that began follow- 5. YC Kim, JY Jung, EL Cohen, SJ Ball-Rokeach, Internet connectedness before and after 578 505 ing these popular accounts after November 1, 2019. Because Twitter returns september 11 2001. New media & society 6, 611–631 (2004). 579 6. J Bar-Ilan, A Echermane, The anthrax scare and the web: A content analysis of web 580 506 follower lists in reverse chronological order, we can infer when an account pages linking to resources on anthrax. Scientometrics 63, 443–462 (2005). 581 507 started following another account (46). We then use the location field to iden- 7. GE Marcus, MB MacKuen, Anxiety, enthusiasm, and the vote: The emotional underpin- 582 508 tify which of these 38,050,454 followers are from mainland China or Hong nings of learning and involvement during presidential campaigns. Am. Polit. Sci. Rev., 583 509 Kong (see SI Appendix S2 for more details). We downloaded all new follow- 672–685 (1993). 584 510 ers of non-pornography accounts and all new followers of a random selection 8. MH Ribeiro, et al., Sudden attention shifts on wikipedia following covid-19 mobility 585 511 of 200 pornography accounts (the majority of the accounts were pornogra- restrictions. arXiv preprint arXiv:2005.08505 (2020). 586

8 | Chang et al. 587 9. ME Kahn, The death toll from natural disasters: the role of income, geography, and 53. T Mikolov, I Sutskever, K Chen, GS Corrado, J Dean, Distributed representations of 671 588 institutions. Rev. economics statistics 87, 271–284 (2005). words and phrases and their compositionality. NIPS, 3111–3119 (2013). 672 589 10. AQ Flores, A Smith, Leader survival and natural disasters. Br. J. Polit. Sci., 821–843 590 (2013). 591 11. E Lazarev, A Sobolev, IV Soboleva, B Sokolov, Trial by fire: a natural disaster’s impact 592 on support for the authorities in rural russia. World Polit. 66, 641 (2014). 593 12. D Stasavage, Democracy, autocracy, and emergency threats: Lessons for covid-19 from 594 the last thousand years. Int. Organ., 1–17 (2020). 595 13. Y You, Y Huang, Y Zhuang, Natural disaster and political trust: A natural experiment 596 study of the impact of the Wenchuan earthquake. Chin. J. Sociol. 6, 140–165 (2020). 597 14. J Lee, The social impact of natural hazards: a multi-level analysis of disasters and forms 598 of trust in mainland China. Disasters 45, 158–179 (2021). 599 15. E Morozov, The Net Delusion: The Dark Side of Internet Freedom. (PublicAffairs, New 600 York), (2011). 601 16. R MacKinnon, Consent of the Networked: The Worldwide Struggle For Internet Free- 602 dom. (Basic Books, New York), (2012). 603 17. R Deibert, J Palfrey, R Rohozinski, J Zittrain, Access Contested: Security, Identity, and 604 Resistance in Asian Cyberspace. (MIT Press, Cambridge), (2011). 605 18. S Sanovich, D Stukal, JA Tucker, Turning the virtual tables: Government strategies for 606 addressing online opposition with an application to Russia. Comp. Polit. 50, 435–482 607 (2018). 608 19. D Stockmann, ME Gallagher, Remote control: How the media sustain authoritarian rule 609 in china. Comp. Polit. Stud. 44, 436–467 (2011). 610 20. R Enikolopov, M Petrova, E Zhuravskaya, Media and political persuasion: Evidence 611 from russia. Am. Econ. Rev. 101, 3253–3285 (2011). 612 21. M Adena, R Enikolopov, M Petrova, V Santarosa, E Zhuravskaya, Radio and the rise of 613 the nazis in prewar germany. The Q. J. Econ. 130, 1885–1939 (2015). 614 22. D Yanagizawa-Drott, Propaganda and conflict: Evidence from the Rwandan genocide. 615 The Q. J. Econ. 129, 1947–1994 (2014). 616 23. D Stockmann, Media Commercialization and Authoritarian Rule in China. (Cambridge 617 University Press, Cambridge), (2012). 618 24. H Huang, Propaganda as signaling. Comp. Polit. 47, 419–444 (2015). 619 25. ME Roberts, Censored: Distraction and Diversion inside China’s Great Firewall. (Prince- 620 ton University Press), (2018). 621 26. J Pan, AA Siegel, How saudi crackdowns fail to silence online dissent. Am. Polit. Sci. 622 Rev. 114, 109–125 (2020). 623 27. SC Jansen, B Martin, Making censorship backfire. Counterpoise 7, 5–15 (2003). 624 28. Z Nabi, Censorship is futile. First Monday 19 (2014). 625 29. N Hassanpour, Media disruption and revolutionary unrest: Evidence from Mubarak’s 626 quasi-experiment. Polit. Commun. 31, 1–24 (2014). 627 30. WR Hobbs, ME Roberts, How sudden censorship can increase access to information. 628 Am. Polit. Sci. Rev. 112, 621–636 (2018). 629 31. C Gläßel, K Paula, Sometimes less is more: Censorship, news falsification, and disap- 630 proval in 1989 east germany. Am. J. Polit. Sci. (2019). 631 32. L Boxell, Z Steinert-Threlkeld, Taxing dissent: The impact of a social media tax in 632 Uganda. Working Paper (2019). 633 33. Y Chen, DY Yang, The impact of media censorship: 1984 or brave new world? Am. 634 Econ. Rev. 109, 2294–2332 (2019). 635 34. GE Marcus, WR Neuman, M MacKuen, Affective Intelligence and Political Judgment. 636 (University of Chicago Press, Chicago), (2000). 637 35. M MacKuen, J Wolak, L Keele, GE Marcus, Civic engagements: Resolute partisanship 638 or reflective deliberation. Am. J. Polit. Sci. 54, 440–458 (2010). 639 36. B Albertson, SK Gadarian, Anxious politics: Democratic citizenship in a threatening 640 world. (Cambridge University Press), (2015). 641 37. Z Tai, T Sun, Media dependencies in a changing media environment: The case of the 642 2003 sars epidemic in china. New Media & Soc. 9, 987–1009 (2007). 643 38. Q Cao, The limitations of internet censorship for authoritarian information con- 644 trol:evidence from china. Working Paper (2020). DRAFT 645 39. C Buckley, SL Myers, As new coronavirus spread, china’s old habits delayed fight. The 646 New York Times (2020). 647 40. A Qin, V Wang, Wuhan, center of coronavirus outbreak, is being cut off by chinese 648 authorities. The New York Times (2020). 649 41. J Griffiths, T John, S George, Unprecedented lockdown on 10 cities and 30million 650 people. CNN (2020). 651 42. J Griffiths, A Woodyatt, 80 million people in china are living under travel restrictions 652 due to the coronavirus outbreak. CNN (2020). 653 43. P Belluck, Coronavirus death rate in wuhan is lower than previously thought, study 654 finds. New York Times (2020). 655 44. P Mozur, Twitter Users in China Face Detention and Threats in New Beijing Crackdown 656 (2019). 657 45. L Zhou, Chinese officials have finally discovered twitter. what could possibly gowrong? 658 South China Morning Post (2020). 659 46. ZC Steinert-Threlkeld, Longitudinal Network Centrality Using Incomplete Data. Polit. 660 Analysis 25, 308–328 (2017). 661 47. N Nazeri, C Anderson, Citation Filtered: Iran’s Censorship of Wikipedia. Iran Media 662 Program (2013). 663 48. I Yamada, et al., Wikipedia2Vec: An Efficient Toolkit for Learning and Visualizing the 664 Embeddings of Words and Entities from Wikipedia. EMNLP, 23–30 (2020). 665 49. E Lazarev, A Sobolev, IV Soboleva, B Sokolov, Trial by fire: A natural disaster’s impact 666 on support for the authorities in rural Russia. World Polit. 66, 641–668 (2014). 667 50. MUG Kraemer, et al., The effect of human mobility and control measures on the covid- 668 19 epidemic in china. Science 368, 493–497 (2020). 669 51. P Mozur, China Presses Its Internet Censorship Efforts Across the Globe (2018). 670 52. E Kinetz, Army of fake fans boosts China’s messaging on Twitter (2021).

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 9 673 Supplementary Materials for “Crisis is a Gateway to Censored Information”

674 S1. Twitter Activity by Province

675 Figure A1 shows the number of unique, geolocating users who are tweeting in Chinese by province. The x-axis is the number of 676 months before (negative) or after (positive) the initial coronavirus lockdown in Hubei province. The blue line is a pre-lockdown 677 average for x less than 0 and a five term polynomial regression for x greater than or equal to 0 (where 0 is the first day of Hubei’s 678 lockdown). The points in Figure 2 are the values of the blue line by province for x equals 1/30 (first day of lockdown) and x 679 equals 1 (day 30 of lockdown).

Fig. A1. Increases in Geolocated Twitter Activity by Province (modeled)

Guangdong Anhui Beijing Fujian Gansu

● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● 50 ● ● ● ● ●●● ● ● ●● ● ● 20 ● ● ● ● ● ● ● ● ● ● ●●●● ●●●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ●●●●● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ●●● ● ● ●●● ● 175 ● 200 ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●●●● ●● ● ● ●● ●● ● ● ●● ● ● ●● ● ● ● 50 ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● 40 ● ● ● ● ● ● ● ● ● ● ● ● ●● 15 ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ●● ● ● ●● ● ● ● ● ●●●●●● ●● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● 150 ● ● ● ● ● ● ● ● 150 ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ●● ●● ●●● ●●● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ●● 40 ● ● ● ●● ● ● ● 30 ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ●● ●● ●●● ● ● ● ● ●● ● ● ● 10 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●● ●●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ● ●●●●● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ●● ● ● ● ● ● ●● ● ● ●● ● ●● ● ●● ●●● ● ● ●●●●●● ● ● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ●●●● ● ● ●●● ●● ● ●● ● ● ● ● ● ● ● 100 ● ● ● ●● 125 ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ●●● ● ● ●●●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ● 20 ● ● ● ●● ● ● ● ● 30 ● ● ● ● ●● 5 ● ● ● ● ●●●● ● ●●● ●● ●● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● 50 ● ● 100 ● ● ● 10 ● ● ● 20 ● ● 0 Guangxi Guizhou Hainan Hebei Heilongjiang

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 25 ● 20 ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● 50 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 40 ●● ●●●● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ● ●● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● 30 ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● 20 ● ● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● 15 ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 40 ●●● ● ●● ● ● ● ● ● ● ●●● ● ● ●●● ● ● ● ● ● ●● ● 30 ●● ● ● ● ● ●● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ● ●●● ●● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ●●● ●● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●●● ● ● ●●●● ● ●●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ●● ●● ● ● ● ● ●● ●● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● 15 ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ●● ●● ●●● ●●●● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ●●●●●● ● ● ● 20 ● ● ● ● ●● ● ●● ● ● ● ● ●● ●● ● ●●● ● ● ● ● ● ● ● ● ●● ● ● 30 ● ● ● ●● ● ● ● ● ● ● ● 20 ● ●● ● 10 ●●● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ●●●●● ●● ●●● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●●● ●●● ● ● ● ● ● ●● ●● ●● ●●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ●● ●● ● ●● ● ● ●●● ● ● ● ● ●● ●●●● ●●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ●● ● ● ● ●●●● ● ● ●● ●● ●●● ● ● ● 10 ● ● ● ●●● ● ●●● ● ●● ●● ●●●● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ●● ● ● ● ●●● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ● ● ●● ● ●● ● ● ●● ●●● ● ● ●● ● 20 ● ● ● ●● ● ● ● ● ●●●●● ● ●● ● 10 ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● 5 ● ● ● ●● ● ● ● ● ● ●● 10 ●●● ● 0 ● ● 5 ● 10 ● ● ● ● Henan Hubei Hunan Inner Mongolia Jiangsu

● ● ● ●● 70 ● 120 ● ● ● ● 70 ● ● ● ● ● ●● ● ● ● ● ● ● 80 ● ● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ●●● ● ● ● 60 ● ● ● ● ● ● ● ●● ● ● ● ●●● ● ● 60 ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 20 ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●●●● ● ● 100 ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 50 ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● 50 ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●●● ● 60 ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ●● ● ● ●● ●●●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ●● ● 40 ● ● ●● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ●●●●●● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● 40 ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●●● ●● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●●●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ●●● ●● ● ● ● ● ● ●●● ● ● ● ● ● ● ●● 80 ● ● ● ●● ● ● ● ● ●● ● ● ● 10 ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ●●● ●● ● ● ● ●●● ●● ●● ●● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ● ●●●●●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 40 ● ● ●●● ● ● ●● ● ● ● ●●● ● ● ●● ● ● ● ●● ●● ● ● ● ● 30 ● ●●● ● ● ●● ● ● ● 30 ● ● ● ●●● ●● ● ●● ● ●● ●● ● ● ● ● ●● ●● ● ●●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ●●●● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● 20 ● ●●● ● ● 20 ● 60 ● 20 0 Jiangxi Jilin Liaoning Ningxia Qinghai

● ● ● ●● ● 10 ● ● ● ● 30 ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● 50 ● 10 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ●● ●● ● ● 20 ●●● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● 5 ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ●● ● ● ●● ●● ● ●●● ●● 20 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ●●●● ●● ●●●●● ●●● ● ●● ●●●●●●●● ● ●●●● ● ● ●●●●●●● ●●●●● ● ●● ●●● ●● ● ●●● ● ● ●●● ● ● ● ● ● ● ● ● ● 40 ● ●● ● ●● ● ● ● ● ●● ● ● ●● ●●●● ●● ●● ●●● ●●● ●● ● ●● ● ●● ● ● ● ●●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●●●●● ● ●● ● ● ● ● ● ●● ● ● ● ● ●●●●●●●●● ●●●● ●● ●●●●● ●●●●●● ● ●●●●● ●● ●●●● ●●● ●●●● ● ●●●● ●● ●●●●●●●●●●● ●●●●● ●●●● ●●●●●●●●● ●● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ● ● ● ● ●● ● ● 5 ● ● ● ●● ● ● ●● ●● ● ● ●●● ● ●●● ●● ●●● ● ● ●●●●●● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●●●● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ●●● ● ● ●● ● ● ●●●● ●●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●●●● ●●●●●●● ●●●●● ●●● ●●●● ●●●●●● ●● ● ● ●● ●●● ● ●●●● ●● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ●●● ● ●● ● ● ● ● ● ●●● ● ●●●● ● ●● ●●● ●● ● ● ● ● ● ● ● ●●● ●●●●●●● ● ● ● ● ● ●● ●● ●● ● ●●● ●● ●● ● ● ● ●●● ●●● ●● ●●● ●● ● ● ● ● ● ●● ● ●● ●●● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● 0 ● ● ●● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●● ● 30 ● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ●●●● ●●● ● ●●● ● ● ● ●● ●● ● 10 ● ● ●● ●● ● ● ● ● ● ● ●● ● ● ● 10 ● ●●● ● ● ● ●● ● ●● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0 20 ● −5 Shaanxi Shandong Shanghai Shanxi Sichuan

● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● 125 ● ● ●● ● ●● ● ● ●● ● ● ● 180 ● ● ● ●● ●● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● 30 ● ● ● ● ● ● ● ●● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●●● ● ●● ●●● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ●● ● ● ● ●● 40 ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● 100 ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● 80 ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● 160 ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 75 ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● 20 ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ● ●● ● ● ● ● ● ●●● ● ● 140 ● ● ● ● ● ● ● ● ● ● ● ● 30 ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● 60 ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ●● ●● ● ●●● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ●●● ●●● ● ●● ● ● ● ● ● ●● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● 50 ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ●●● 120 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● 10 ●● ● ● ● ● ●● ● 20 ● ● ● ● ● ● ● ● ● ● ●● ● ● ● 40 ● ● ● ● ● ● ● ● ● 25 ● ● ● ● ● ● ● ● 100 ● ● ● ● 0 Tianjin Tibet Xinjiang Yunnan Zhejiang

● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● 35 ● ● ● ● ● ● ● ● ● ● ● ● 10 30 ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ●● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● 100 ●● ● ● ● ● 30 ● ● ● ●● ● ● ● ● 10 ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ●●● ● ●● ●● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● 5 ●● ● ● ● ●● ● ● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ●● ●● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ●●● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ●● ● ● ● ● ●●●●● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ●● 20 ● ● ● ● ● ● ● ● ● ● ● ●● ● 25 ● ● ● ● ● ● ●●● ● ●● ● ● ● ● ●● ●●●●●●●●●●● ●● ●●● ●●●●● ●● ●●●●●● ● ●●● ●●●●● ● ●● ●●●● ● ●●●●●● ●● ●●● ●●●●●●●● ●●● ● ● ● ●● ●● ●● ● ●● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ●● ● ●● ●● ● ●●● ●●● ● ●●● ● ●● ● ● ● ● ● ●●● ● ●●●●●●● ●●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● 80 ● ● ● ● ●●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ●●●● ●● ● 5 ● ●●● ●● ●●● ● ●●● ● ● ● ●●●●● ● ●● ● ●●●● ●●● ●● ●● ●●● ● ●●● ●●● ●● ●●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●●●●● ●●● ● ●●● ● ● ●●●● ●●●● ●●● ●● ● ● ● ● ●● ●●●● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● 0 ● ● ● ● ● ● ● ●●●●● ● ●● ● ● ● ● ● ● ● ● ● 20 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●●●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● 10 ● ● ● ● ● ● ● ● ● ● ● Number of Unique Users Tweeting in Chinese Tweeting Number of Unique Users 15 ● ● −5 0 60 ● ● ● 10 ● −5 0 ● ● −2 0 2 4 −2 0 2 4 −2 0 2 4 −2 0 2 4 −2 0 2 4 MonthsDRAFT Since Wuhan Quarantine

10 | Chang et al. S2. Twitter Data 680

From a global sample of tweets with GPS coordinates, we found the 1,448,850 tweets from China from December 1, 2019 681 through June 30, 2020, 367,875 of which are in Chinese. This corpus contains 101,553 unique users, 43,114 of whom had 682 names or descriptions in Chinese. These dates were chosen to encompass a baseline period and the height of COVID-19 in 683 China. This corpus is used for Figures 2 and 4, evaluating the impact of the lockdown on tweeting behavior. 684 For the follower analysis (Figures 5 and 6), we sample 5,000 of these 43,114 accounts. For these 5,000 random users in China, 685 we download who they follow, their “friends” in Twitter parlance. From these friends, we identify the 5,000 most commonly 686 followed accounts that are either a Chinese language account or have Chinese characters in their name or description field. Of 687 these 5,000 most common friends, the vast majority were pornography accounts. We therefore hand-categorized the accounts 688 into pornography or not pornography. We keep the 354 non-porn accounts and sample 200 from the remaining 4,646 porn 689 accounts. 690 We then download the followers of these 554 accounts. We identify 38,050,454 total followers. For each, we identify the 691 location of the users. Because very few of these followers have geolocated information, we rely on the language of their Twitter 692 status and their self-reported location to distinguish between mainland and overseas followers. We only include users whose 693 status language is Chinese in order to study only Chinese language followers of these accounts. Followers are classified as 694 Mainland Chinese if the location field contains the name of a Chinese city, town, or province. Followers are classified as from 695 Hong Kong if the location field contains the name of a district in Hong Kong. Followers are classified as Taiwanese if the 696 location field contains the name of a Taiwanese city, county, or district. Followers are classified as US if if the location field 697 contains the name of states or state abbreviations (in capital letters). 698

DRAFT

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 11 699 S3. Mobility and Twitter Usage

700 To better understand the relationship between lockdown and Twitter usage, we scrape the publicly-available human mobility 701 data from Baidu Qianxi (https://qianxi.baidu.com/2020/), which tracks real-time migration (including moves in & out of 702 provinces and within city movements) across China during the Lunar New Year period in both 2020 and 2019.

703 Figure A2 plots the average within city movement index in both 2020 (real black line) and 2019 during the same period in 704 the Chinese Lunar New Year (dotted line). Specifically, since the New Years day is on February 5 in 2019 and January 25 in 705 2020, we shifted the dates in 2019 backwards for 12 days to match the dates in 2020. Red vertical line indicates the day of 706 Wuhan lockdown. One can see that almost all provinces experienced a huge decrease in human mobility after January 23 in 707 2020, compared to the same period in 2019. In 2019, we only see significant decreases in mobility in Beijing, Shanghai, and 708 Tianjin.

Within City Movement Index (Black: 2020, Dotted: 2019, Dates Adjust for Lunar New Year) Guangdong Anhui Beijing Fujian Gansu 5 5 6 6 4 5 5 4 4 4 3 4 3 3 3 3 2 2 2 2 2 Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Guangxi Guizhou Hainan Hebei Heilongjiang 6 6 5 5 5 5 5 4 4 4 4 4 3 3 3 3 3 2 2 Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Henan Hubei Hunan Inner Mongolia Jiangsu 6 6 6 6 6 5 5 5 5 4 5 4 4 4 3 4 3 3 3 2 2 3 2 Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Jiangxi Jilin Liaoning Ningxia Qinghai 6 6 5 4.0 5 5 5 3.5 4 4 4 4 3.0 3 3 3 3 2.5 2 2 2 2.0 Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Shaanxi Shandong Shanghai Shanxi Sichuan 6 7 6 5 6 5 5 5 4 5 4 4 4 4 3 3 3 3 2 3 2 2 2 Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Jan Feb Mar Tianjin Tibet Xinjiang Yunnan Zhejiang 6 3.5 4 5.0 6 5 3.0 4.5 5 3 4.0 4 4 2.5 3 2 3.5 3 2 2.0 1 3.0 2 2.5 Jan Feb Mar Jan Feb DRAFTMar Jan Feb Mar Jan Feb Mar Jan Feb Mar

Fig. A2. Within city movement index by Province (black: 2020, dotted: same period in 2019). Note: Real black line indicates the time series for the average within city movement index by province in 2020. Dotted line indicates the average within city movement index by province during the same Chinese Lunar New Year period in 2019. Red vertical line indicates the day of Wuhan lockdown. Chongqing is excluded since it is not counted in Twitter’s geolocation map.

709 We also validate that the increase in geolocated Twitter users is correlated with the decrease in human mobility. The left panel 710 of Figure A3 plots this correlation. Let Mi,t denote the mobility index for province i on date t. The x-axis plots the decrease in 711 within city movement index from January 22 (the day before the Wuhan lockdown) to February 22, Mi,Jan22 − Mi,Feb22. The 712 y-axis plots the increase in geolocated Twitter users for 30 days after the Wuhan lockdown compared to the average number of 713 geolocated Twitter users in a province before the Wuhan lockdown. This shows that the more reduction in human mobility, the 714 more increase in geolocated Twitter users, comparing to the levels before the lockdown. Hubei province experience the most 715 reduction in mobility and most increase in the number of geolocated Twitter users.

716 In Figure A1 we see increases in geolocated Twitter users in most provinces except Beijing and Shanghai. One explanation 717 for this is that Twitter users in Beijing and Shanghai left the cities during the outbreak. Mobility data supports this explanation. 718 The right panel of Figure A3 plots the relationship between moving out of a province on January 23, the day of Wuhan lockdown, 719 and the increase of number of geolocated Twitter users on the same day, compared to the average number of geolocated Twitter 720 users in a province before the lockdown. One can see that the more people moving out, the less jump in Twitter user on the day

12 | Chang et al. of lockdown. Beijing, Shanghai, and Guangdong all experienced large outflows of individuals on the day of Wuhan lockdown. 721

2 Inner Mongolia 1.9 Tibet Hubei 1.8 2.9 Shanxi

Gansu Xinjiang Hebei 1.7 Shanxi Anhui 2.5 Gansu Anhui 1.6 Inner Mongolia Guangxi Shandong Average Number of 2.1 JiangxiHunanHubei Average Number of Guangxi 1.5 Users Geo−Locating HeilongjiangJilin Users Geo−Locating Hebei Heilongjiang Hunan QinghaiNingxia Shandong Guizhou Henan Each Day After 1.7 Hainan Henan Each Day After 1.4 Wuhan Lockdown Liaoning Wuhan Lockdown Jiangxi Fujian 1.3 Shaanxi a 50 Guizhou Jiangsu a 50 Jilin YunnanShaanxi Liaoning Xinjiang 1.3 Hainan Sichuan 100 Tianjin Guangdong 100 1.2 Sichuan a Zhejiang a Ningxia Qinghai Fujian a 150 a 150 1.1 Yunnan Guangdong Jiangsu 0.9 Zhejiang 1 Beijing

Jump in Geo−located Tweets on Jan 23 on Jan in Geo−located Tweets Jump Shanghai Tianjin Jump in Geo−located Users after 30 days Jump Beijing 0.9 Shanghai

1 2 3 4 0 10 20 30 Decrease in Within City Movement Index (Feb 22 − Jan 22, Average across Cities) Moving Out Index on Jan 23

Fig. A3. Reduction in within city movement and increase in geolocated Twitter users during the month of Wuhan lockdown (left); degree of moving out and increase in geolocated Twitter users on the day of Wuhan lockdown (right). Note: The left panel plots the correlation between decreased mobility and increased geolocated Twitter users during the first 30 days of Wuhan lockdown. The x-axis plots the decrease in within city movement index from January 22 (the day before Wuhan lockdown) to Feb 22. The y-axis plots the increase in geolocated Twitter users for 30 days after the Wuhan lockdown compared to the average number of geolocated Twitter users in a province before the Wuhan lockdown. The right panel plots the relationship between moving out of province and the increase of geolocated Twitter users on January 23, the day of Wuhan lockdown. Estimates and day of lockdown are drawn from a five term polynomial regression on the number of unique geolocated Twitter users per day after the lockdown. These province-by-province polynomials are displayed over the raw data in Figure A1.

Since the period of Wuhan lockdown overlaps with the Chinese Lunar New Year, increased Twitter usage could partly be 722 due to general boredom during the New Year. To explore New Year versus pandemic effects, we normalize both mobility and 723 number of Twitter users in 2020 by those in the same period in 2019. To do so, we first adjust the dates in 2019 backwards for 6 724 days to match the dates of 2020 Lunar New Year. Then, we create normalized mobility and Twitter usage. Specifically, denote 725 Mi,y,t the mobility index and Ti,y,t the Twitter usage for province i in year y on date t. The normalized mobility index would 726 be 727 Mi,2020,t/Mi,2019,t 728 and the normalized Twitter usage would be 729 Ti,2020,t/Ti,2019,t. 730

We then plot the weekly change in mobility and Twitter usage after Wuhan lockdown, comparing to the period before Wuhan 731 lockdown. Figure A4 shows the plots. In mathematical notations, for the first week of Wuhan lockdown, we plot 732

Mi,2020,Week 1/Mi,2019,Week 1 733 Mi,2020,Week 0/Mi,2019,Week 0 on the x-axis and 734 Ti,2020,Week 1/Ti,2019,Week 1 735 DRAFTTi,2020,Week 0/Ti,2019,Week 0 on the y-axis. This is shown in the top left panel in Figure A4. The other panels shows the corresponding ratios for the 2nd, 736 3rd, and 4th week, respectively (all relative to the week before lockdown). 737 Figure A4 shows that we still find effects of reduced mobility on increased Twitter usage, after adjusting for the decrease in 738 movement driven purely from New Year, in the early periods of lockdown (at least for the first week of lockdown, the correlation 739 for the second week is not statistically significant). In the 3rd and 4th weeks, we find a general increase in Twitter usage in most 740 provinces, regardless of the relative decrease in mobility in these weeks. In other words, the mobility-induced effect specific to 741 Wuhan lockdown fades out in around 2 weeks, and there’s a general increase in Twitter usage across China that is not related 742 to reduced human mobility. This pattern suggests that the increase in Twitter usage is not driven only by people’s staying at 743 home because, if that is the case, we expect to see a continued relationship between relative reduction in mobility and increase 744 in Twitter usage, as other Provinces started to announce stay-at-home orders. This pattern is also not driven only by New Year 745 because we should not expect to see an overall increase in Twitter usage after normalizing with the same New Year period in 746 2019. 747

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 13 Reduced Mobility vs. Twitter Users (1st Week) Reduced Mobility vs. Twitter Users (2nd Week) Relative Change (Ratios) 1st Week After Wuhan Lockdown, Adjusting for Same Period in 2019 Relative Change (Ratios) 2nd Week After Wuhan Lockdown, Adjusting for Same Period in 2019

1.5 Jiangxi Jiangxi 3 Corr = −0.12 Guangxi 95% CI = [−0.46,0.25] Shanxi Anhui Hubei Beijing Sichuan Fujian Guangdong Guizhou Zhejiang Hebei Xinjiang Qinghai Yunnan 1.0 Shaanxi Shanghai Jiangsu 2 Shanxi Heilongjiang Shandong Gansu Heilongjiang Anhui Tianjin Hainan Inner Mongolia Qinghai Fujian Guizhou Henan Tibet Hubei Henan Hebei Hunan Jilin Gansu Yunnan Shandong Shaanxi Liaoning Hunan Guangxi Beijing Zhejiang Jiangsu Liaoning 0.5 1 Sichuan Shanghai Inner Mongolia Guangdong Hainan Tianjin

Geolocated Twitter Users (Relative Change) Users (Relative Geolocated Twitter Change) Users (Relative Geolocated Twitter Xinjiang Corr = −0.44 Jilin 95% CI = [−0.69,−0.1] Tibet Ningxia Ningxia 0.5 0.6 0.7 0.8 0.9 0.3 0.4 0.5 0.6 0.7 Within City Movement (Relative Change) Within City Movement (Relative Change) Reduced Mobility vs. Twitter Users (3rd Week) Reduced Mobility vs. Twitter Users (4th Week) Relative Change (Ratios) 3rd Week After Wuhan Lockdown, Adjusting for Same Period in 2019 Relative Change (Ratios) 4th Week After Wuhan Lockdown, Adjusting for Same Period in 2019

Jiangxi Tibet 5 Shanxi Tibet

Shanxi

3 4 Anhui Ningxia

Guizhou 3 Jiangxi Guangxi Heilongjiang 2 Hubei Yunnan Henan Qinghai Heilongjiang Gansu Henan Hebei Guangxi Fujian Hainan 2 Anhui Shandong Inner Mongolia Hubei Gansu Shandong Fujian Guizhou Zhejiang Sichuan Jiangsu Jiangsu Shaanxi Qinghai Shanghai Shaanxi Xinjiang Hebei Sichuan 1 Guangdong Hunan Guangdong Yunnan Geolocated Twitter Users (Relative Change) Users (Relative Geolocated Twitter Beijing Change) Users (Relative Geolocated Twitter Beijing Shanghai Inner Mongolia 1 Hainan Xinjiang Liaoning Zhejiang Ningxia Liaoning Tianjin Jilin Tianjin Jilin Hunan 0.3 0.4 0.5 0.6 0.7 0.2 0.3 0.4 0.5 0.6 0.7 Within City Movement (Relative Change) Within City Movement (Relative Change)

Fig. A4. Weekly changes in within city movement and geolocated Twitter users relative to pre-lockdown period, after adjusting for the same period in 2019. Note: We plot the weekly relative change in mobility and Twitter usage after Wuhan lockdown, adjusting for the same period in 2019 and comparing to the period before Wuhan lockdown. Specifically, denote Mi,y,t the mobility index and Ti,y,t the Twitter usage for province i in year y on date t. For the first week of Wuhan lockdown, we plot (in the top M /M T /T left panel) i,2020,Week 1 i,2019,Week 1 on the x-axis and i,2020,Week 1 i,2019,Week 1 on the y-axis. The other panels shows the same for the 2nd, 3rd, and 4th Mi,2020,Week 0/Mi,2019,Week 0 Ti,2020,Week 0/Ti,2019,Week 0 weeks, respectively.

Geo−located Twitter Users by Day 2020DRAFT versus 2019 aligned by lunar calendar Lunar New Year (day of)

1400 2020 1000

2019 600 Jan 16 Jan 20 Jan 24 Jan 28 Geolocated Twitter Users Geolocated Twitter Date

Note also that the first week of Twitter use in 2020 was not much higher than in 2019 because 2019 saw a very large number of posts on Chinese Lunar New Year. This increase was presumably related to New Year related posts, and these celebratory posts did not increase to the same extent during the start of the COVID-19 pandemic.

14 | Chang et al. S4. Effect Size 748

S4.1. New Twitter Users. This section provides rough estimates of absolute increases in Twitter use in China, and sections 749 below expand it to consider increased Twitter followings and increased Wikipedia use. Note that these are estimates for increased 750 usage on only these sites, which require that new users from Mainland China (where these sites are blocked) 1) create an account 751 to view Twitter content and 2) use cookies (to be recorded in the Wikipedia unique device data). Other sites that do not require 752 accounts could have seen larger increases, and the Wikipedia unique device counts are underestimates. 753 The top panel of Figure 2 shows a 10% long-term increase in the number of geolocating users from China. In 2019, as 754 reported in (44), Professor Daniela Stockman of the Hertie School of Governance surveyed 1,627 internet users in China and 755 found .4% of them use Twitter; the article reports that number as 3,200,000. Roughly, if the same 10% increase applies to all 756 users from China and the long-term increase reflects a new pool of users (the number of unique geotagging users in our sample 757 in May 2020 was around 10% higher than in December 2019), then 320,000 new users joined Twitter because of the crisis. 758 We can assess this estimate by considering 1) the fraction of (posting) users who geotag and 2) the number of unique geo- 759 tagging users in our data. For 1), using a sample of 100 hours of non-geolocated tweets from 2019.01.01-2020.12.31, we found 760 37,957 in Chinese. Assigning location to these tweets using the same code that was used to assign location to followers of the 761 most commonly followed accounts, we then found that 1.79% of tweets and 1.95% of users from China geotag. For 2), we 762 find that 47,389 unique Twitter users geotagged (in Chinese and in China) in our sample (note, however, that our 1% sample 763 captures approximately 56% of tweets that are geotagged). Dividing this number by 0.0195 gives us 2.4 million Twitter users, 764 2.4 suggesting that somewhere around 70% ( 3.2×1.1 ) of geotagging Twitter users in China publicly geotagged posts and were in 765 our sample. 766 Though this number is small in the context of China’s 1.4 billion inhabitants, it is nonetheless important for three reasons. 767 First, the effects in this paper are a minimum effect size for Twitter since accounts do not have to use geolocation or provide 768 an accurate self-reported location in their profile. Second, the effects documented herein focus only on one banned platform 769 (Twitter) and website (Wikipedia), and there is no reason to think the same behavior did not occur on other banned platforms 770 like Facebook, Telegram, and Instagram as well as banned websites such as Reddit or The New York Times. Third, the Chinese 771 government behaves as though these relatively small numbers threaten it. Since 2018, it has become increasingly repressive 772 in response to comments its citizens make on platforms unavailable in China. Recently, several individuals from China have 773 been arrested for comments made on platforms such as WhatsApp (owned by Facebook, unavailable inside the Great Firewall) 774 and Twitter (51). The government has also started large influence campaigns on social media platforms that are unavailable 775 domestically, including Twitter (52). If the behaviors documented in this paper were immaterial, then we believe the government 776 would not put such a priority on attempting to control speech on these platforms. 777

S4.2. Followers. In addition to causing new people to join Twitter, the crisis caused more people to follow accounts posting 778 sensitive content. Here, we estimate the number of surplus followers from China and show that they persist after the crisis, 779 perhaps at greater rates than users who follow after. 780 Figure A5 shows the absolute number of excess followers (top) and its ratio (bottom). The absolute number is the total 781 number of new followers minus the total number of predicted new followers based on the December daily average growth rate per 782 category; the bottom panel divides the new follower count by the predicted number of new followers. Several interesting patterns 783 emerge. First, the crisis clearly causes all account types to gain followers; some, such as pornography and international news 784

DRAFT 785 agencies, may even have served as early warning indicators since they receive excess followers before the Wuhan lockdown. Second, the categories with the most excess followers, citizen journalists/political bloggers and international news agencies, are 786 exactly those people would seek out in a crisis. By the end of March, 53,860 more accounts follow citizen journalists/political 787 bloggers than would have happened without the crisis; for international news agencies, 52,144. Third, normalizing for the 788 expected number of new followers reinforces that attention was paid to sensitive categories. Extra, early attention is paid 789 to the citizen journalists and activist categories (which received almost 4 times as many new followers during the lockdown 790 as we would expect based on December’s following rate), while international news agencies’ importance decreases to third 791 place. Normalizing emphasizes the increased attention activists receive since they have relatively fewer followers than the 792 other categories. Fourth, Chinese accounts increase their following of state media or Chinese officials once Hubei’s lockdown 793 lifts, though from a low base. 794 Importantly, these excess followers persist a year after the lockdown. To make this claim, we crawled the follower list of the 795 same popular accounts starting on May 31, 2021, more than one year after the first crawl, and assigned location using the same 796 procedure as before. Comparing the 2021 follower lists to 2020 shows which followers stopped following the popular accounts. 797 We then calculate the percentage of the 2020 followers that persist in 2021 by account type, follower location, and date. Table 798 A1 shows these results. 799 Accounts from China that start following the popular accounts during the lockdown period persist at the same to slightly 800 higher rates than those that start following before or after then. 87.31% of accounts from China that start following international 801 news agencies before the lockdown persist versus 89.09% that start following after the lockdown. The difference is especially 802

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 15 DRAFT

Fig. A5. Excess Followers, absolute (top) and ratio normalized by category growth rate (bottom). Growth rate is calculated based on the December 2019 average number of new followers by category.

803 stark for citizen journalists/political bloggers. Finally, since older followers should have a lower persistence rate since more time 804 has passed, it is striking that accounts that start following during lockdown have higher persistence rates than newer accounts, 805 those that start following during the seventeen days after the lockdown ends. The increased exposure to sensitive content persists 806 after the crisis passes at rates equal to or greater than for non-crisis periods.

807 S4.3. Number of unique devices accessing Wikipedia with cookies enabled. Wikipedia tracks the number of unique 808 devices that have accessed its site each day and month using a ‘privacy-sensitive access cookie’ (https://dumps.wikimedia.

16 | Chang et al. Table A1. Persistence of Followers by Account Type and Period Following Starts

Pre-Lockdown Lockdown Post-Lockdown

China Hong Kong Taiwan China Hong Kong Taiwan China Hong Kong Taiwan International News Agencies 87.31 87.71 85.51 90.80 90.19 83.20 89.09 87.35 88.70 Citizen Journalists / Political Bloggers 72.84 79.58 78.73 87.49 86.35 85.00 81.69 83.01 79.64 Activists or US / Taiwan / Hong Kong Politics 78.27 76.74 76.45 88.02 86.40 83.83 85.82 85.46 83.99 Pornography Accounts 85.56 84.28 83.00 88.84 87.51 89.52 86.32 87.19 86.88 State Media or Chinese Officials 82.72 81.38 84.62 87.99 86.36 84.94 85.90 86.61 82.45 Non-Political Bloggers or Entertainment Accounts 73.75 72.94 65.66 87.80 87.01 87.19 81.90 85.13 83.61 Note: Each cell is the percent of followers from April 2020 that still follow the six account types (row) in May 2021, by follower location and period the follower started following the account. The lockdown period is January 23, 2020 - March 13, 2020. Post-lockdown refers to March 14-April 1.

org/other/analytics/). By design, this number does not count devices not accepting cookies through private browsing (as we 809 might expect from users accessing Wikipedia from within Mainland China) and so underestimates access (see https://diff. 810 wikimedia.org/2016/03/30/unique-devices-dataset/). However, this estimate still provides some perspective on the number 811 of individuals who might be accessing the Chinese language version of Wikipedia over time. For the Chinese language version 812 of Wikipedia, 40.8 million devices accessed the site during December 2019 and 42.8 million per month during January, February, 813 and March 2020, an increase of approximately 2 million devices. 3.3 million devices accessed the Chinese language Wikipedia 814 per day in December 2019 and 3.54 million, an increase of approximately 300 thousand devices. These differences are somewhat 815 smaller when comparing to the last half of 2019 (during ongoing protests in Hong Kong) – an increase of 1 million unique devices 816 monthly during lockdown compared to July through December 2019, and an increase of 200 thousand devices daily. 817

DRAFT

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 17 818 S5. Robustness Checks

819 In this section, we assess whether the result is driven by (1) a misspecified treatment period, (2) the choice of comparison group, 820 or (3) an increase of followers due to only a few accounts. 821 Figure A6 plots the estimates based on regressions for each week before and after the lockdown. We do not see pre-treatment 822 increases in number of followers in China, and the increase starts precisely on the week of lockdown. 823 Figures A7 and A8 verify that the results in Figure 5 are not due to choosing Hong Kong for the denominator. Figure A7 uses 824 accounts from Taiwan for the denominator, and Figure A8 uses accounts in the United States. These accounts are from any user 825 using Chinese and their self-reported location is in Taiwan or the United States. Figure A9 reports the regression estimate for 826 the relative ratio of number of new followers (akin to a Difference-in-differences design with December 2019 as control period 827 and Hong Kong/Taiwan/China as control group). The result is not driven by Hong Kong-specific trend of news cycles. 828 One might also curious about whether new users stayed on Twitter at different rates. Figure A10 plots the daily unique active 829 users since their sign up dates in 2020. We don’t find that users from one location stayed on Twitter longer than others.

DRAFT

18 | Chang et al. Fig. A6. Increases in Twitter Followers from mainland China versus Hong Kong by Week

Relative Size of New Followers by Week, China / Hong Kong

International News Agencies Citizen Journalists Political Bloggers Activists or US / TW / HK Politics

1.5x

1x

0.5x -3 -2 -1 0 1 2 3 4 5 6 7 8 9 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 Pornography Accounts State Media or Chinese Officials Non-Polical Bloggers or Entertainments

1.5x Mean and 95% Confidence Interval

1x

0.5x -3 -2 -1 0 1 2 3 4 5 6 7 8 9DRAFT-3 -2 -1 0 1 2 3 4 5 6 7 8 9 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 Weeks into Lockdown

Note: Incidence rate ratios shown above are from Negative Binomial regressions of number of daily new followers on the interaction between dummy for each week and China, with December 2019 as control period and Hong Kong as control group.

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 19 Fig. A7. Increases in Twitter Followers from China versus Taiwan

New Followers Compared to Baseline, China / Taiwan

International News Agencies Citizen Journalists / Political Bloggers

3x 1.28x 3x 1.38x

2x 2x

1x 1x

0.7x 0.7x

0.4x 0.4x

Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 Removal Removal

Activists or US / Taiwan / Hong Kong Politics Pornography Accounts

3x 1.18x 3x 1.25x

2x 2x

1x 1x

0.7x 0.7x

0.4x 0.4x

Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 Removal Removal

State Media or Chinese Officials Non-Political Bloggers or Entertainment Accounts

3x 1.43x 3x 0.93x 2x DRAFT2x

1x 1x

0.7x 0.7x

0.4x 0.4x

Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 Removal Removal

Note: Gain in followers from mainland China compared to Taiwan across six types of popular accounts, relative to December 2019 average. A value greater than 1 means more followers than expected from mainland China than from Taiwan. Accounts creating sensitive, censored information receive more followers than expected once the Wuhan lockdown starts. Fewer Taiwanese users follow Chinese state media or government officials than Hong Kong users do.

20 | Chang et al. Fig. A8. Increases in Twitter Followers from China versus US

New Followers Compared to Baseline, China / US

International News Agencies Citizen Journalists / Political Bloggers

3x 1.14x 3x 1.44x

2x 2x

1x 1x

0.7x 0.7x

0.4x 0.4x

Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 Removal Removal

Activists or US / Taiwan / Hong Kong Politics Pornography Accounts

3x 1.26x 3x 1.02x

2x 2x

1x 1x

0.7x 0.7x

0.4x 0.4x

Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 Removal Removal

State Media or Chinese Officials Non-Political Bloggers or Entertainment Accounts

3x 1.47x 3x 0.99x 2x DRAFT2x

1x 1x

0.7x 0.7x

0.4x 0.4x

Dec Wuhan First Hubei Apr Dec Wuhan First Hubei Apr 2019 Lockdown Lockdown 2020 2019 Lockdown Lockdown 2020 Removal Removal

Note: Gain in followers from mainland China compared to US across six types of popular accounts, relative to December 2019 average. A value greater than 1 means more followers than expected from mainland China than from the US. Accounts creating sensitive, censored information receive more followers than expected once the Wuhan lockdown starts. Fewer US users follow Chinese state media or government officials than Hong Kong users do.

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 21 Fig. A9. Increases in Twitter Followers from China versus Others (Regression Estimate)

Relative Size of New Followers, China / Control Group

Control Group Hong Kong Taiwan US

International News Agencies

Citizen Journalists / Political Bloggers

Activists or US / Taiwan / Hong Kong Politics

Pornography Accounts

State Media or Chinese Officials

Non-Political Bloggers or Entertainment Accounts

0.75x 1x 1.25x 1.5x 1.75x Mean and 95% Confidence Interval

Note: Incidence rate ratios shown above are from negative binomial regressions of number of new followers on the interaction between indicator variables for ‘in lockdown period’ and ‘in mainland China’, with December 2019 as the control period.

Fig. A10. New Users Stay on Twitter at the Same Rates across Locations

Decay of Daily Unique User Activity

Location China 2020 Hong Kong 2020 Taiwan 2020

1.00 DRAFT

0.75

0.50

0.25

0.00 Number of Tweets (Normalized, max = 1) 0 100 200 300 Days Since Join Twitter

Note: This figure plots the daily unique active users since their sign up date using the user panel across locations. A user is considered active between their sign up date and the last day they tweet (before July 2020). We find that users stay on Twitter at the same rate across locations.

22 | Chang et al. S6. Wikipedia Page-by-Page Analyses and Country Comparisons 830

Page view data analyzed in this paper is publicly available and hosted here: https://dumps.wikimedia.org/other/ 831 pagecounts-ez/merged/. In replication materials, we will additionally provide processed and aggregated versions of the page 832 view data so that this paper’s findings can be more quickly replicated than would be possible with the above page view files. 833 Below, we show the top Wikipedia pages by relative and absolute increases in page views within each of the categories we 834 analyzed in the main text, as well as pages about the coronavirus and COVID-19 (pages considered: coronavirus, COVID-19, 835 ventilator, flu, pneumonia, fever). The largest relative increases among these pages and for current leaders were related to 836 coronavirus – the COVID-19 pandemic Wikipedia page and the head of China’s National Health Commission. Top increases 837 for pages that were blocked prior to the introduction of https on Wikipedia (after which China blocked all pages) were for an 838 activist who criticized China’s pandemic response. 839

Table A2. Top relative increases for Wikipedia pages January 24 through March 13 compared to December 2019. Note: Labels are limited to: blocked, leader, historical leader, COVID/coronavirus. All other pages are aggregated as “rest of Wikipedia”. In Figure A11 , we show the trajectories DRAFT for categories matching those analyzed for China – current leaders (using offices 840 listed in the CIA World Factbook), historical leaders, and, in Iran, pre-https blocked Wikipedia pages ( 47). 841 Russia, Germany, and Italy (none of which block Wikipedia) saw increases in current leader views without accompanying 842 increases in historical leader views. Germany and Italy did see spikes views of in historical leader pages in the weeks leading 843 up to the relaxation of lockdowns in early May, but saw no change during the initial crisis. 844 German and Russian political pages also saw an increase in political leader page views prior to their own lockdown, and 845 approximately at the same time as the announcement of widespread lockdown in Italy (see Figure A11). 846

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 23

Table A3. Top absolute daily increases for Wikipedia pages January 24 through March 13 compared to December 2019. Note: Studying average daily increases standardizes the different lengths of time before versus after the Wuhan lockdown. Labels are limited to: blocked, leader, historical leader, COVID/coronavirus. All other pages are aggregated as “rest of Wikipedia”.

Country Lockdown Start Lockdown End Historical Leaders

China January 24, 2020 March 13, 2020 Paramount Leader Hubei Lockdown

DRAFT Iran March 20, 2020 April 18, 2020 President, Supreme Leader Nowruz - Tehran Easing

Russia March 28, 2020 May 12, 2020 President Non-Working Period General Secretary (Soviet Union) Chairman, Council of Ministers (1953)

Germany March 22, 2020 May 6, 2020 Chancellor National Social Distancing

Italy March 9, 2020 May 18, 2020 Prime Minister National Quarantine

Table A4. Lockdown dates

Note: This table lists the time periods we use to estimate the effects of crisis lockdowns on Wikipedia page views, along with the offices considered for the historical leaders analysis. Each country’s lockdown involved various levels of lockdown for different parts of the countries, and so there is no single time period for us to analyze. Figure A11 displays Wikipedia page views with solid, vertical gray lines for the periods listed above.

24 | Chang et al. FA: Change in Page Views per Day FA: Page Views by Category

Lockdown Lockdown Current Leader Pages Historical Leader Pages 4

1.4 Blocked Pages (pre−https) Rest of Wikipedia (=1) 1.3 Ratio of Mobile to Desktop Views (7−day MA) 2 1.2 1.1 Total Views Total 1 1 Wikipedia Page Views Wikipedia Page 0.9 Ratio to Rest of Wikipedia 0.5 (Compared to Pre−Covid Level) (Compared to Pre−Covid Jan Mar May Jul Jan Mar May Jul Date Date

RU: Change in Page Views per Day RU: Page Views by Category

Lockdown Lockdown in Italy Lockdown Current Leader Pages 1.4 Historical Leader Pages 4 Rest of Wikipedia (=1) Ratio of Mobile to Desktop Views (7−day MA) 1.2 1.1 2 1 Total Views Total 0.9 1 Wikipedia Page Views Wikipedia Page Ratio to Rest of Wikipedia 0.5 (Compared to Pre−Covid Level) (Compared to Pre−Covid Jan Mar May Jul Jan Mar May Jul Date Date

DE: Change in Page Views per Day DE: Page Views by Category

Lockdown Lockdown in Italy Lockdown Current Leader Pages Historical Leader Pages 4 1.5 Rest of Wikipedia (=1) Ratio of Mobile to Desktop Views (7−day MA) 1.3 2 1.2 1.1 Total Views Total 1 1 Wikipedia Page Views Wikipedia Page 0.9 Ratio to Rest of Wikipedia 0.5 (Compared to Pre−Covid Level) (Compared to Pre−Covid Jan Mar May Jul Jan Mar May Jul Date Date

IT: Change in Page Views per Day IT: Page Views by Category LockdownDRAFT Lockdown Current Leader Pages Historical Leader Pages Rest of Wikipedia (=1)

4 Ratio of Mobile to Desktop Views (7−day MA) 1.4 2 1.2 1.1 Total Views Total 1 1 Wikipedia Page Views Wikipedia Page 0.9 Ratio to Rest of Wikipedia 0.5 (Compared to Pre−Covid Level) (Compared to Pre−Covid Jan Mar May Jul Jan Mar May Jul Date Date

Fig. A11. Views of Blocked, Current Leader, and Historical Leader Wikipedia Pages in Other Countries

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 25 847 S6.1. Analysis of an expanded set of historical political pages and ‘politically sensitive’ pages using 848 Wikipedia2vec. We replicated our analyses of historical Wikipedia pages and “politically sensitive” (pages specifically blocked 849 in China and Iran prior to the introduction of https) Wikipedia pages by expanding the original set of pages to a much larger set 850 of related pages. We expanded these lists of pages using Wikipedia2vec (48). This analysis assesses 1) whether the increase 851 in views of Chinese historical leaders (and the lack of increase for other languages) was a relatively narrow effect or much 852 broader one than what we see for that small set of pages and 2) whether a broader set of ‘politically sensitive’ pages are able to 853 uncover increases in page views in Iran and Russia. Because, unlike China and Iran, Russia did not provide a list of politically 854 sensitive pages (by blocking specific pages on Wikipedia), we assess Russian views of political opposition pages related to a) 855 Alexei Navalny (arguably the most prominent opposition leader in Russia) and b) a list of opposition-related pages which we 856 mine to discover increases in views – after this, we then looked to closely related pages to assess whether single page increases 857 represented broader trends or were isolated and potentially random occurrences. 858 Wikipedia2vec finds similar pages (along with other entities and words) on Wikipedia by analyzing the network of page links, 859 the co-occurrence of words, and the occurrences of specific words on pages. This analysis is accomplished using the same ap- 860 proach as in word2vec (53). At a high level, this approach involves placing words and entities into a shared n-dimensional 861 space such that words and entities are placed closed together if they frequently share contexts (e.g. page links or co-occurring 862 words). Shared contexts must occur beyond what would be expected from the frequency of a word or page, which is accom- 863 plished through ‘negative sampling’ – predicting the co-occurrence of words and entities against frequency weighted sampling 864 of negative cases. Once in the n-dimensional space, we can find the most similar entities (pages) for any given entity (or the 865 mean of a set of entities’ projections) using cosine similarity – and we can incorporate dissimilar entities in this calculation 866 by flipping the sign those entities’ locations when calculation the mean of a set of entities. Wikipedia2vec can be run with 867 hyperparameters that affect the size of the n-dimensional space and the exact weighting scheme used in negative sampling. 868 Our estimations for each language used the same default settings as the wikipedia2vec pre-trained embeddings provided at 869 https://wikipedia2vec.github.io/wikipedia2vec/pretrained/ with the number of dimensions set to 100. 870 For each set of pages (historical leaders, blocked pages, current leaders, Russian opposition pages), we found the top 100, 871 250, 500, and 1,000 pages that were most similar according to Wikipedia2vec. For historical leaders, we expanded to historical 872 leader related pages not related to the current leader – using the current leader as a dissimilar case – and we expanded to current 873 leader related pages not related historical leaders. With these sets, we re-estimated the changes in views during the first 30 days 874 of lockdown. This excludes the late lockdown spikes in historical leader views visible for German and Italian (visible in Figure 875 7 in the main text and in Figure A11 above). Note that the German increases in views of historical leaders (in those figures and 876 in the results below) began well prior to the German lockdown (in February). 877 For Alexei Navalny specifically, we also manually collected a list of Wikipedia pages closely related to his opposition 878 activities, and re-estimated changes in lockdown for each of these pages. The list of Russian opposition-related pages checked 879 for increases is shown in Table A5. 880 For the pages previously blocked in Iran, (47) provided labels for the category of each blocked page: academic, artis- 881 tic/cultural, drugs or alcohol, human rights, media and journalism, other, political, profane non-sexual, religious, and sex and 882 sexuality. We also replicated our analyses for the Persian language set subset to page categories human rights, media and 883 journalism, and political. 884 The findings from these analyses are displayed in Figures A12, A13, and A14, and we also show findings for current leaders 885 in Figure A15. In each cluster of estimatesDRAFT in the top panels, the first is the estimate for the seed pages (and is colored yellow 886 for historical pages, red for blocked/‘politically sensitive’ pages, purple for current leaders). These exclude estimates for seeds 887 which we mined for increases (i.e. selected them only because we saw increases during lockdowns – after a Bonferroni multiple 888 testing correction). Given many tests when looking for increases, these pages have estimates that could very likely reflect 889 random variation in page views, even though we are relatively certain that the increases were not zero, given the multiple 890 testing correction. 891 Across the results, we see 1) that the increase in historical leader page views in Chinese also applies to a much larger set of 892 pages and page views (bottom panel) and 2) we do not see comparable increases in historical leader pages or politically sensitive 893 page views in other languages, despite increased interest in current leaders across almost all languages analyzed. 894 In the manual Alexei Navalny analysis, we see that views for his page specifically did rise and that this rise was comparable 895 to what we see for historical leaders in Chinese. However, unlike the broad increase in views in China, we did not see similar 896 increases for any other Navalny-related pages – and only one of the 9 considered showed a statistically significant increase 897 without a multiple-testing correction (falling just short of significance at a 0.05 level after a Bonferroni correction for 9 tests).

26 | Chang et al. Relative increase in historical leader views: first 30 days of lockdown Leader set order: Seeds 1.50 n most similar pages − 100 not related to 250 1.25 current leader 500 1,000

1.00

0.75

(compared to overall increase in views) (compared to overall Chinese German Italian Persian Russian Increase in Views of Historical Leader Pages Increase in Views Number of Historical Leader Page Views by Page Set

30.0 M

Top 1k pages 500

250 10.0 M 100

3.0 M Seed pages DRAFT

Number of Historical Leader Page Views Number of Historical Leader Page Chinese German Italian Persian Russian

Fig. A12. Changes in views of historical leader Wikipedia pages (expanded set of pages). German increases in views of historical leaders began in February (see Figure A11 above)

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 27 Relative increase in "sensitive" political page views: first 30 days of lockdown Leader set order: Seeds 1.50 n most similar pages − 100 1.33 250 500 1.00 1,000

0.75

0.50

(compared to overall increase in views) (compared to overall Chinese Persian Persian Persian Russian Russian Blocked In China: Blocked In Iran: Blocked In Iran: Blocked In Iran: Alexei Navalny Biggest

Increase in Views of "Sensitive" Political Pages Political of "Sensitive" Increase in Views Pre−Https Biggest Pre−Https Pre−Https, Political Increase Seeds Increase Seeds

Number of Sensitive Political Page Views by Page Set

Seed pages 100.00 M

10.00 M

Top 1k pages 1.00 M 500

250 0.10 M 100 DRAFT

Number of Sensitive Political Page Views Page Political Number of Sensitive Chinese Persian Persian Persian Russian Russian Blocked In China: Blocked In Iran: Blocked In Iran: Blocked In Iran: Alexei Navalny Biggest Pre−Https Biggest Pre−Https Pre−Https, Political Increase Seeds Increase Seeds

Fig. A13. Changes in views of ‘politically sensitive’ Wikipedia pages (expanded set of pages).

28 | Chang et al. Relative increase in Alexei Navalny related page views (in Russian): first 30 days of lockdown

1.50 1.33

1.00

0.75

0.50

Alexei Navalny 2017−2018 2019 He Is Not Russia of Russian Russian protests Moscow protests Dimon to You the FutureOpposition (compared to overall increase in views) (compared to overall 2011−2013 2018 Russian Anti−Corruption Party of crooks Coordination Russian protests pension protests Foundation and thieves Council Increase in Views of Pages Related to Alexei Navalny Related to Alexei of Pages Increase in Views Number of Page Views

1.000 M

0.100 M

Number of Page Views Number of Page DRAFT 0.010 M Alexei Navalny 2017−2018 2019 He Is Not Russia of Russian Russian protests Moscow protests Dimon to You the FutureOpposition 2011−2013 2018 Russian Anti−Corruption Party of crooks Coordination Russian protests pension protests Foundation and thieves Council

Fig. A14. Changes in views of Alexei Navalny related Wikipedia pages. The Alexei Navalny-related pages in this figure are listed in alphabetical order.

Chang et al. PNAS | July 1, 2021 | vol. XXX | no. XX | 29 2011-2013 Russian protests He Is Not Dimon to You 2014 anti-war protests in Russia Human rights in Russia 2017-2018 Russian protests List of journalists killed in Russia 2018 Russian pension protests Media freedom in Russia 2019 Moscow protests Mikhail Khodorkovsky Alexander Litvinenko Novaya Gazeta Alexei Navalny Open Russia Anna Politkovskaya Opposition to Vladimir Putin in Russia Anti-Corruption Foundation Party of crooks and thieves Assassination of Anna Politkovskaya Pussy Riot Assassination of Boris Nemtsov Russia of the Future Boris Berezovsky (businessman) Russian Opposition Coordination Council Boris Nemtsov Sergei Magnitsky Corruption in Russia Sergei Yushenkov Table A5. List of opposition-related pages in Russian that were checked for significant increases during lockdown.

Note: Russia did not block specific Wikipedia pages prior to Wikipedia’s introduction of https. Because of this, we do not have a government-provided list of politically sensitive or objectionable content. As an alternative, we mine a manual list of government opposition-related pages, and then check whether for those increases were narrow and perhaps random (i.e. only occurred for those specific pages) or represented broad increases similar to those seen for historical and previously blocked pages in China. This table lists those Wikipedia pages (translated) that were checked for significant associations during the Russian lockdown period when compared to December 2019. Pages with statistically significant increases (p < 0.05) after a Bonferroni multiple testing were used as seeds when expanding with Wikipedia2vec. These “biggest increase seeds” are in bold above.

Relative increase in current leader views: first 30 days of lockdown

Leader set order: 3.00 Seeds n most similar pages − 100 2.00 not related to 250 historical leaders 500 1.50 1,000

1.00

0.75

(compared to overall increase in views) (compared to overall Chinese German Italian Persian Russian Increase in Views of Current Leader Pages Increase in Views Number of CurrentDRAFT Leader Page Views by Page Set 20.0 M

10.0 M

Top 1k pages

500 250 Seed pages100 5.0 M

Number of Current Leader Page Views Number of Current Leader Page Chinese German Italian Persian Russian

Fig. A15. Changes in views of current leader Wikipedia pages (expanded set of pages).

30 | Chang et al.