Rumors were defined as reports that disseminate false information—information that is inconsistent with existing scientific evidence. In this study, all COVID-19 rumors were confirmed to be incorrect based on current scientific evidence by health authorities. In 3 rounds of discussions, we categorized the rumors that we identified into 8 groups: (1) prevention, (2) diagnosis/treatment/assistance, (3) origin and spread of COVID-19, (4) consequences of COVID-19, (5) disease statistics, (6) return to work or back to school, (7) imported cases, and (8) all others. We arranged a trained researcher to categorize each rumor into 1 of the 8 categories. One-third of the rumors were randomly extracted for a second evaluation by 3 other independent researchers. Consistency between both sets of evaluations was excellent (κ=0.96, P<.001).

Data sources were derived through 4 steps. First, based on preliminary search results, we identified 20 prominent and publicly accessible web-based platforms available in China to search for COVID-19–related rumors. These platforms included 9 websites, 5 official Sina Microblog accounts, and 6 WeChat public accounts (Table S1 in Multimedia Appendix 1).

Second, we retrieved all media reports regarding 4 major rumors and classified the media reports into 3 categories: those disseminating rumors, those providing correct information, and those with ambiguous information. We conducted this classification using the ZhongQing HuaYun web-based public service platform (CYYUN), a free public database platform licensed by the Central Committee of the Communist Youth League. CYYUN automatically retrieves media reports from WeChat accounts officially certified by governmental or business entities, Microblog accounts, public forums, news websites, print media, blog accounts, videos, news apps, major overseas media sources, and other mainstream national news sources. CYYUN data are updated approximately every 5 minutes.

Third, Sina Microblog Topic was used to gather readers’ comments in response to media reports of the rumor event with the most public comments. Sina Microblog Topic is a webpage that summarizes personal microblogs with a tag (“#topics-related keywords#”), allowing readers to post personal comments below the microblogs. Comment data from Sina Microblog Topic are freely accessible in China.

Last, the number of daily confirmed COVID-19 cases at the national and provincial levels were derived from the official website of the National Health Commission of China.

We implemented 2 strategies to assess the effectiveness of authorities’ refutation of rumors.

We compared the distribution of the 2 types of media reports (those perpetuating the rumor and those correcting the rumor) before and after the rumor was formally refuted in the media. Because there is no standard criterion to determine the launch time of a formal rumor-refuting effort, we studied the temporal distributions of the number of correct reports and the number of rumor reports per hour for all 4 cases (Figure S1 in Multimedia Appendix 1). Based on these empirical distributions, we found that the number correct reports per hour generally increased substantially in subsequent hours once it reached over 50 reports. Thus, we considered the refutation as formally launched from the first hour in which more than 50 reports refuting the rumor were posted. We considered 3 phases for our analysis: prerefutation, refuting the rumor, and postrefutation.

We examined differences in the distribution of positive, neutral, or negative emotions in comments responding to reports disseminating the rumor and in reports correcting the rumor. We focused this analysis on case A, the rumor that received the most reader comments and was sustained the longest, thus having the largest social impact among all COVID-19 rumors during the early stages of the pandemic in China. To conduct the analysis, we examined both the frequency of reader comments and the time intervals between the first and last comments from the same readers when more than 2 comments were posted.

Based on publicly accessible Microblog users’ unique identity accounts, we compared the distribution of comments in response to case A from readers who posted 2 comments or more. For each user, the first and last comments during the study period were considered, and the pairs were divided into 4 groups: Group A, both comments occurred in response to reports disseminating correct information; Group B, both comments occurred in response to reports disseminating rumors; Group C, the first comment occurred in response to a report disseminating a rumor, and the last in response to correct information; and Group D, the first comment occurred in response to a report disseminating correct information and the last in response to a rumor. This grouping allowed us to examine the potential effect of the order of reading media reports (first reading reports disseminating rumors and then reading correct reports or vice versa).

All statistical analyses were performed using the R software package (version 4.0.0). Proportions and Spearman rank correlation coefficient with 95% confidence intervals were calculated. The chi-square test was used to test the difference in the proportion of rumor reports and correct reports between prerefutation and postrefutation. All tests were 2 tailed, and P<.05 was considered statistically significant.

This study was approved by the ethics committee of Xiangya School of Public Health (XYGW-2020-43). This study used open-access social media data and excluded all personal information. The ethics committee determined that this study was exempt from requiring informed consent.

After screening 19,683 recorded rumors on 20 websites and social media accounts, we obtained 1829 unique COVID-19–related rumors in China that appeared from December 1, 2019 to April 15, 2020 (Figure S2 in Multimedia Appendix 1).

The frequency of rumors began with a quiescent period (<6 rumors per day before January 20, 2020), but then rose rapidly to reach an initial peak on January 25, 2020 (n=75). After a second peak on February 7, 2020 (n=82), the frequency started to decrease gradually (Figure 1a). “Pneumonia,” “WeChat,” and “spread” were the most commonly occurring words, with high-frequency words varying across the 6 stages of the COVID-19 epidemic (Table S2 and Figure S3 in Multimedia Appendix 1).

Figure 1b shows the large variation in the number of rumors across different provinces during the study period. Of the 1829 rumors, 399 (21.8%) rumors involved more than 1 province, but the majority (1430 rumors) involved a single province. The largest number occurred in Hubei (n=186), followed by Zhejiang (n=137) and Guangxi (n=121). Ningxia, Tibet, Xinjiang, Qinghai, Hainan, and Taiwan each had fewer than 10 rumors.

The content of the rumors evolved greatly across the 6 stages of the COVID-19 epidemic in China (Figure 1c). Rumors related to prevention and disease statistics were most common, accounting for 72.2% (39/54), 73.3% (335/457), and 64.8% (414/639) of all rumors in stages 2, 3, and 4, respectively. The percentage of rumors involving the categories diagnosis/treatment/assistance (χ₁²=6.352, P=.01), return to work or back to school (χ₁²=148.094, P<.001), and imported cases (χ₁²=126.04, P<.001) significantly increased during stages 5 and 6.

We considered in greater detail the 4 most widespread rumor cases (case A, the novel coronavirus can be prevented with “Shuanghuanglian”; case B, the novel coronavirus is the SARS coronavirus; case C, medical supplies intended for assisting Hubei were detained by the local government; and case D, asymptomatically infected persons were regarded as diagnosed COVID-19 patients with symptoms in official counts). Most reports were from personal social media accounts for case A (33,870/53,798, 63.0%), case B (21,234/24,436, 86.9%), and case C (6411/7101, 90.3%). For case D, most reports were from websites and social media accounts of news agencies (1850/3436, 53.8%) (Figure 2). Personal social media accounts were the major source of rumor reports (case A: 7648/10,664, 71.7%; case B: 14,696/15,902, 92.4%; case C: 3911/3943, 99.2%; case D: 322/323, 99.7%) and ambiguous reports (case A: 3116/4968, 62.7%; case B: 2582/2592, 99.6%; case C: 382/383, 99.7%; case D: 156/156, 100.0%).

In addition, there was a strong correlation between the frequency of rumors and the daily number of newly confirmed COVID-19 cases at the national level from December 30, 2019 to April 15, 2020 (ρ=0.88, 95% CI 0.81-0.93, P<.001). A spatial correlation between the frequency of rumors and the cumulative number of confirmed cases up to April 15, 2020 also emerged at the provincial level (ρ=0.83, 95% CI 0.61-0.94, P<.001) (Figure S4 in Multimedia Appendix 1).

Totals of 53,798 media reports (case A), 24,436 media reports (case B), 7101 media reports (case C), and 3436 media reports (case D) were obtained after eliminating duplicates (n=3788) and irrelevant reports (n=4036) (Figure S2b in Multimedia Appendix 1).

The proportion of rumor reports significantly decreased after the rumors were refuted by health authorities for case A (before: 534/666, 80.2%; after: 5685/45,657, 12.5%), case C (before: 1978/2101, 94.1%; after: 908/3796, 23.9%), and case D (before: 120/860, 14.0%; after: 192/2431, 7.9%) (P<.001) (Table 1); however, the proportion of rumor reports increased from 10.8% (50/463) to 60.5% (12,004/19,846) for case B after the official clarification with correction information was released (P<.001).

For case A (the novel coronavirus can be prevented with “Shuanghuanglian”), we identified a total of 60,744 comments posted by 54,290 microblog readers in response to 42 articles disseminating rumors and 78 articles refuting rumors and providing accurate information. Most readers posted once (49,535/54,290, 91.2%) or twice (3789/54,290, 7.0%) (Figure 3a), and most readers’ comments expressed emotions with a negative connotation (eg, anger, anxiety, fear) in response to both types of microblog articles. However, the proportion of comments with negative emotions for articles disseminating correct information was significantly higher than the proportion for articles disseminating rumors (correct: 42,485/52,811, 80.5%; rumor: 5271/7933, 66.4%, χ₁²=804.55, P<.001) (Figure 3b). Among readers who posted multiple comments, 60.5% (2875/4755) posted their first and last comments within half an hour (Figure 3c), and the proportion of negative comments posted by those readers did not change significantly when comparing their first versus last posts, no matter what type of article they read (both correct: χ₁²=0.315, P=.58; both rumors: χ₁²=0.025, P=.88; first rumor and last correct: χ₁²=1.287, P=.26; first correct and last rumor: χ₁²=0.033, P=.86) (Figure 3d).

This study yielded 3 key findings. First, the contents and the number of rumors were strongly associated with the development of the COVID-19 epidemic in China, both temporally and geographically, and reports from personal social media accounts accounted for 71.7% (7648/10,664), 92.4% (14,696/15,902), 99.2% (3911/3943), and 99.7% (322/323) for cases A, B, C, and D, respectively, of all rumor reports. Second, releasing news to correct rumors decreased the dissemination of media articles for 3 of the 4 selected rumors (cases A, C, and D) but did not reduce news disseminations for case B. Third, the likelihood that readers posted negative comments in response to rumor reports and to reports disseminating correct information differed insignificantly (both correct reports: χ₁²=0.315, P=.58; both rumors: χ₁²=0.025, P=.88; first rumor and last correct report: χ₁²=1.287, P=.26; first correct report and last rumor: χ₁²=0.033, P=.86).

The findings of previous studies, regarding fluctuations in the number of rumors reported on a daily basis, the geographic variations for those reports across China [14], and major topics of the COVID-19 rumors [14,15], were generally concordant with our findings. The strong and positive correlation between reported rumors and the development of the COVID-19 epidemic likely reflects extreme tension, fear, and anxiety about the novel contagious disease among the population, especially residents of the regions most affected by the epidemic [16], such as Hubei, Zhejiang, and Guangxi. The likely reason for this was that the rapidly growing epidemic provoked emotional responses among inhabitants, leading them to scramble for any information they could find to alleviate their anxieties and fears. When the residents in those regions were unable to obtain timely official answers about major public concerns related to the COVID-19 epidemic, rumors emerged and were disseminated quickly [17]. For example, during the early stages of the COVID-19 epidemic in China, the question of how to dispose of used face masks emerged as a major public health concern. The question aroused 2031 microblog posts, but only 10 official Sina Microblog accounts released authoritative information on the issue before January 23, 2020 [18]. As a result, a total of 14 versions of misleading rumors emerged immediately on this topic, and 4740 articles disseminated improper strategies to dispose of discarded masks, such as heating them in the oven, scrubbing them with water, or cutting them into pieces [18].

This study presents a unique finding that personal social media accounts were the dominant source of rumor reports. This may reflect the combined effect of flourishing personal social media accounts [19], an inability for average citizens to detect misinformation [20], and amplification effects through (ie, an “echo chamber”—people are susceptible to peer influences when engaging in a homogeneous online cluster with others who share similar interests) [21].

Unexpectedly, we found that websites or news-based social media accounts were the second most common source of rumor reports. This may be a result of inadequate fact-checking of contents by the news industry or deliberate reports designed to attract public attention to these news media outlets [22,23]. As an example, the most widely spread rumor (Case A, the novel coronavirus can be prevented with “Shuanghuanglian,” a traditional Chinese medicine used to treat the common cold) was first released by Weibo of Xinhua, an official news platform with nearly 100 million subscribers [24]. That release quickly led to a Sina Microblog topic titled “The novel coronavirus can be prevented with Shuanghuanglian,” attracting 1.55 billion readings and 423,000 reader comments [25] and leading to panic buying of Shuanghuanglian.

The release of news to correct rumors was successful in reducing the spread of rumors in cases A, C, and D. This confirms that it is possible to mitigate the spread of rumors through authoritative releases of correct information in a timely manner [26]. Case B did not follow the same pattern, probably due to an unintentional slip of tongue by a government spokesperson at the 19th press conference in Hubei province on February 9, 2020 [27]. The spokesperson unintentionally and incorrectly confirmed the rumor, leading to a rapid increase in the number of rumor reports concerning Case B (from 41 to 15,767; Figure S5 in Multimedia Appendix 1) within 3 hours of the start of the press conference.

Strikingly, most reader comments in response to articles were negative, whether the articles reported rumors or correct information, and the proportions of negative comments were similar between both types of articles. These results match those of previous reports [28,29] and may reflect the fact that official correction news tends to disseminate correct information but largely overlooks any attempt to provide emotional support or empathy to affected people. Such factual presentation without empathy leads to unproductive dialogue with the emotional public, who release their negative sentiments in writing [29].

Our findings have 3 major implications. First, they underscore the importance of releasing scientifically accurate information in response to major public health concerns in a timely manner. Such releases help prevent the spread of rumors. To be effective, they require close cooperation between government officials, scientists, and the media.

In emerging health crises, such as the COVID-19 pandemic, public concerns inevitably arise that cannot always be answered immediately with accurate scientific evidence. In the absence of guidelines to satisfy public angst and curiosity for critical public health information while science progresses, basic guidance might be established by the WHO or academic experts to direct the release of available information through multiple channels, including government websites, news websites, and official and personal social media accounts. Such releases might reduce tension and anxiety among the public.

Second, the official release of correction articles tended to reduce rumor reports in China, suggesting similar effects could be expected in other countries. Success with this strategy would require a real-time web-based monitoring system to detect rumor reports and close cooperation between government, media companies, and public health experts to release authoritative correction articles that refute rumors, with the caveat that such a strategy may not work well in countries where government and public health experts provide conflicting advice, such as in the United States for much of the situation with COVID-19 in 2020 [30].

Last, our findings suggest rigorous public health research is needed to generate solutions to unsolved questions, including how to effectively detect rumor reports as early as possible, which solutions might be best for releasing and disseminating correction articles, and how to effectively alleviate negative sentiments held by the public. In addition, the effectiveness of alternative postrumor approaches such as fact-checking and retractions (eg, deleting posts) should be evaluated.

This study has limitations. First, our data collection approach likely underestimated the number of COVID-19 rumors, by missing some comments in response to Sina Microblog topics, since some were likely quickly detected and deleted by platform providers as part of efforts to control rumors [31]. Second, the lack of detailed data limited our ability to explore why most reader comments in response to media articles that disseminated correct information were negative and how to alleviate such negative sentiments to promote public support on fighting the COVID-19 pandemic. Further research is needed to address this critical public health matter. Third, we could not determine if any rumors we studied were published artificially by robots. As readers generally do not know who posted the reports, the role of robots is unlikely to influence our findings substantially. However, future studies may collect data on readers’ responses to rumors spread by robots versus those spread by humans to quantify potential differences in their social impact and to develop specific interventions to deal with them separately.

The number of rumors about COVID-19 fluctuated greatly in China during the early months of the pandemic. The frequency of rumors was highly correlated with the magnitude of the epidemic. Rumor-disseminating reports emerged primarily from personal social media accounts, and releasing rumor-correcting news substantially reduced the spread of rumor reports. Reader comments in response to media articles disseminating rumors and to articles disseminating correct information expressed nearly the same percentage of negative emotions.

Individual governments and relevant international organizations such as the WHO should take immediate actions to develop regulations or guidelines to help social media companies and users of social media prepare and release reports properly and efficiently to prevent and control rumors, especially when scientific knowledge is limited during a health crisis such as the COVID-19 pandemic. In the meantime, our results suggest refutation reports were generally successful in reducing spread of rumors, and governments should encourage timely release of correct information to minimize the impact of rumors.