https://orcid.org/0000-0002-4550-5411
Figures
Abstract
Finding ways to improve police legitimacy and police-community relations has for long been an important social issue in the United States. It becomes particularly urgent following the murder of George Floyd on May 25th, 2020. An emerging area that holds potential in remediating police-community relations pertains to the use of social media by police. Yet, this body of research stays highly exploratory (e.g., case studies based on a small sample of agencies) and different viewpoints exist regarding the objectives of police social media usage. The current study identified 115 large police departments in the U.S. and collected their tweets over a 4-month period between 4/1/2020 and 7/31/2020. We investigated how police agencies (both individually and as an aggregate) leveraged social media to respond to the nationwide protests directed at the police and community reactions to such responses. We found that police agencies tweeted more frequently in the immediate aftermath of the murder and posted an increased number of civil-unrest related tweets. The public showed a greater interest in engaging with law enforcement agencies (i.e., average favorite and retweet counts) following the murder. A great variability emerged across agencies in their responses on social media, suggesting that examining only a handful of agencies or a particular dimension of social media usage would limit our understanding of police behaviors and citizen interactions on social media. In conclusion, we suggested a few avenues for future research (and practices) on responsible and effective use of social media by police, while pointing out the challenges associated with such inquiries.
Citation: Dong B, Wu X (2022) Reaching and engaging people: Analyzing tweeting practices of large U.S. police departments pre- and post- the killing of George Floyd. PLoS ONE 17(7): e0269288. https://doi.org/10.1371/journal.pone.0269288
Editor: Rogis Baker, Universiti Pertahanan Nasional Malaysia, MALAYSIA
Received: January 4, 2022; Accepted: May 17, 2022; Published: July 14, 2022
Copyright: © 2022 Dong, Wu. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The study’s minimal underlying dataset was uploaded to a stable, public repository. Please see https://www.kaggle.com/datasets/bdresearch/pd-tmline.
Funding: The authors received no specific funding for this work.
Competing interests: Xiaoyun Wu is an employee of the National Policing Institute (formerly the National Police Foundation), a not-for-profit organization committed to conducting research, training, and technical assistance in policing. This does not alter our adherence to PLOS ONE policies on sharing data and materials.
Introduction
On May 25, 2020, George Floyd, a 46-year-old Black man arrested on suspicion of using a counterfeit $20 bill, was murdered by Derek Chauvin, a 44-year-old white police officer in the Minneapolis Police Department who knelt on Floyd’s neck for almost 10 minutes. Following the killing of George Floyd, the social protest movement across the United States (U.S.) in the summer of 2020 led to a new round of contentious debates on police work, with many calling for fundamental police reform (e.g., to “defund the police”) and enhanced accountability. Inquiries on ways to improve police legitimacy and police-community relations have not been more urgent, as law enforcement agencies continue to represent the primary source of social control and the public rely on them for protection in a turbulent social environment (e.g., the recent spike of homicides amidst an ongoing pandemic) [1–5].
An emerging area that holds potential in remediating police-community relations pertains to the use of social media by police [6]. Throughout the protests, we saw police departments across jurisdictions engaging in public conversations and other activities on social media, some of which garnered extensive public attention. Despite some preliminary evidence suggesting the potential benefits of a police presence on social media, research on police use of social media has been relatively scant, as well as mixed [7–10]. Advocates for increasing police presence on social media suggested a myriad of potential benefits, including improving community outreach, investigation, and crime prevention [11, 12]. Inversely, police social media usage was argued to mainly fulfill a function of socialization (i.e., people internalize how police think and what police value) or legitimation (i.e., police justify contested actions through direct information sharing), thereby mediating public pressure for reform [13]. Additionally, there are concerns that police mostly engage in shallow, non-dialogical interactions with the public on social media [14–16].
In light of these dissimilar viewpoints of police social media usage, the current study seeks to understand how police in the U.S. leveraged social media to respond to the killing of George Floyd and to the nationwide protests that ensued. The study is motivated by 1) an increasing social media presence of law enforcement agencies [17], 2) ongoing frictions between police and disadvantaged and minority communities, 3) perceived benefits and challenges of social media usage among police practitioners, and 4) limited research covering police use of social media and its impact across the nation. The scale and intensity of the protest, amidst a global pandemic that ushered in a period of rapid growth in digital communication generally [18] and in policing [19], provide us with an exceptional opportunity to examine police social media usage and community reactions to it.
Police use of social media
Police presence on social media has become growingly prevalent in the U.S. and other countries over the past decade. In a recent law enforcement use of social media survey in the U.S., Kim and colleagues noted that about 96% of their agency respondents (N = 539) affiliated with the International Association of Chiefs of Police (IACP) have a social media account as of 2016, with most agencies adopting social media usage between 2010 and 2014 [17]. Social media is thought to also serve as a technological driver of open government initiatives. The Open Government Directive of the Obama administration propels government agencies to provide more information to the public and to establish mechanisms through which public feedback can be collected and used to evaluate and improve government performance [9, 20, 21]. This trend continues to be facilitated by the COVID-19 pandemic, which has prompted digitization of government communications and transactions at an unprecedented rate and urged many police agencies to shift their community engagement activities online through social media platforms [19].
As a direct communication channel, social media allows the police to bypass traditional news media and reach a wider audience at a low cost and with greater efficiency [22, 23]. Policing scholarship has established that law enforcement agencies commonly seek to gather intelligence, enhance crime prevention and investigation, humanize the agency, engage in image-building activities, and improve their relations with the public through social media usage, which are consistent with the overall goals of community-oriented policing [11, 14, 24, 25]. At times of immediate crisis, police social media usage has the advantages over traditional news outlets to deliver instant messages to the mass. By exerting authority and providing immediate responses under exceptional circumstances, police agencies’ social media accounts often become the trusted source for information and can garner wide societal attention. Such instances were found during natural disasters, demonstration and social riots, terrorist attacks, among others [8, 26–28].
Nonetheless, the actual impact of social media usage in transforming police work and remediating police-community relations well depends on the way in which police agencies use it. Prior research suggests a variability across law enforcement agencies in social media usage, depending on agency organizational goals and pre-existing communication strategies [9]. This may be particularly evident in crisis situations. The crisis communication literature suggests that image-making and repair are one main motivation behind individual or organizational responses to crises [29]. Image is considered threatened when an organization or individual has committed or was responsible for an offensive act. Specifically, image repair theory identifies several approaches in response to accusations or damages including denial, evasion of responsibility (e.g., provocation, defeasibility, accident, or good intentions), reducing offensiveness of event (e.g., bolstering, minimization, differentiation, etc.), and mortification and corrective action attempt to repair an image without directly dealing with blame or offensiveness [30].
While social media may be used to promote a more open culture in police departments [31], social media platforms such as Twitter may also be used to publicize police-curated content unfiltered by traditional mass media, serving the purposes of deflecting institutional change (e.g., through socialization and/or legitimation) and mediating public pressure for reform. In a case study examining the New York Police Department (NYPD)’s daily Twitter posts in 2018 and an in-depth analysis of public reactions on Twitter to a contested NYPD shooting (i.e., the killing of Saheed Vassell), Cheng concluded that police social media usage represents “selective transparency” and mainly provides police with the technological capacity to “shape social memories while avoiding various forms of public accountability” (p.413) [13]. In the case of George Floyd, police as a profession have received heavy criticism for the long-standing racial disparities in policing outcomes and a string of fatal encounters between police and black citizens in recent years (e.g., Michael Brown, Freddie Gray, Breonna Taylor, etc.). As such, most U.S. police agencies would feel compelled to respond to the killing of George Floyd and associated protest activities through an image repair angle (e.g., emphasizing their role in fighting crime to ease public outrage or devoting an increased share of social media posts to racial justice related posts as a corrective action).
Current study
The bulk of research on police use of social media has emerged within the past decade, scattered in such areas as criminology, sociology, public administration, communication, and information technology. It is far from clear what constitutes responsible and effective police public engagement on social media and whether actual uses of social media by police live up to the ideal of a community-oriented policing approach or mainly serve self-interested purposes [14, 16, 32, 33]. This body of research stays highly exploratory and is conducted typically on a small sample of agencies that limits their generalizability [7, 24, 33–35]. Narrowing this research gap has important implications to the study of innovative, sustainable ways through which police improve their engagement with targeted groups and the broader audiences. Expanding research on police use of social media also propels understanding of the utility (or lack thereof) of social media as a communication strategy for public or government agencies like law enforcement.
The current study seeks to examine how law enforcement agencies across the nation reacted on social media following a major legitimacy crisis. This inquiry is further situated within the context of a global pandemic that has pushed digital communication to the forefront. Specifically, we identified 115 large police departments in the U.S. with a regular presence on Twitter. We collected their tweets over a 4-month period between 4/1/2020 and 7/31/2020, covering critical periods before and after the killing of George Floyd and the nationwide protests that followed. With this, we aim to understand whether and how police agencies (both individually and as an aggregate) leveraged social media to respond to the social protests directed at the police. By combining a host of items that capture police activities on social media and public reactions to their activities, we created a single index to indicate how well police agencies engaged (or governed in a more neutral sense) the public on Twitter during the George Floyd protests.
Materials and methods
Data and sample
Using the 2016 Law Enforcement Management and Administrative Statistics (LEMAS) survey [36], we identified 139 large law enforcement agencies in the U.S. with more than 300 sworn officers and serving a population of 300,000 people and more. We focus on large agencies because they are most likely to have a regular social media presence, thus providing sufficient social media posts for analysis. The screening criteria were used in efforts to attain a meaningful and manageable sample of large U.S. police departments. We located Twitter handles of 137 (out of 139) law enforcement agencies. For police agencies with multiple official Twitter accounts, we selected only the main account with the greatest number of followers (also tweets and replies). All tweets from these agencies were fetched through Twitter’s Application Programming Interface (API) using R package rtweet on August 25th, 2020 [37]. Twitter handles of each of the 137 law enforcement agencies were used in the get_timeline function. Our data collection method complied with Twitter’s terms and conditions. We set the study period from April 1st, 2020 to July 31st, 2020 (approximately two months before and after the killing of George Floyd). This four-month study period allowed us enough data to analyze police Twitter usage before and after the killing of George Floyd and lessened the influence of the onset of the COVID-19 pandemic on police presence on social media. We excluded law enforcement agencies posting fewer than 50 tweets (including replies but excluding retweets) during this period. The final analysis sample included 115 law enforcement agencies and 38,701 tweets over a 4-month period. A complete list of these law enforcement agencies is included in S1 Table.
It is worth noting that we only examined police use of Twitter in the study due to data availability constraints. Twitter reaches between one-fifths and one-quarter of the U.S. population, and its users are younger, more likely to identify as Democrats, more highly educated and have higher incomes than U.S. adults overall [38]. Thus, the findings should be interpreted with caution due to the non-coverage of other social media platforms used by law enforcement agencies (e.g., Facebook or Nextdoor) and the demographics of Twitter users.
Data analysis
Data analysis proceeded in three main steps. First, descriptive patterns were presented to show the frequency, public reactions (i.e., favorite and retweet counts), and emotions (characterized by pre-existing sentiment lexicon and metric) expressed in the tweets of the 115 large U.S. police departments before and after the killing of George Floyd. Second, a supervised machine learning algorithm was trained to categorize each of these tweets according to a 7-category scheme. The 7 categories are: 1) civil unrest related; 2) COVID-19 related; 3) police gathering of information; 4) police communication of administrative and mundane information; 5) police communication of traffic information; 6) police communication of case updates; and 7) community engagement and outreach. The categorization scheme was constructed based on previous studies of police social media usage, consultation of leading policing scholars and practitioners, and the focus of the current study [6, 22, 39]. Detailed categorization and exemplary tweets can be seen in S2 Table. Changes in the categories or focal issues of police departments tweets before and after the killing were analyzed. Specifically, to train the multiclass classifier, a random subset of 5,000 tweets were sampled and manually labeled into one of the 7 pre-defined categories. Each author independently labeled these tweets and the intercoder reliability was about 0.70. Discrepancies were identified, discussed, and resolved (i.e., agreement on the final categorizations). The random forest classifier was evaluated with the labeled tweets (with a 75/25 split) and then applied to the entire set of tweets for classification. Additional technical details can be found in S3 Table. Third, to assess adjustments of Twitter usage made by each of the 115 law enforcement agencies before and after the killing, we analyzed and ranked changes in the frequency, public reactions, and proportions of different categories of tweets. In addition, rankings for individual items were averaged to derive an overall ranking gauging police agencies’ performance on Twitter following the George Floyd protests. Data analysis was performed using R, version 4.0.2 in 2021.
Results
Descriptive analysis of aggregate police tweets
Fig 1 shows that the 115 law enforcement agencies in our sample posted substantially more tweets in the week following the killing of George Floyd on May 25th, 2020. Yet, the number of tweets dropped to the pre-killing level after one week. Figs 2 and 3 present the average number of favorites and retweets received per tweet by the 115 law enforcement agencies before and after the killing. There was a noticeable increase in citizen reactions to police tweets immediately after the killing, and this trend lasted at least until the end of July 2020. To reduce the influence of “outliers”, tweets that received a favorite or retweet count exceeding three standard deviations above the mean were excluded. In the robustness check, substantively similar patterns were observed without excluding the outliers.
Using the Bing sentiment lexicon—a widely used general purpose English lexicon that detects the sentiment of words through a dictionary lookup and classifies words as being “positive” or “negative” [40], Fig 4 shows that police-generated tweets during the study period were more likely to include words indicating a negative emotion than words expressing a positive emotion. The negative-to-positive words ratio further increased following the killing of George Floyd. Fig 5 depicts sentence-level emotional valence (i.e., the value associated with a stimulus as expressed on a continuum from pleasant to unpleasant or from attractive to aversive) in the tweets using Rinker’s sentimentr package. The package balances accuracy (e.g., considering valence shifters) and speed in calculating text polarity sentiment in the English language at the sentence level [41]. Consistent with Fig 4, there was a decrease in the “pleasantness” or “attractiveness” expressed in the tweets over the study period. Exemplary tweets illustrating sentence-level pleasant or attractive versus unpleasant or aversive emotion can be seen in S4 Table.
Focal issues of police tweets
Table 1 reports the accuracy and kappa of the multiclass random forest classifier. Ten-fold cross-validation indicates an overall accuracy of 0.814 and a kappa value of 0.767. When applying the classifier to the split test set, the accuracy was 0.808 and the kappa value equaled to 0.758, indicating a reasonably high accuracy. By-class accuracy was also acceptable for all sub-categories in cross-validation and when applying to the split test set. Supplementary details about the results of the multiclass random forest classifier can be found in S5 Table. With the trained classifier, 37,899 tweets were classified into the 7 pre-defined categories. The number of tweets reduced from 38,701 to 37,899 because text pre-processing removed tweets that only contained hyperlinks, digits/numbers, images, videos, or stop words. Table 2 shows the number of tweets by categories. Consistent with prior research, the most frequent categories were for community engagement and outreach purpose and for case updates.
Fig 6 displays the proportions of focal issues mentioned in the tweets before and after the killing of George Floyd. There was a significant increase in the proportion of tweets related to civil unrest in the post-killing period and a significant decrease in the proportion of tweets that were COVID-19 related. More tweets were posted about case updates in the post-killing period, whereas fewer tweets were posted for community engagement and outreach purpose. Fig 7 further displays the distribution of police departments in relation to the changes in the focal issues of tweets. For instance, most police departments increased their posting of civil unrest related tweets (focal issue #1) and decreased the posting of COVID-19 related tweets (focal issue #2) in the post-killing period, whereas the distribution is more bell-shaped when looking at changes in tweets of community engagement and outreach (focal issue #7). Fig 8 shows public reactions by focal issues before and after the killing. The left panel shows that tweets related to civil unrest and community engagement and outreach received the highest average number of favorites per tweet. The pattern became more evident in the post-killing period. The right panel illustrates that tweets related to civil unrest and police gathering of information received the highest average number of retweets per tweet. Again, police audiences on Twitter were more likely to disseminate such information in the post-killing period.
Focal issues: (1) civil unrest related; (2) COVID-19 related; (3) police gathering of information; (4) police communication of administrative and mundane information; (5) police communication of traffic information; (6) police communication of case updates; and (7) community engagement and outreach.
Focal issues: (1) civil unrest related; (2) COVID-19 related; (3) police gathering of information; (4) police communication of administrative and mundane information; (5) police communication of traffic information; (6) police communication of case updates; and (7) community engagement and outreach.
Focal issues: (1) civil unrest related; (2) COVID-19 related; (3) police gathering of information; (4) police communication of administrative and mundane information; (5) police communication of traffic information; (6) police communication of case updates; and (7) community engagement and outreach.
Adjustments of individual police departments Twitter usage
Tables 3–12 illustrate how each of the 115 law enforcement agencies adjusted their Twitter usage before and after the killing of George Floyd. To adjust for baseline levels of tweeting practices across police departments, the 115 law enforcement agencies were divided into two groups. Across the 115 law enforcement agencies, the mean was 155 and the median was 114 tweets in the pre-killing period. We made the cut-point at 110 tweets in the pre-killing period to create the two groups. The first group included the more active agencies, namely, the 60 agencies which posted, on average, at least 2 tweets per day during the pre-killing period (i.e., the higher-use group). The second group included the other 55 agencies which were less active on Twitter during the pre-killing period (i.e., the lower-use group). By dividing the agencies into the higher-use and lower-use groups, we balanced the raw and percentage changes when ranking the agencies and partially adjusted for potential influences of agency/personnel size and jurisdiction population (i.e., agency-level factors) on police social media usage.
Tables 3 and 4 rank agencies in the higher- and lower-use group based on how substantially they increased or decreased their tweeting practices pre- and post-killing. For instance, the Aurora Police Department (ranked #1 in the higher-use group for this dimension) posted 152 tweets in the pre-killing period and 477 tweets in the post-killing period. The average number of posts per day was 2.76 (152 tweets/55 days) and 7.12 (477 tweets/67 days) pre- and post-killing. This translates to a raw frequency change of 4.36 (7.12–2.76) and a percentage change of 158% (7.12/2.76–1). Tables 5 and 6 rank the agencies based on the increase or decrease in the average number of favorites they received per tweet pre- and post-killing. For instance, tweets from the Venture County Sheriff’s Office (ranked #1 in the lower-use group for this dimension), on average, received 11.5 favorites in the pre-killing period and 139 favorites in the post-killing period, a raw favorite change of 127.5 (139–11.5) and a percentage change of 1117% (139/11.5–1). In a similar vein, Tables 7 and 8 rank the agencies based on the increase or decrease in the average number of retweets they received per tweet pre- and post-killing.
Moreover, Tables 9 and 10 rank the agencies based on the increase or decrease in posting civil unrest related tweets pre- and post-killing. For instance, the Portland Police Department (ranked #1 in the higher-use group for this dimension) posted a total of 303 tweets in the pre-killing period and none of the tweets were civil unrest related. Yet, in the post-killing period, they posted a total of 900 tweets, 535 of which were civil unrest related. Thus, the proportion change was 0.594 (535/900–0/303). Tables 11 and 12 rank the agencies based on the increase or decrease in posting community engagement and outreach tweets pre- and post-killing. For example, the NYPD (ranked #1 in the higher-use group for this dimension) posted a total of 663 tweets in the pre-killing period and 179 of the tweets were for community engagement and outreach purpose; in the post-killing period, they posted a total of 428 tweets, 205 of which were for community engagement and outreach purpose. Thus, the proportion change was 0.209 (205/428–179/663).
Finally, we combined the five dimensions above (Tables 3 through 12) and constructed an overall ranking to gauge which law enforcement agencies may have more effectively reached and engaged (or governed) citizens through Twitter. To offer a straightforward understanding, the five dimensions were assumed equal weights in our attempt and their corresponding ranks were averaged to derive an overall ranking. Tables 13 and 14 illustrate the overall rankings for the higher- and lower-use group. For example, the Charlotte-Mecklenburg Police Department (in the higher-use group) ranked 5th in posting more tweets per day, 3rd in receiving more favorites per tweet, 4th in receiving more retweets per tweet, 5th in posting a higher percentage of civil unrest related tweets, and 30th in posting a higher percentage of community engagement and outreach tweets before vs. after the killing. The mean equaled 9.4 ((5+3+4+5+30)/5) across the five rankings and placed the Charlotte-Mecklenburg Police Department 1st in the overall rank. It is worth noting that none of the police departments ranked (very) high on all five dimensions. In particular, if they increased their posting of civil unrest related tweets in the post-killing period, they were likely to reduce posting community engagement and outreach tweets.
Discussion
The current study explored police tweeting practices in a sample of 115 large agencies in the U.S., approximately two months before and after the killing of George Floyd that sparked the nationwide protests directed at the police in 2020.
In line with image repair theory, our analyses provided insights into the specific activities police agencies engaged in on social media in response to image damage and public reactions to those activities. Specifically, law enforcement agencies tweeted more frequently in the immediate aftermath of the killing and posted an increased number of civil-unrest related tweets. Police also continued to communicate case updates, perhaps directing public attention to their traditional role and responsibilities in fighting crime and maintaining order. On the other end, the public (at least those who were exposed to police departments tweets) showed a greater interest in engaging with law enforcement agencies. The rate at which the public favorited or retweeted a police tweet went up significantly following the George Floyd incident and stayed higher than before throughout the rest of the study period. Changes in the focal issues of police tweets (and potentially an increased attention to police behavior) may partially explain the increases in favorites and retweets received per tweet despite the police being in a legitimacy crisis. In particular, police tweets related to civil unrest, on average, received public reactions between twice and 20 times more than those of other categories of tweets received. By channeling and amplifying public energy towards this issue, social media provides opportunities for law enforcement to respond, engage, and rectify any misinformation with high efficiencies.
It may not be surprising that agencies that had the largest increases in public reactions (i.e., average favorite and retweet counts) or protest-related posts were from cities that saw major protest and riot activities, which aligns with the image repair thesis. However, we cannot conclude whether these increases were results of police engagement efforts aiming to genuinely improve police-community relations or socialization/legitimation efforts of reputation management. For instance, category 1 (or civil unrest related) tweets covered such topics as operational responses to the protest, crime and violence committed during the riots, challenges to racial justice in policing, and injuries and hostilities to police. These subcategories tap genuine concerns about racial injustice in the U.S. but also governance of citizens. We aimed to further distinguish subcategories within a focal issue (e.g., our labeled data included such information). Yet, the very low frequency of some subcategories in the labeled data and the limits of our prediction models prohibited us from pursuing this route. Nonetheless, it is clear that individual police agencies varied vastly in their social media usage. Not every agency was actively using Twitter to reach and engage or govern people. Twenty-four of the initial 139 large police agencies identified did not have a regular presence on Twitter. The final sample of 115 large police agencies also demonstrated tremendous variability in how often they tweeted, the types of tweets they tended to post, and public reactions to their social media content both at the baseline level and during the protest. Examining only a handful of agencies or a particular dimension of social media usage seems unlikely to provide a complete picture of police behaviors and citizen interactions on Twitter.
We contributed to the literature of police use of social media by creating a single indicator that combined measures of changes in the quantity of tweets, composition of tweets, and public responses to those tweets. While prior studies on police use of social media have looked at the number and types of police posts and the diffusion of those posts individually, few engaged in efforts to show where law enforcement agencies are relative to each other with respect to different sub-metrics and the overall standing of Twitter use. We controlled for baseline social media activity levels by separating our sample into higher versus lower activity agencies, thus adjusting for potential influences of agency-level factors on social media usage and its changes (e.g., agency size and jurisdiction population). Our combined indicator may serve as a useful first step toward cultivating responsible and effective use of social media by police. That said, the current study represents a preliminary effort at quantifying and understanding police social media usage in the context of the George Floyd protests and does not represent a comprehensive measurement of police performance on social media overall.
Given the great variability in police social media usage observed in our study and different possible interpretations of these efforts (e.g., engagement vs. socialization/legitimation), we do not recommend the deactivation of all police Twitter accounts as suggested by some [13]. Instead, we suggest a few avenues for future research (and practices) on responsible and effective use of social media by police, while pointing out the challenges associated with such inquiries.
First, future studies should explore why (and how) changes in police social media usage occur before and after a major social event. A few important challenges remain. We are uncertain of who are interacting with police on social media, which has important implications to its impact on police legitimacy or police-community relations. Much of the legitimacy crisis reflects ongoing frictions between police and disadvantaged/minority communities, the dynamic of which may not be captured by examining the overall responses received by police tweets. For instance, rather than reflecting improvements in community engagement activities or citizen trust, the increases in favorites and retweets of police-generated content might reflect more active reactions from a pre-existing pro-police audience. Police engagement efforts, however, are most needed towards minority and disadvantaged groups who are regularly contacted by law enforcement agencies. Whether and through what strategies police social media usage can target, reach, and respond to those groups would largely determine the efficacy of online police-community interactions, especially in repairing harm and (re)gaining trust. Otherwise, police communications on Twitter may not be fundamentally different from traditional means of communication and mainly fulfill a function of socialization/legitimation or appeal to those who already endorse police value and activities.
In addition, it is necessary to further investigate the detailed content generated by police on social media. While categorization, as in our case, is helpful in understanding shifts in general directions of police social media usage, topic modeling in natural language processing may uncover themes from a large corpus of tweets and assign individual tweets to different themes, better illustrating police motives for social media usage [42, 43]. Adopting computer-assisted techniques to analyze (at a large scale) hyperlinks, images, and videos contained in police tweets should further improve our understanding of police social media usage [44]. Moreover, it would be helpful to investigate what organizational characteristics are associated with agency-level police motives for social media usage and adjustments after a major challenge.
Second, future studies should assess the impact of police social media usage on other performance measures, including public receptivity, police legitimacy and trust, crime investigation and clearance rate, community informal social control, among others. Such undertaking is challenging given the nature of these inquiries and the data needed for answering these questions yet important. Citizens’ experiences with the police affect their overall assessment of the police, but the vast majority of the American public do not have face-to-face contact with a police officer in any given year [45, 46]. The extension from physical interaction with the police to social media platforms is worth further investigation.
Of note, although some consider liking and reposting behaviors less engaging or dialogical than “real” engagement activities such as community meetings, police-community collaborations, and joint problem-solving efforts, metric-driven engagement has an important meaning in and of itself in an algorithmic environment of social media in which information is curated and disseminated based on their relative popularity. Recognizing the limitations (e.g., ambiguous motives of police social media usage and messages not necessarily reaching targeted groups), scholars have argued that these metrics should be used to guide the development of social media strategies of law enforcement agencies [47], similar to how favorites and retweets are commonly used as indicators of success of a marketing strategy in the private industry. That said, agencies should be cautious not to seek reactions by posting content simply to appeal to their audiences. Authenticity and communicating negative but honest messages have been found to be key to maintaining police credibility on social media [15]. This helps explain the findings from our sentiment analysis. Police-generated social media content exuded greater negative than positive emotions following the George Floyd incident, but public reactions (i.e., average favorite and retweet counts per tweet) also went up tremendously during this time.
The study has limitations. First, police agencies may use social media platforms other than Twitter (e.g., Facebook or Nextdoor) to reach and engage (or govern) people during the same study period. Second, we did not explicitly investigate two-way police-citizen interactions on Twitter. Official police agency Twitter accounts often replied to other non-public Twitter accounts (e.g., a police chief’s account or police precinct account). Given the scale of the current study, manually checking each replying tweet was not feasible. Thus, we could not accurately assess the proportion of two-way police-citizen interactions on Twitter. Our preliminary check (excluding self-replying tweets) indicated that approximately 13% of all included 38,701 tweets were replies and that there were great variabilities in the proportion (e.g., over half of the Denver Police Departments tweets during the study period were replies, whereas several police agencies did not post any replying tweets during the same period of time) and the way official police agency Twitter accounts posted replying tweets. Additionally, we only analyzed the text content of police tweets, yet image or video content (also URLs) may meaningfully affect public reactions to police tweets. Moreover, retweeting does not necessarily reflect agreement with original content (e.g., retweets with users’ own negative reactions). In this sense, our findings show increased public participation in dialogues on public safety and social justice issues, not necessarily increased support for police-generated content online. Third, our classification algorithm was not perfect, but its accuracy was acceptable for our purpose. Fourth, the study examined police departments tweets approximately two months before and after the killing. Adjustments of Twitter usage made by police agencies may take longer to carry out. The scale and intensity of protests (and disruptions) at different jurisdictions may also affect how local police agencies adjust their social media presence, which we could not explicitly study. Future research should also explore geographic and political influences on police use of social media. Finally, given our focus on large police departments in the U.S., the results may not be generalizable to smaller agencies or agencies in other countries in their use of social media during a social crisis event.
Conclusion
The utility of social media in policing and public governance remains an understudied area, where case studies and qualitative evidence predominate. Through examining Twitter usage by 115 large U.S. police agencies following a major legitimacy crisis, we conclude that police reacted to the George Floyd incident on social media and that the public paid attention to and seemingly held positive attitudes toward those changes. Police agencies in our sample tweeted more frequently following the killing of George Floyd and posted more tweets related to civil unrest as well as case updates. These tweets received greater public reaction (through favorites and retweets), which persisted throughout the study period.
Nonetheless, a great variability emerged across agencies in their responses on social media (e.g., different rates and focuses of use), and the motives for the observed changes pre- and post-event were inconclusive. Future efforts are called for to address the limitations and ambiguities uncovered by this study about police use of social media (e.g., characteristics of those who interact with police on social media, communications that go beyond favorites and retweets, and police behaviors on social media platforms other than Twitter), and to find ways for police to responsibly and effectively utilize various communication platforms in the era of “big data”. For instance, a guideline or protocol of best practices for police social media usage may be developed and made public for comments prior to its approval and implementation, through which “selective transparency” may be curbed.
Supporting information
S1 Table. A complete list of the 115 law enforcement agencies included in the study.
https://doi.org/10.1371/journal.pone.0269288.s001
(DOCX)
S2 Table. Detailed categorization scheme used in the study.
https://doi.org/10.1371/journal.pone.0269288.s002
(DOCX)
S3 Table. Random forest and multiclass boosted trees classifier.
https://doi.org/10.1371/journal.pone.0269288.s003
(DOCX)
S4 Table. Exemplary tweets illustrating sentence-level pleasant or attractive vs. unpleasant or aversive emotion.
https://doi.org/10.1371/journal.pone.0269288.s004
(DOCX)
S5 Table. Supplementary details about the results of the multiclass random forest classifier.
https://doi.org/10.1371/journal.pone.0269288.s005
(DOCX)
Acknowledgments
The authors would like to thank Cynthia Lum, Laurie Robinson, James Willis, Charlotte Gill, and David Weisburd for their helpful comments on the earlier draft of the manuscript. The opinions expressed in this article are those of the authors and do not reflect the positions of any of the organizations with which the authors are affiliated.
References
- 1. Hinton E. Policing social unrest and collective violence. Science. 2021;374(6565):272–4. pmid:34648335
- 2.
National Academies of Sciences, Engineering, and Medicine. Proactive policing: effects on crime and communities. National Academies Press; 2018.
- 3. Beard JH, Jacoby SF, Maher Z, Dong B, Kaufman EJ, Goldberg AJ, et al. Changes in shooting incidence in Philadelphia, Pennsylvania, between March and November 2020. JAMA. 2021;325(13):1327–8. pmid:33566087
- 4. Rosenfeld R, Abt T, Lopez E. Pandemic, social unrest, and crime in US cities. Council on Criminal Justice; 2020.
- 5.
Crime Data Explorer. Federal Bureau of Investigation. https://crime-data-explorer.app.cloud.gov/pages/explorer/crime/crime-trend. Trend of homicide from 2010 to 2020. Accessed October 23, 2021.
- 6.
Hu X, Lovrich NP. Electronic community-oriented policing: theories, contemporary efforts, and future directions. Lexington Books; 2020.
- 7. Beshears ML. Effectiveness of police social media use. American Journal of Criminal Justice. 2017;42(3):489–501.
- 8. Crump J. What are the police doing on Twitter? Social media, the police and the public. Policy & Internet. 2011;3(4):1–27.
- 9. Meijer A, Thaens M. Social media strategies: understanding the differences between North American police departments. Government information quarterly. 2013;30(4):343–50.
- 10. Walsh JP, O’Connor C. Social media and policing: a review of recent research. Sociology Compass. 2019;13:e12648.
- 11. O’Connor CD. The police on Twitter: image management, community building, and implications for policing in Canada. Policing and Society. 2017;27(8):899–912.
- 12.
International Association of Chiefs of Police (IACP). Social Media. https://www.theiacp.org/sites/default/files/2019-05/Social%20Media%20Paper%20-%202019.pdf Accessed March 6, 2021.
- 13. Cheng T. Social media, socialization, and pursuing legitimation of police violence. Criminology. 2021;59(3):391–418.
- 14. Kudla D, Parnaby P. To serve and to tweet: an examination of police-related Twitter activity in Toronto. Social Media + Society. 2018;4(3):2056305118787520.
- 15. Bullock K. (Re) presenting ‘order’ online: the construction of police presentational strategies on social media. Policing and Society. 2018;28(3):345–59.
- 16. Brainard LA, McNutt JG. Virtual government–citizen relations: informational, transactional, or collaborative? Administration & Society. 2010;42(7):836–58.
- 17.
Kim K, Oglesby-Neal A, Mohr E. 2016 Law enforcement use of social media survey. International Association of Chiefs of Police & the Urban Institute. 2017.
- 18. De R, Pandey N, Pal A. Impact of digital surge during Covid-19 pandemic: a viewpoint on research and practice. International Journal of Information Management. 2020;55: 102171. pmid:32836633
- 19.
Lum C, Maupin C, Stoltz M. The impact of COVID-19 on law enforcement agencies (Wave 2). International Association of Chiefs of Police & the Center for Evidence-Based Crime Policy, George Mason University. Washington, DC. 2020. https://www.theiacp.org/sites/default/files/IACP_Covid_Impact_Wave2.pdf. Accessed December 6, 2021.
- 20. Tat‐Kei Ho A. Reinventing local governments and the e‐government initiative. Public Administration Review. 2002;62(4):434–44.
- 21. Kumar N, Vragov R. Active citizen participation using ICT tools. Communications of the ACM. 2009;52(1):118–21.
- 22. Dai M, He W, Tian X, Giraldi A, Gu F. Working with communities on social media: varieties in the use of Facebook and Twitter by local police. Online Information Review. 2017;41(6):782–96.
- 23. Mawby RC. Police corporate communications, crime reporting and the shaping of policing news. Policing and Society. 2010;20(1):124–39.
- 24. Mayes L. Social media and community-oriented policing: examining the organizational image construction of municipal police on Twitter and Facebook. Police Practice and Research. 2021;22(1):903–20.
- 25. Brayne S. Big data surveillance: the case of policing. American Sociological Review. 2017;82(5):977–1008.
- 26. Martinez-Rojas M, del Carmen Pardo-Ferreira M, Rubio-Romero JC. Twitter as a tool for the management and analysis of emergency situations: a systematic literature review. International Journal of Information Management. 2018;43:196–208.
- 27.
Denef S, Bayerl PS, Kaptein NA. Social media and the police: tweeting practices of British police forces during the August 2011 riots. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems; 2013.
- 28. Davis EF, Alves AA, Sklansky DA. Social media and police leadership: lessons from Boston. Australasian Policing. 2014;6(1):10–16.
- 29. Liu BF, Fraustino JD. Beyond image repair: suggestions for crisis communication theory development. Public Relations Review. 2014;40(3):543–6.
- 30. Benoit WL. Image repair discourse and crisis communication. Public Relations Review. 1997;23(2):177–86.
- 31. Bertot JC, Jaeger PT, Grimes JM. Using ICTs to create a culture of transparency: E-government and social media as openness and anti-corruption tools for societies. Government Information Quarterly. 2010;27:264–71.
- 32. Lieberman JD, Koetzle D, Sakiyama M. Police departments’ use of Facebook: patterns and policy issues. Police Quarterly. 2013;16(4):438–62.
- 33. Procter R, Crump J, Karstedt S, Voss A, Cantijoch M. Reading the riots: what were the police doing on Twitter? Policing and Society. 2013;23(4):413–36.
- 34. Harms J, Wade L. Reconnecting with the public: police use of social media. Journal of Law Enforcement. 2017;6(1):1–10.
- 35. Kelly A. Managing the risks of public discourse on the New South Wales Police Force Facebook site. Salus Journal. 2014;2(1):19–42.
- 36.
Bureau of Justice Statistics, U.S. Department of Justice. Law Enforcement Management and Administrative Statistics (LEMAS), 2016. https://bjs.ojp.gov/data-collection/law-enforcement-management-and-administrative-statistics-lemas. Accessed October 2, 2021.
- 37.
Kearney MW, Heiss, A, Briatte F. Package ‘rtweet’. https://cran.r-project.org/web/packages/rtweet/rtweet.pdf. Accessed October 23, 2021.
- 38.
Wojcik S, Hughes A. Sizing up Twitter users. Pew Research Center; 2019.
- 39. Hu X, Dong B, Lovrich N. “We are all in this together:” police use of social media during the COVID-19 pandemic. Policing: An International Journal. 2021;online first.
- 40.
Liu B. Sentiment Analysis: mining sentiments, opinions, and emotions (2nd edition); Cambridge University Press; 2020.
- 41.
Rinker T. Package `sentimentr’. https://cran.r-project.org/web/packages/sentimentr/sentimentr.pdf. Accessed October 23, 2021.
- 42.
Yan X, Guo J, Lan Y, Cheng X. A biterm topic model for short texts. Proceedings of the 22nd International Conference on World Wide Web; 2013.
- 43. Qiang J, Qian Z, Li Y, Yuan Y, Wu X. Short text topic modeling techniques, applications, and performance: a survey. IEEE Transactions on Knowledge and Data Engineering; 2020.
- 44. Imran M, Ofli F, Caragea D, Torralba A. Using AI and social media multimodal content for disaster response and management: Opportunities, challenges, and future directions. Information Processing & Management. 2020;57(5):102261.
- 45.
Gallagher C, Maguire ER, Mastrofski SD, Reisig MD. The public image of police: final report to the IACP by the George Mason University Administration of Justice Program. International Association of Chiefs of Police; 2001.
- 46.
Harrell E, Davis E. Contacts between police and the public, 2018–statistical tables. Bureau of Justice Statics Report, NCJ. 2020;255730.
- 47. Wood MA. Policing’s ‘meme strategy’: understanding the rise of police social media engagement work. Current Issues in Criminal Justice. 2020;32(1):40–58.