2.1. Corporate fake news

We build a comprehensive database of news about U.S. firms that has been debunked by prominent fact-checking organizations (Snopes.com, Factcheck.org, Politifact.com, and Truthorfiction.com). Each fact-checking site has its own classification scheme (e.g., Snopes.com classifies articles into six categories, whereas Politifact.com has nine categories). We normalize the verdicts across different sites by mapping these classes to fake and non-fake news categories (see S2 Appendix in S1 File for details), and we do not include mixed news in the analysis (i.e., news classified as neither true nor false). In doing so, we aim to identify a broad spectrum of corporate fake news over a long time period. Whereas previous studies focus only on the financial information of public companies [11, 36], our approach allows us to identify a comprehensive sample of fake news that targets businesses but is not necessarily financial in nature.

First, we automatically scrape the fact-checking websites to collect all the fact-checked articles and parse the publication date, title, claim, body of the text, and fact-checking ratings for each piece of news. Because a naïve company name keyword search might not fully capture a firm’s products, executives, or subsidiaries (e.g., news about Oreo Cookies may be linked to Mondelez International), we use Named Entity Recognition (NER) (i.e., NLTK, TextBlob, and SpaCy) methods to create a filtered sample of news potentially related to firms. NER methods help us identify an initial subset of news referring to a company (or product) such as KFC, Facebook, or Pfizer or a person such Bill Gates or Steve Jobs. However, because the automated algorithms are trained on non-business textual data, they may generate false positives [37]. Therefore, we manually read the filtered subsample to exclude any non-firm-related news. After reading 100 randomly selected articles, we develop the labeling rules presented in S3 Appendix in S1 File. For example, if the news is about a CEO’s arrest, we include the news in our sample because of its importance to the firm’s operations (see, e.g., https://www.snopes.com/fact-check/italy-bill-gates-arrest/). However, if the news is about a CEO’s private life (or her charitable activities), we do not include it in our sample (see, e.g., https://www.snopes.com/fact-check/bill-gates-planned-parenthood/ or https://www.politifact.com/factchecks/2020/may/14/facebook-posts/no-evidence-gates-foundation-will-profit-coronavir/). We also exclude conspiracy theories, satire, and false statements by politicians [7] and keep only the negative-sentiment news, using the sentiment score determined by the word list in [37] (e.g., we classify news as negative if it contains more negative than positive words and manually read the news to determine the sentiment if the difference between positive and negative words does not exceed 1% of total words).

Second, we manually classify news content to identify the specific topics in the text. We prefer manual annotation (instead of automated annotation algorithms) to develop a deeper understanding of the features of the text given the high level of domain expertise required in our setting. To do this, we start by determining categories based on 50 randomly selected subsets of news. Then, we train three research assistants to annotate the same text independently and compare results. After ensuring the degree of consensus regarding the topics, research assistants continue to independently annotate a larger set of news and identify major categories of news topics. When the research assistants disagree (on a new topic category), the authors discuss the news before reaching an agreement. This way, we identify a broad range of topics including firms’ products and services, operations, data privacy, politics, and founders and executive management. Table 1 shows several examples of corporate fake news classified into the various topical categories.

Finally, we identify the origin of a story by relying primarily on URL links mentioned in fact-checking articles. Fact-checking articles often discuss and cite the source of a claim while debunking the claim. We manually read each URL link in fact-checked articles to determine the origin of the claim and identify its publication date. If there is more than one relevant source, we keep the earliest published source.

2.2. Twitter data

We collect historical tweet- and user-level data about corporate news using Twitter Academic API. The API grants academics full access to historical tweets dating back to 2006 (except for deleted tweets and accounts). We identify the original (or source) tweets spreading fake and non-fake corporate news in two steps. First, we collect all the tweets that contain a link to a fact-checking website that evaluates the veracity of a corporate news story. We exclude the original tweets containing a link to a fact-checking website, because our goal is to identify the spread of unverified and contested information, not information verified by fact-checking organizations. The remaining tweets are replies to an original tweet or replies to replies. We also remove tweets that do not directly reply to an original tweet to ensure that a reply containing a link to a fact-checking website is in fact addressing the original tweet.

Second, we extract the URL links to external articles mentioned in the above original tweets (spreading fake and non-fake news) or the URL links mentioned in the fact-checking articles. After manually reading the extracted articles, we identify the URL links about the (fake and non-fake) news. We then extract the original tweets containing a link to any of these articles (which are not necessarily fact-checked through replies). In our search, we transform links to canonical URLs by removing http://, https:// and analytic tracking parameters (i.e., Urchin Tracking Module parameters), and we transform short URLs to the expanded form (to merge different links referring to the same article). Through this process, we identify a sample of original tweets mentioning a fact-checked corporate news story on Twitter. We define the sender of the original tweet (i.e., the source of the Twitter cascade) as the rumor source.

These sample filters leave us with 342,818 original tweets (i.e., no retweets or replies) mentioning corporate news verified by a fact-checking organization. Tweet-level variables include the tweet text, timestamp, tweet language, and number of retweets. We also separately query for author-level data—user name, user description, the number of followers, the number of followees, and account creation date—for each unique Twitter user authoring one of the collected tweets. We obtain data for 189,158 unique Twitter users. Finally, we manually match Twitter data to other academic datasets (e.g., COMPUSTAT) using company names. Our code and dataset are available at https://github.com/alperdarendeli/corporatefakenews. The collection and analysis methods comply with the terms and conditions for the source of the data. Our study was also reviewed and approved by the Institutional Review Board of Nanyang Technological University (IRB-2022-349).

2.3. Twitter location-prediction model

The geolocation of a rumor source on social media is hard to pin down, either because users do not disclose their home location or do not enter data that correspond to their actual locations [14]. For example, users frequently enter fake locations or sarcastic comments in their profiles (e.g., Jupiter, Outta Space, Out of this world, etc.), making it difficult to infer user location solely from self-declared profile data. The profile locations could also be inaccurate because individuals choose not to publicly share their country of location or intentionally obfuscate the origin of their tweets, as is frequently the case with user-generated input. Hence, location information on Twitter is often far from complete and reliable. To tackle this issue, we employ an LTSM model for location prediction, which is particularly suitable for processing reasonably long and sequential text data [15].

The LSTM model is a type of RNN that can utilize information further back in a sequential chain. To do this, LSTM models use mechanisms called gates that regulate the flow of information being passed from one step to the next. Unlike a vanilla RNN, LSTM includes a forget gate that determines the sort of information that is passed across the sequence. The operations within an LSTM allow the model to keep relevant information (and forget irrelevant information) from the previous steps no matter what the length of the sequence is. Thus, LSTM models do not suffer from the vanishing gradient problem (i.e., the short-term memory problem). This makes LSTM suitable for processing sequential data with long-term dependencies (such as text). We use the model to predict the geographic location of Twitter users [16].

For prediction, we use tweet content in combination with user metadata because the recent literature shows that the metadata can contribute substantially to predictive accuracy and provide valuable location signals [15]. We use a large sample of English and non-English geolocated tweets between 2014 and 2019 as our labeled dataset. We use the locations in geo-tagged tweets as our ground-truth because they are based on the GPS coordinates of mobile devices, which are reliable and difficult to manipulate. The training data consist of 3,927,563 geotagged tweets covering 149 countries and 2,187 cities (see S1 Appendix in S1 File for details).

First, we extract the following text features from the geotagged tweets: tweet text, tweet language, user-declared location, user description, and user name. Other features such as time zone, UTC offset, URL links, and messenger source (e.g., iPhone or Android) can also help predict locations. Time zone and UTC offset, however, cannot be extracted from Twitter API at the time of our study (https://twittercommunity.com/t/upcoming-changes-to-the-developer-platform/104603). The previous literature also shows the limited benefit of URL links and messenger source for the prediction task [16]. We then clean the text by (i) removing links, user name, punctuation, and extra spaces, (ii) separating emoticons, (iii) making all text lower case, and (iv) concatenating the tweet features. We concatenate as follows, inserting special tokens at the front of each text field. A [BLANK] token is also inserted if the specific text field is blank.

‘[TEXT] <cleaned_text> [LANG] <tweet_lang> [LOC] <cleaned_user_declared_location> [DESC] <cleaned_user_description> [NAME] <cleaned_user_name>‘

We then feed the concatenated text into an LSTM model to predict geographic locations. By and large, we follow the model architecture in [16] and pose the geolocation-prediction task as a multi-label classification task of 2,187 cities. In the model, the concatenated text is represented as word embeddings because it better captures words with similar locational semantics in low-dimensional vectors (and produces a more efficient representation of words than one-hot encoding). After transforming the features into machine-readable vectors, we train an LSTM model with a prediction layer. We split 80% of the data into training, 10% into validation, and 10% into test sets. We use stratified sampling to ensure that the geographical distribution of each set is approximately equal. The model learns a weight (or coefficient) for each tweet feature and uses Adam optimization to minimize cross-entropy loss over all possible weights. We fine-tune the model using different hyperparameters on the validation set. Table 2, Panel A reports the selected model hyperparameters. S1 Appendix in S1 File provides technical details about data pre-processing, parameter tuning, and model training steps. The model predicts the geographic location of a tweet at the city level (after which we map the prediction to the related countries). We do not use city-level predictions because the predictive accuracy at the city level is much lower than the accuracy at the country level (51.4% versus 88.8%). And, more importantly, our research question, which examines the prevalence of fake news from foreign accounts, is cast at the country level.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Twitter location-prediction model.

https://doi.org/10.1371/journal.pone.0301364.t002

We then evaluate the accuracy of the model at the country level using the test set.

The individual performance of different features is presented in Table 2, Panel B. If we had always predicted that the United States would be the most frequently observed country in the training data, our prediction would be 21% correct. As a simple baseline, we assess our model’s performance against this majority predictor. We start our bottom-up analysis with individual tweet features. We find that tweet text is the most relevant feature for location prediction. Using the text alone, we can correctly predict the location of 72.48% of all tweets. Consistent with the prior literature, we find that augmenting text with user metadata improves accuracy. For example, combining text with user-declared location improves accuracy to 86.34%. When we concatenate all features (tweet text, user-declared location, user description, and user name), the model achieves the best predictive accuracy of 88.76%.

Next, we perform a leave-one-out feature importance analysis to rank the factors in terms of their contribution to model performance. The analysis relies on the idea that if a feature is not important, excluding it from the predictor set should not noticeably decrease the model’s out-of-sample performance. We iteratively remove one feature at a time and evaluate accuracy. Specifically, we calculate the decrease in the predictive accuracy when a feature is excluded from the predictor set. We then scale the decrease by the predictive accuracy when all predictors are used. Surprisingly, while tweet text separately has the largest influence on the prediction of user locations, we find, in Fig 1, that user-declared location is the top explanatory predictor in the feature importance analysis. This implies that user-declared location may not carry a large weight on its own, but combining it with other features might help improve the predictive accuracy of the model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Feature importance analysis.

This figure ranks the features of the location-prediction model in terms of their contribution to model performance. For the feature importance analysis, we use the following equation: The numerator of the equation calculates the decrease in predictive accuracy when a feature is excluded from the predictor set. We then scale the decrease in predictive accuracy by the predictive accuracy of the model when all features are used. The reported numbers are in percentage points.

https://doi.org/10.1371/journal.pone.0301364.g001

Overall, our model significantly outperforms a naïve majority predictor in predicting the geographic locations of tweets. Our model accuracy is also comparable to that of other location-prediction models in the literature. For example, a maximum entropy classifier [38] predicts the country origins of tweets with 88.90% accuracy using tweet language, user-declared location, user language, time zone, offset, user name, user description, and tweet text. A linear classifier model [39] predicts the country origins of tweets with 87.34% accuracy using tweet text, profile location, time zone, and time (in UTC time). Our model’s predictive accuracy is comparable despite our inability to use additional features such as time zone, offset, and user language, which were not accessible via Twitter API at the time of this study.

As the authors of [38] show that geo-located and non-geolocated tweets have similar characteristics, we can use a model trained on geolocated tweets to predict the locations of users without geo-tagged data. Therefore, we use our fine-tuned model to predict the locations of tweets disseminating fact-checked corporate news. Fig 2 illustrates the location-prediction procedure at the user level. We use metadata and multiple tweets of each user to predict the user’s country location.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Twitter user location-prediction model.

This figure illustrates the location prediction of a Twitter user. We implement a two-step procedure. First, we use our model to predict locations at the tweet level. Specifically, we apply our fine-tuned model’s weights to produce a score (or multiclass logit) for each tweet location. Second, we take the average probabilities of tweet location at the user level to predict the country with the highest score as the location of the Twitter user.

https://doi.org/10.1371/journal.pone.0301364.g002

We apply the fine-tuned model’s weights to our sample to produce a score for each tweet’s location. Specifically, we use the softmax function to convert these weights to multiclass logits (or probabilities) for each city and country. For each tweet, the model generates a vector of probabilities Pr(Cu_k) = {pr(C_1uk), pr(C_2uk),…, pr(C_Luk)}, where pr(C_1uk) is the probability that the user u_k will be assigned to country 1, and so on. Then, we take the average probabilities of tweet locations at the user level and predict that the Twitter user is located in the country with the highest score.

3.1 Empirical construct

We use the model’s predictions to construct Foreign Corporate Fake News (%) as the key empirical construct of our study. Our model predicts the geographical locations of Twitter users across 139 countries. We exclude, however, the countries where Twitter is blocked (i.e., China, North Korea, Russia, Iran, Uzbekistan, Turkmenistan, and Belarus), which account for 0.62% of the users in our sample. Our findings are robust to including these countries in the analysis.

We categorize a corporate fake news story as foreign originated if its rumor source (i.e., the Twitter user who initiates the cascade) is outside the U.S. To put it differently, we define a continuous variable—Foreign Corporate Fake News (%)—as the percentage of original tweets spreading fake news initiated by an account in a foreign country. For example, if we have 10 original tweets spreading a fake news story, and five of the tweets are initiated by foreign accounts, the value of Foreign Corporate Fake News (%) is 50%. The measure accounts for the fact that fake news can be spread via multiple cascades on Twitter. In robustness tests, we replace Foreign Corporate Fake News (%) with Foreign Corporate Fake News (dummy), as binning the data can reduce measurement errors caused by a noisy continuous variable [40]. We define Foreign Corporate Fake News (dummy) as an indicator variable equal to one if at least one Twitter account spreading fake news is located in a foreign country, and zero otherwise. We also use corporate non-fake news to construct a corresponding benchmark. We define Foreign Corporate Non-fake News (dummy) as an indicator variable equal to one if at least one Twitter account spreading non-fake news is located in a foreign country, and zero otherwise. Similarly, Foreign Corporate Non-fake News (%) is the percentage of original tweets spreading non-fake news initiated by an account outside the U.S. We construct the measures at the news level. For much of the analysis, however, we aggregate the measures to the country-year (Section 3.3) and firm-year (Section 3.4) level.

3.2. Sampling design and summary statistics

We begin our analysis by reporting descriptive statistics. First, we report data across fake and non-fake corporate news. We identify 541 fake news stories (about 126 unique firms) and 144 non-fake news stories (about 67 unique firms) over the period from January 2012 to June 2021. Within each news category, we tabulate news content and provide statistics about the characteristics of rumor starters. Table 3, Panel A presents the results. Most of the fake news revolves around the firm’s politics, product, or operations. 37.7% of corporate fake news involves political discussions (e.g., support for social movements such as Black Lives Matter, funding of political parties, government contracts, etc.), whereas product-related news (e.g., contaminated products, product discontinuation, product safety, etc.) accounts for 22.6%, and operations-related news (e.g., downsizing, employee policy, etc.) accounts for 16.8% of the fake news. The distribution of topics for non-fake news follows a similar pattern, and the difference across news categories is not statistically significant.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Characteristics of corporate fake news.

https://doi.org/10.1371/journal.pone.0301364.t003

In Panel B, we identify the individual characteristics of the rumor starters. We use Twitter metadata to extract user profiles. Specifically, we check the number of people who follow the user on Twitter (i.e., followers), the number of people whom the user follows in Twitter (i.e., followees), the age of the user’s account (measured in years), and the percentage of bot accounts. We find that starters of fake news (i.e., Twitter users who initiate the news) have more followers than starters of non-fake news. They are also relatively new accounts. We also use Botometer, a popular public tool, for bot detection on Twitter (see S4 Appendix in S1 File for details). Using this tool, we calculate a Bot Activity score between zero and one for each Twitter user. A higher score indicates a higher likelihood that a Twitter account is a bot. We find that bot score is higher for fake news accounts (0.51 versus 0.46), suggesting that fake news spreaders are more likely to exhibit bot-like behavior. We also find that bot-like activity is significantly higher for fake news than for non-fake news, which indicates that a large group of automatically operated accounts may be promoting the fake news (t = 36.48, p<0.00). Finally, we find that corporate fake news, on average, is retweeted around 480 times, whereas non-fake news is retweeted around 20 times. This finding is consistent with the faster diffusion of fake news on social media, as documented by [10].

In Panels C and D, we report the distribution of news categories across industries and over time. At the industry level, the wholesale, retail, business equipment, and consumer nondurable industries are more likely to be targeted by fake news (columns 1 and 2). Fake news, however, is more likely to be foreign originated in the finance, chemicals, and consumer durables industries (column 3). In temporal trends, the fake news peaks in 2015, after which it gradually starts to decline (Panel D). Panel E reports the distribution of news by fact-checking organizations. Most of the fact-checked news is collected by Snopes (67%). We also calculate the number of days it takes for a fact-checking organization to check a claim. To do this, we manually read and identify the original date of the claim (i.e., the source news). We then subtract the publication date of the fact-checking article from the date when the source news began to disseminate news in the public domain. We find Politifact to be faster than other organizations in fact-checking the claims. We also manually identify the platforms where the source news is disseminated. Panel F reports the results. We find that the fake news is disseminated mostly on social media (e.g., Twitter, Facebook, Youtube). In contrast, non-fake news originates mostly from news sites, consistent with the filtering role that the editorial process plays in traditional media.

3.3 Country-level analysis

We next explore the distribution of corporate news at the country level. Our analysis is motivated by the fact that ideologically motivated foreign actors may spread misinformation to attack the reputation of a country’s firms. For example, a recent study finds evidence of social media manipulation campaigns in 70 countries organized by government agencies to shape public attitudes [41]. Several state actors target foreign countries to influence global audiences, amplify hate speech, or harass political figures or journalists, and countries use troll farms in the Middle East and Africa to spread rumors about target countries (see, e.g., https://www.theguardian.com/technology/2020/mar/13/facebook-uncovers-russian-led-troll-network-based-in-west-africa.) The country-level analysis may help shed light on the geographical origin of rumor spreaders (and potential foreign involvement) in capital markets.

Because the number of Twitter users varies across countries, we begin our analysis by normalizing the number of fake (and non-fake) news stories originating from a country by the total number of users in that country in a given year. We use the number of users in the randomly collected geo-tagged tweet dataset as a proxy for the total number of users in that country (as we could not obtain the historical country-level user statistics from Twitter or external data providers). To reduce outlier effects, we include countries with at least 300 tweets in the training data. The top five countries with the highest number of Twitter users are the United States, Indonesia, Brazil, Turkey, and Great Britain. We take the average of the yearly normalized Foreign Corporate Fake News (%) at the country-year level to construct our measure at the country level.

Table 4 reports the results. In Panel A, we show that corporate fake news originates primarily from Middle Eastern and African countries. Oman, Jordan, Morocco, Qatar, and Ghana are the top five countries from which most tweets with corporate fake news originate. The probability of each Omani user originating corporate fake news is 2.01%, while the probabilities for Jordanian and Moroccan users are 1.76% and 1.47%, respectively. In contrast, non-fake news originates mostly from Western countries. The probability of each Austrian user originating non-fake news is 0.63%, while the probabilities for Finnish and Polish users are 0.60% and 0.59%, respectively. In Fig 3, we plot the geographical distribution of 10 countries spreading fake (Panel A) and non-fake (Panel B) news about U.S. public firms. In Panel B, to account for the general business interests of users from a specific country, we construct an adjusted measure by subtracting the probability of disseminating non-fake news from that of disseminating fake news. Using this measure, we still find that the foreign fake news originates primarily from the Middle East and Africa (i.e., Oman, Jordan, Qatar, and Morocco), consistent with the anecdotal evidence that foreign actors use troll farms in this region. For example, a Russian-led network of professional trolls (operated by local residents) targeting the U.S. was discovered in Ghana and Nigeria [42]. There is also evidence of Lebanese, Nicaraguan, and Moroccan governments running disinformation campaigns for political motives (see, e.g., https://www.newarab.com/analysis/disinformation-and-electronic-armies-lebanons-elections for Lebanon, see https://www.bbc.com/news/world-latin-america-59129894 for Nicaragua, and see https://www.accessnow.org/how-pro-government-media-in-morocco-use-fake-news-to-target-and-silence-rif-activists/ for Morocco).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Geographical origins of corporate fake news.

This graph plots the geographical location of users spreading corporate news about U.S. public firms. We construct Foreign Fake News (%) and Foreign Non-Fake News (%) at the country level in two steps. First, we normalize the number of users spreading (fake or non-fake) news on Twitter by the total number of users in each country. We then take the averages of this normalized country-year-level measure to convert it to a country-level measure. In Panel A, we plot the top 10 countries spreading corporate fake news about U.S. public firms. In Panel B, we plot the top 10 countries spreading corporate non-fake news about U.S. public firms. The map in the figure is made with Natural Earth (https://www.naturalearthdata.com/about/terms-of-use/).

https://doi.org/10.1371/journal.pone.0301364.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Geographical distribution of corporate fake news.

https://doi.org/10.1371/journal.pone.0301364.t004

Next, we employ the KL divergence metric to compare the country distribution of the percentage of Twitter accounts spreading fake and non-fake news. KL divergence is a measure of how one probability distribution differs from a second distribution. Intuitively, we examine how far away the geographical distribution of the percentage of accounts spreading fake news is from the geographical distribution of the accounts spreading non-fake news. If the two distributions match perfectly, KL divergence is zero; otherwise, it can take values between zero and infinity. We find that the KL divergence is 0.62 and statistically significant (SE = 0.13, 95% CI[0.37,0.88]), which suggests that the geographical distribution of fake news is different from that of non-fake news. We also compare the concentration of geographic locations using a median relative polarization index (MRP). A positive MRP implies a more uneven distribution of news across countries relative to a benchmark. We take the distribution of locations spreading non-fake news as our benchmark and compute the MRP. We find that the MRP is relatively higher within countries spreading fake news (MRP = 0.33, SE = 0.11, 95% CI[0.12,0.54]), suggesting that the distribution of countries spreading fake is more concentrated than that of countries spreading non-fake news.

Finally, we examine if the probability of foreign fake news is more pronounced during periods of high geopolitical risks. To do this, we plot a time series of the news originating from foreign accounts and geopolitical risks. We use two proxies for geopolitical risks. First, we employ a news-based measure of the Geopolitical Risk Index (GPR) developed by [43]. The authors of [43] (on page 1195) define geopolitical risks “as the threat, realization, and escalation of adverse events associated with wars, terrorism, and any tensions among states and political actors that affect the peaceful course of international relations.” They construct an index at the country-year level by counting the share of articles mentioning adverse geopolitical events in leading newspapers in the U.S. A higher level of GPR corresponds to escalated geopolitical tensions facing the U.S. Second, we use the Global Database on Event, Location, and Tone (GDELT) to construct an Interstate Conflict Index. GDELT is the most widely used database for studying international relations and conflicts. It contains more than 200 million geolocated events compiled from international news sources [44]. The sources include AfricaNews, Agence France Presse, Associated Press, BBC Monitoring, United Press International, The Washington Post, The New York Times, and Google News. After machine coding the relevant information in the text of a news story (e.g., related countries, type of event, intensity of conflict or cooperation) into events, GDELT merges all duplicate events into a single event record. GDELT then provides the Goldstein scale [45], measuring the impact of each event from -10 (most conflictual) to +10 (most cooperative). We calculate the annual average Goldstein scale between the U.S. and other countries and reverse the scale to make it an increasing function of the conflicts that the U.S. faces.

We use these two indices—GPR and the Interstate Conflict Index—to explore the relation between the news originating from foreign accounts and geopolitical tensions. We calculate the annual average of Foreign Corporate Fake News (%) (at the U.S. level) to capture the prevalence of foreign accounts targeting U.S. firms. For ease of interpretation, we standardize Foreign Corporate Fake News (%) and the indices. We plot the temporal evolution of the variables in Fig 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Foreign-originating fake news and geopolitical tensions.

This figure plots the time series of corporate news originating from foreign countries and geopolitical tensions. Foreign Fake News (%) is the percentage of original tweets spreading corporate fake news initiated by a foreign (non-U.S.) Twitter account. We reconstruct this measure by aggregating all the corporate news at the yearly level. Panel A plots the relation between Foreign (Non-fake) Fake News (%) and Interstate Conflict Risk. Interstate Conflict Risk is an index measuring interstate conflicts (based on the Goldstein scale) using daily reported events in the global news media. Panel B plots the relation between Foreign Fake (Non-fake) News (%) and Geopolitical Risk Index. Geopolitical Risk Index is the share of articles mentioning adverse geopolitical events in leading newspapers in the U.S. For ease of interpretation, we standardize the indices and news measures.

https://doi.org/10.1371/journal.pone.0301364.g004

Panel A shows the close comovement of Foreign Corporate Fake News (%) and Interstate Conflict Index. There is a positive correlation (0.40) between the proportion of fake news from foreign accounts and interstate conflict risk. The correlation, however, is weaker (0.17) for Foreign Corporate Non-fake News (%). We observe a similar pattern in Panel B using GPR Index. The correlation between GPR and Foreign Corporate Fake News (%) is positive (0.09), but that between GPR and Foreign Corporate Non-fake News (%) is negative (-0.48). Overall, the increased fake news originating from foreign countries during periods of high geopolitical tensions suggests a link between foreign accounts and state actors. However, given our data limitations and descriptive research design, we interpret our evidence as merely suggestive.

3.4 Firm-level analysis

It is an empirical question what kind of firms are more likely to be targeted by fake news.

At the country level, we show a comovement between heightened geopolitical risks and foreign-originating fake news. At the firm level, we predict that firms in strategic industries (i.e., the telecommunication, pharmaceutical, semiconductor, military, and computer industries) and firms that are leaders in their industries are more likely to be targeted by foreign actors, given the recent proliferation of foreign-originated cyberattacks that have damaged the reputations of targeted firms in strategic industries (see, e.g., https://nyti.ms/3jsdGSR, https://bit.ly/3twMIho). Similarly, foreign actors can strategically disseminate fake news on social media to tarnish the reputations of strategically important firms, as reputation is often considered a firm’s most important intangible asset [46]. A significant body of research shows the effect of reputation on asset prices [47, 48], firm sale [49], risk and financial policy [50], investor preferences [51], and consumer behavior [52, 53]. Ideologically motivated actors may use Twitter to strategically target important industries, as reiterated news conveyed via multiple channels can reach a wide audience of individual and institutional investors [20].

To empirically test this prediction, we construct Foreign Fake News (%) by identifying the proportion of fake news originating from foreign accounts for a specific firm in a given year. For example, if 100 accounts initiate fake news about a firm in a given year, and 40 of these accounts are from a foreign country, Foreign Fake News (%) is 40%. We employ both univariate and regression analyses to examine the characteristics of firms targeted by fake news. Table 5 Panel A reports univariate results. Our firm-level analysis does not include 22 (11) private firms that are subject to fake (non-fake) news because we do not have financial data for this subset of firms. In the sample, we have 59 unique public firms (465 firm-years) targeted by fake news and 17 unique public firms (145 firm-years) subject to non-fake news. The firms targeted by fake news are significantly more profitable and have greater foreign sales. Target firms have higher pooled average Return on Assets (ROA) (0.12 versus 0.08) and lower Book-to-Market ratio (0.30 versus 0.50). Industry competition (as measured by TNIC HHI) also increases the probability of a firm being targeted by fake news. The TNIC HHI of target firms (0.35) is higher than that of firms with non-fake news (0.24). Firms in strategic industries (i.e., pharmaceutical, semiconductor, computer, defense, and telecommunication) are more likely to be targeted by fake news (0.22 versus 0.12). More importantly, the fake news is more likely to originate from foreign accounts. 14.21% of fake news is initiated by a non-US Twitter account versus 6.37% of non-fake news (see S5 Appendix in S1 File for the variable definitions).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Determinants of being targeted by foreign-originating fake news.

https://doi.org/10.1371/journal.pone.0301364.t005

In Table 5 Panel B, we estimate the following model to examine the likelihood of a firm being targeted by fake news originating from foreign Twitter accounts: (1) where Foreign Fake (Non-fake) News (%) is the percentage of fake (non-fake) news spread by foreign accounts for a specific firm a given year. i, t, and s denote the firm, year, and industry subscripts, respectively. We use year and industry fixed effects to control for temporal trends and time-invariant industry characteristics. Because we are interested in the across-industry variation, we exclude industry fixed effects when estimating the model with Strategic Industry and Industry Leader. Standard errors are clustered at the firm level to avoid underestimation [54]. A firm can be exposed to both fake and non-fake news in a given year. Therefore, in the empirical analysis, we separately analyze firm-years where there is fake or non-fake news (and compare them with firm-years without fact-checked news).

In the model, we include a vector of firm-level characteristics, including size (Total Assets), profitability (Return on Assets), leverage (Leverage), book-to-market ratio (Book-to-Market), dividend (Dividend Dummy), and a loss dummy (Loss). We do not have a clear directional prediction for these variables. To capture a firm’s visibility in foreign markets, we also control for Foreign Sales, defined as a dummy variable equal to one if at least 10% of a firm’s sales are to foreign (non-U.S.) markets. In addition, we control for TNIC HHI and Product Similarity to capture market competition. TNIC HHI is calculated as the sum of the squared market shares of all firms operating in the same industry, using the time-varying Text-based Network Industry Classification (TNIC) developed by [55]. Product Similarity is a firm-level measure based on product descriptions from 10-K filings [55]. We include these measures because peer firms can spread disinformation about their rivals, especially in competitive markets [25]. For example, the authors of [25] document negative peer disclosure (NPD) as an emerging corporate strategy firms use to publicize adverse news about their industry peers on social media. Moreover, NPD propensity increases with product market rivalry, as highly competitive industries provide greater incentives to spread negative news.

We employ Institutional Ownership and Return Volatility to control for the relation between the information environment and the fake news. Institutional Ownership represents the percentage of a firm’s stock held by institutional investors. Return Volatility captures uncertainty regarding a firm’s underlying fundamentals, which we measure as the standard deviation of a firm’s daily stock returns during a fiscal year. A poor information environment can create incentives for rumormongers to spread fake news, as investors are more likely to be influenced by news when access to alternative information sources is limited. We use these variables to control for the relation between a firm’s information environment and the likelihood of being targeted by fake news.

Table 5, Panel B summarizes the results. In column 1, we use Foreign Fake News (%) as the dependent variable in the baseline model. In column 2, we estimate a benchmark model using Foreign Non-fake News (%) for comparison. We find that larger firms with foreign sales and growth opportunities are more likely to be targeted by foreign fake news. Column 3 compares the coefficients across the baseline and benchmark models. The coefficient estimates of target firms are also significantly different from those of non-target firms. Second, we find that firms operating in less competitive markets (i.e., industries with higher TNIC HHI) are more likely to be targeted by foreign fake news. The coefficient estimate on TNIC HHI, however, is not statistically different from that on non-fake news (χ² = 0.00, p<0.99).

Next, we examine whether firms with an uncertain information environment (i.e., higher return volatility) and a less sophisticated investor base (i.e., high retail ownership) are more prone to foreign-originating fake news. Information frictions may slow the price-discovery process and can cause prices to deviate from intrinsic values for prolonged periods [56, 57], which increases the influence of rumormongers on stock prices. Theoretical literature suggests that rumors create profit opportunities for rumormongers [58], and recent work shows that corporate fake news can affect stock prices [11, 59, 60]. Consistent with our conjecture, we find the incidence of fake news to be higher for firms that have a less robust information environment. The likelihood of being targeted by a foreign source is higher for firms with lower institutional ownership (-0.010 versus -0.006, χ² = 5.63, p<0.02) and higher return volatility (0.027 versus 0.010, χ² = 3.07, p<0.08).

In columns 4 and 5, we examine whether fake news from foreign accounts is concentrated in strategically important firms. To test this, we use two variables. First, we define Industry Leader as an indicator equal to one if a firm is the largest member of its industry in terms of revenue, and zero otherwise. Second, we define Strategic Industry as an indicator equal to one if a firm is in the computer, telecommunication, pharmaceutical, semiconductor, or defense industry, and zero otherwise. We predict and find that firms that are leaders in their industries, as well as firms operating in strategic industries, are more likely be targeted by fake news originating from foreign countries. Being a member of a strategic industry increases the proportion of fake news from foreign countries by 2.9% (0.004/0.137). The coefficient estimate on Strategic Industry for target firms (0.004) is statistically different from the coefficient estimate for non-target firms (0.002) with χ² = 2.63 and p<0.10. Industry leaders are also more likely to be targeted by foreign accounts (0.021 versus 0.016). However, the difference of coefficient estimates is not statistically significant between firms subject to fake and non-fake news (χ² = 0,58 and p<0.45).

Overall, the firm-level analysis complements our country-level findings by showing that strategic industries have a higher probability of being targeted by fake news originating from foreign countries. That said, as discussed in the Introduction, we cannot cleanly attribute our findings to a foreign state actor or an intentional disinformation operation. Other potential reasons for disseminating negative fake news include the financial incentives of investors holding short positions or participating in other schemes that would benefit from a negative market reaction. In addition, inter-firm competition incentives may induce competitors to spread false news about their rivals. Finally, social media algorithms may unintentionally amplify certain content, including misinformation, based on factors like user engagement or click-through rates. While these are all plausible possibilities, they do not line up well with our country- and firm-level findings (i.e., the positive correlation between foreign-originated fake news and heightened geopolitical risks, as well as the targeting of strategic industries). We acknowledge, however, that we cannot fully rule out these alternative explanations, given the aforementioned limitations and the lack of granular data.

3.5 Additional analysis

In this section, we conduct three additional tests. First, we exclude political corporate fake news from our sample as politics is the most common topic found in corporate fake news (37.7%). Table 6, Panel A summarizes the results. Our analysis shows that the results hold in the sample of fake and non-fake corporate news (except that some differences in the coefficient estimates are statistically insignificant).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Additional analysis.

https://doi.org/10.1371/journal.pone.0301364.t006

Second, in the analysis above, we use Foreign Fake News (%) as a continuous measure to capture the prevalence of foreign-originated fake news. Alternatively, we create an indicator variable Foreign Fake News (dummy) that is equal to one if a firm is targeted by foreign-originated fake news, and zero otherwise. Unlike the continuous measure, Foreign Fake News (dummy) takes a value of one if at least one initiator of fake news is a foreign account. Panel B summarizes the results. The findings are robust to the alternative use of this indicator variable.

Third, our inferences are robust to alternative clustering of standard errors, such as two-way clustering at the firm and year level (untabulated).