1.1 The Problem of Fake News

There is no simple definition of fake news. The term can be used to refer to any news that is suspected to be inaccurate, biassed, misleading, or fabricated. This includes news originating from across the news media landscape, from anonymous blogs to mainstream newspapers. The term is often used by the public, politicians, and the news media to attack news, journalists, and news outlets deemed to be problematic. It is even common for allegations of fake news made by one outlet to be labelled as fake news by another. During the 2016 American presidential election, Hillary Clinton, the Democrats, and the mainstream press claimed that fake news from social media accounts, right-wing news outlets, and foreign governments was propelling Donald Trump to victory, while Trump, the Republicans, and the right-wing press claimed that Clinton, the Democrats, and the mainstream press were spreading fake news about these and other scandals to undermine Trump’s campaign (Reference Allcott and GentzkowAllcott & Gentzkow 2017). Fake news became the focus of the news, with news organisations arguing over whose news was faker.

Given this situation, how can the public judge what news is real and what news is fake? We cannot trust the news media to lead public inquiry into its own practices, nor can we trust the government or industry to monitor the news media, as they are most often the subject of the news whose validity is being debated. Academic research on fake news is therefore especially important, but it is also difficult to conduct (Reference Lazer, Baum and BenklerLazer et al. 2018). Researchers must define fake news in a specific and meaningful way and then apply this definition to identify instances of real and fake news for analysis. This is a challenging task. Any piece of news communicates a wide range of information, some of which can be true, some of which can be false, and all of which can be an opinion. Often the only way to verify if news is real or fake is to conduct additional independent investigation into the events being covered. Crucially, even if fake news is defined precisely and in a way that is acceptable to most people, researchers must still label individual pieces of news as fake that a substantial proportion of the public believe are real. The study of fake news therefore quickly becomes politicised, further eroding public confidence, and encouraging researchers to define fake news in such a way that data can be collected easily and uncontroversially, often moving research further away from the central problem of fake news.

To understand the central problem of fake news, it is important to consider the history of fake news. Although most current research focuses on the very recent phenomenon of online fake news, reviewing the history of deception in the news media can help researchers understand what communicative events are considered fake news, how these different forms of fake news are related, and which types of fake news should be of greatest scholarly concern. The history of fake news can also point us to specific cases for further analysis, depoliticising the study of fake news by allowing researchers to focus on news coverage of events that are of less immediate consequence.

The history of fake news is almost as old as the history of news itself. In Europe, the precursor to the modern newspaper were the avvisi, handwritten political newsletters from Italy that circulated across the continent during the sixteenth and seventeenth centuries (Reference Infelise, Signorotto and ViscegliaInfelise 2002). Unlike personal letters, the avvisi were intended to report general information and to be widely read. Unsurprisingly, we can find reports almost immediately of authorities questioning the veracity of the information being presented and the motives of their authors, who were generally anonymous. For example, in 1570, Pope Pius V executed one suspected author, Niccolo Franco, for defaming the church. Alternatively, the Italian scholar Girolamo Frachetta considered whether the avvisi could be used in wartime to spread false information to the enemy in his 1624 political treatise A Seminar on the Governance of the State and of War (Reference Infelise, Signorotto and ViscegliaInfelise 2002).

Indeed, there are many cases of fake news being used to mislead foreign populations and governments, exactly as Frachetta suggested. Many historical examples come from the Cold War, especially the Soviet use of dezinformatsiya, the purposeful spread of false information, which was often spread via the foreign press (Reference Cull, Gatov, Pomerantsev, Applebaum and ShawcrossCull et al. 2017). The word disinformation only entered the English language in the 1980s due to increased awareness of so-called active measures, a wide range of strategies used by the USSR for undermining foreign countries, including fake news. One of the most famous of these initiatives was ‘Project Infektion’, which involved the Soviets spreading rumours that the United States had engineered AIDS, initiated by a letter published in an obscure Indian newspaper in 1983, titled ‘AIDS may invade India: Mystery disease caused by US experiments’ (Reference BoghardtBoghardt 2009).

It is perhaps more common, however, for fake news during wartime to be directed at one’s own citizens – to encourage support for war and to manage expectations. For example, during World War I, the Committee on Public Information was established in the United States to influence the media and shape popular opinion, especially as President Woodrow Wilson had campaigned on staying out of the war (Reference HollihanHollihan 1984). Similar strategies were used to promote the Vietnam and Iraq wars. Most notably, we now know that reporting on the presence of weapons of mass destruction in Iraq after 9/11 was fabricated to build support for the war, especially to help Tony Blair justify the United Kingdom joining the coalition (Reference RobinsonRobinson 2017).

Fake news is not new, but the nature of fake news has shifted in recent years due to the growth of digital communication and social media (Reference Lazer, Baum and BenklerLazer et al. 2018). The Internet has changed the medium over which news is published and accessed by the public. Consequently, people now have access to a much wider range of news sources, which disseminate information continuously throughout the day, often from very specific perspectives, while social media provides a platform for people worldwide to share and discuss the news. One important effect of this new approach to the production and consumption of news is that people can focus exclusively on the information they want to hear, leading to what has become known as the echo chamber (Reference Del Vicario, Bessi and ZolloDel Vicario et al. 2016). The rise of blogging and social media has also given the opportunity to people from outside the mainstream news media to spread their own message, including potentially fake news.

This type of online fake news has been the focus of much concern in recent years, including in the lead up to the 2016 US Election. Perhaps the most notorious example was the ‘Pizzagate’ conspiracy theory, which went viral in 2016, after Wikileaks published the personal emails of John Podesta, Clinton’s campaign manager (Reference KangKang 2016). Extremist websites and social media accounts reported that the emails contained coded messages related to a satanic paedophile ring involving high-ranking officials, which allegedly met at various locations, including the Comet Ping Pong pizzeria in Washington. Provoked by these reports, a man travelled to the nation’s capital from North Carolina, shooting at the pizzeria with a semi-automatic rifle. The Covid pandemic also offers numerous examples of this type of online fake news (see Reference van Der Linden, Roozenbeek and Comptonvan Der Linden et al. 2020). For example, social media has been used to spread fake news about alternative treatments for Covid that are potentially deadly, including ingesting bleach (Reference OrganizationWorld Health Organization 2020).

Although it seems reasonable to assume that the amount of fake news has increased in recent years, we should not assume that the effect of fake news has become worse. Most notably, reporting from the mainstream news media leading up to the second Iraq War, which predates the rise of social media, was arguably far more damaging than anything that has happened since. In some ways, social media has even made it more difficult for certain types of fake news to spread by increasing public scrutiny of the news media and by amplifying alternative perspectives. An important example is coverage of the murder of George Floyd in May 2020. This event was filmed by a teenager named Darnella Frazier, who was walking to the store for groceries. She posted the video on social media, giving rise to widespread public protest to police brutality towards African Americans – a topic often overlooked by mainstream news media, which can be considered an example of fake news by omission (Reference WenzelWenzel 2019). In recognition of the importance of this act, which was only made possible by the existence of these non-traditional platforms for sharing information, Frazier received a Pulitzer Prize in 2021.

Overall, the problem of fake news is long-standing, pervasive, and potentially of great consequence, even leading to war. The study of fake news is therefore of true societal importance. Fake news is also very diverse, driven by a wide range of specific political, social, economic, and individual factors. In addition, it is clear that fake news, at least in its most troubling instantiations, is not simply characterised by inaccurate reporting: it is intentionally dishonest, designed to deceive as opposed to inform the public.

In this Element, we therefore adopt the view that fake news is most productively analysed as deceptive news, in contrast with most academic research on fake news, which focuses on false news. In other words, we define fake news based on the intent of the author: as opposed to real news, whose primary goal is to inform readers about new and important information that the journalist believes to be true, the goal of fake news is to deceive the public, to make them believe information that the journalist believes to be false. This approach not only forces us to concentrate on the most problematic forms of fake news, but, as we argue, it provides a more meaningful basis for the analysis of the language of fake news, which is the subject of this Element.

1.2 Fake News and Linguistics

Understanding the language of fake news is key to understanding the problem of fake news because most cases of fake news are language. Fake news can involve pictures and other media, but usually an instance of fake news consists primarily of a news text – an article in a newspaper, a report on the radio, a post on social media, an interview on television. The news text is the basic communicative unit of journalism and consequently the basic unit of analysis in most research on fake news. The main questions we pursue in this Element are therefore how can the language of fake news be analysed in a meaningful way? How can we describe linguistic differences between news texts that are real and news texts that are fake? And how can we understand why this variation exists?

Crucially, however, we should not assume that the language of real and fake news differs systematically. There has been considerable research in natural language processing (Reference Oshikawa, Qian and WangOshikawa et al. 2020) where machine learning models are trained to automatically distinguish between real and fake news based on patterns of language use, often achieving relatively high levels of accuracy. This may seem like evidence of variation between the language of real and fake news, but it is important to consider these results with care, especially as this research prioritises the maximisation of classification accuracy over the explanation of patterns of language use. There are two basic reasons for caution. First, the data upon which these systems are trained and evaluated may not isolate variation between real and fake news, especially given the inherent challenges associated with defining these terms and identifying cases of each. Second, these systems often focus more on variation in language content than variation in language structure: the identification of topical trends that tend to distinguish between real and fake news is different from the identification of stylistic variation in the language of real and fake news regardless of topic. In other words, research in natural language processing focuses on the language of fake news, but it does not necessarily focus on the linguistics of fake news.

Although linguistic perspectives on fake news are limited, fake news is fundamentally a linguistic phenomenon, and its analysis should therefore be grounded in linguistic theory. To address this basic limitation with fake news research, we propose a framework for the linguistic analysis of fake news in this Element. Our framework is based on functional theories of language use, drawing especially on research on register variation (Reference Biber and ConradBiber & Conrad 2019), which has repeatedly demonstrated that differences in communicative purpose and context are reflected in linguistic structure. In addition, our framework is based on the distinction between misinformation and disinformation (Reference RubinRubin 2019), which we believe is crucial for understanding what fake news is and why the language of real and fake news should differ. By bringing together these two perspectives for the first time, we provide a basis for the linguistic analysis of fake news – for collecting real and fake news texts, for comparing their grammatical structure, and for understanding why this structure varies depending on whether their author intends to inform or deceive.

To demonstrate how our framework can be used to better understand the language of fake news, in this Element, we focus on one especially famous episode drawn from the history of the news media. This case involves Jayson Blair, a young reporter at The New York Times, who published a series of fabricated news articles in the early 2000s (Reference HindmanHindman 2005). In addition to its notoriety, this case is especially well suited for the linguistic analysis of fake news for three reasons. First, there is a relatively large amount of real and fake news available from one author and from the same time period, which has been validated through an extensive investigation by The New York Times (Reference Barry, Barstow, Glater, Liptak and SteinbergBarry et al. 2003) and acknowledged by Blair himself (Reference BlairBlair 2004). This gives us a controlled context for the study of fake news, where we have substantial amounts of comparable and valid real and fake news data, allowing us to effectively isolate the effects of deception on the language of one journalist. Second, we know much about why Blair wrote fake news, including from his own account and the account of The New York Times, giving us a basis for explaining differences in language use that we observe. Third, this case is relatively uncontroversial, as it is old enough that everyone can agree that the articles in question were faked, regardless of their political outlook – an important factor that has often limited the societal impact of fake news research. In addition, the case reminds us that fake news can be found across the news media, including in one of the most respected newspapers in the world.

The analysis of the language of fake news grounded in linguistic theories and methods also opens up the possibility for a wide range of applications. Although our goal is not to develop systems for fake news detection, the most obvious application of our research is to support the language-based identification of fake news. Most notably, this includes considerable current research in natural language processing concerned with developing systems for automatically classifying real and fake news at scale via supervised machine learning (Reference Oshikawa, Qian and WangOshikawa et al. 2020). As noted above, these systems can achieve good results, but they are not designed to explain why the language of real and fake news differs, and they appear to focus more on the content of fake news than its linguistic structure (Reference Castelo, Almeida and ElghafariCastelo et al. 2019). Our framework is not intended to supplant these types of systems, but it can offer an explanation for why they work, or why they might appear to work, which is necessary to justify the real-world application of such tools. Furthermore, the identification of a principled set of linguistic features for the analysis of real and fake news can be used to enhance existing machine learning systems, which tend to be based on relatively superficial feature sets like the use of individual words and word sequences. These types of insights can be especially useful to improve performance on more challenging cases, which also seem likely to be the most important cases of fake news. In addition, our framework can directly inform how fake news corpora should be compiled in a principled manner for the robust training and evaluation of fake news detection systems, which is a major limitation in much current research on fake news (Reference Asr and TaboadaAsr & Taboada 2018).

The framework we propose is also of direct value to the detailed discursive analysis of individual cases of fake news of sufficient importance to warrant close attention. For example, in a legal context, empirical analysis presented as evidence in court is often required to be based on accepted scientific theory (Reference AllenAllen 1993). Until now, however, there has been no clear explanation for why the linguistic structure of real and fake news should be expected to differ systematically. Our framework also potentially provides a basis for extending discourse analytic methods for deception detection more generally in forensic linguistics, which is relevant across a wide range of areas, including for police interviews (Reference PicornellPicornell 2013). In addition, our framework can be of value for supporting work in investigative linguistics, which is an emerging field of applied linguistics that focuses on the application of methods for the study of language use to make sense of real-world issues currently in the news (Reference Grieve, Woodfield, Coulthard, May and Sousa-SilvaGrieve & Woodfield 2020).

More generally, understanding the language of fake news, and how it differs from the language of real news, is important for understanding the language of the news media, and, through this language, the biases and ideologies that underlie any act of journalism. The current fake news crisis reflects a growing and general distrust of the news media that cannot be rectified simply by developing systems for automatically detecting real and fake news with a reasonable degree of accuracy. An article that obliquely expresses the editorial view of a newspaper in a context that appears to be purely informational is not necessarily fake, but it has real societal consequences. Being able to recognise the motivations of journalists and news outlets through the analysis of their language is an important part of reading the news intelligently and holding the news media accountable. Studying the discourse of fake news is therefore part of the greater enterprise of understanding the expression of information, opinion, and prejudice in the news media – understanding how the language of the news media shapes the world around us and our perceptions of it. We therefore hope that our framework will also be valuable for the critical analysis of the news media (Reference van Dijkvan Dijk 1983).

Finally, our framework and its application can also help us better understand the psychology of fake news (Reference Pennycook and RandPennycook & Rand 2019, Reference Pennycook and Rand2021). Why do people create, share, and believe fake news? These are basic questions whose answers are central to understanding the phenomenon of fake news in the modern world. Most notably, as we demonstrate through our analysis of the case of Jayson Blair and The New York Times, variation in the linguistic structure of fake news reflects the specific communicative goals of authors who consciously write fake news and the production circumstances in which fake news is produced. Appreciating the linguistic structure of fake news can also potentially help us understand why some fake news is more likely to be believed and to be shared, which may be especially important for combating the spread of fake news online.

1.3 Overview

Fake news is a long-standing problem, but it is receiving unprecedented attention today due to the rise of online news and social media, as well as growing distrust of the mainstream news media. Although fake news most commonly involves news texts, the study of the language of fake news has been limited, with researchers focusing more on the automated classification of true and false news than on explaining why the structure of real and fake news differs. It is therefore crucial to extend our understanding of the language of fake news, especially through the detailed analysis of real and fake news texts collected in a principled manner and grounded in linguistic theory.

Given this background, the goal of this Element is threefold. First, we introduce a new linguistic framework for the analysis of the language of fake news, focusing on understanding how the linguistic structure of fake news differs from real news, drawing especially on the distinction between misinformation and disinformation and the concept of register variation. Second, based on this framework, we conduct a detailed analysis of the language of fake news in the famous case of Jayson Blair and The New York Times to identify and explain systematic differences in the grammatical structure of his real and fake news.Footnote ¹ Third, we consider how our results can help address the problem of fake news, including by informing research in natural language processing and psychology.

The remainder of this Element is organised as follows. In Section 2, we present a critical review of research on the language of fake news, before presenting our theoretical framework for the linguistic analysis of fake news, which directly addresses limitations with previous research. In Section 3, we review the case of Jayson Blair, including the background, the scandal, the investigation, and the aftermath. In Section 4, we describe the corpus of Jayson Blair’s writings that we collected, which is the basis of this study. In Section 5, we present our main linguistic analysis, discussing a range of grammatical features that vary across Blair’s real and fake news. We find that Blair’s fake news is written in a less dense style than his real news and with less conviction. We then offer explanations for these findings based on specific factors that led Blair to write fake news. Finally, in Section 6, we consider the implications of our research for our understanding of fake news more generally.

2.1 Defining Fake News

The first major challenge in the study of the language of fake news is to define fake news in such a way that instances of fake news texts can be identified and collected (Reference Tandoc, Lim and LingTandoc et al. 2018; Reference Asr and TaboadaAsr & Taboada 2019). There is, however, no simple or standard definition of fake news, which is better understood as the product of a range of practices that are related to the validity of information being shared by the news media. Researchers must therefore define the specific form of fake news they are interested in studying. Any coherent definition of fake news can be the starting point for meaningful empirical research, but researchers naturally tend to focus on certain types of fake news, depending both on the perceived societal importance of that type of fake news and the feasibility of collecting news texts of that type in a reliable and efficient manner – considerations that are often at odds with each other.

The vast majority of research on the language of fake news has been conducted in natural language processing and has focused on the development of tools for automatically distinguishing real and fake news (e.g. Reference Conroy, Rubin and ChenConroy et al. 2015; Reference Rubin, Chen and ConroyRubin et al. 2015; Reference Shu, Sliva, Wang, Tang and LiuShu et al. 2017; Reference Asr and TaboadaAsr & Taboada 2018; Reference Bondielli and MarcelloniBondielli & Marcelloni 2019; Reference Oshikawa, Qian and WangOshikawa et al. 2020; Reference Zhou and ZafaraniZhou & Zafarani 2020). In general, this research defines fake news as false news – untrue information disseminated by the news media. This has also been the definition that has been adopted in the very limited amount of linguistic research on this topic in discourse analysis (e.g. Reference Igwebuike and ChimuanyaIgwebuike & Chimuanya 2021). Crucially, this definition of fake news is based on the underlying truth of the information being conveyed: to study fake news from this perspective, comparable corpora of true and false news must be compiled. For example, to develop a machine learning system capable of distinguishing between true and false news requires that many true and false news texts be collected so that the system can be trained and tested on this dataset.

A major advantage of this veracity-based approach to fake news research is that it allows fake news to be collected with relative ease. Most commonly this involves drawing on the work of fact-checking organisations and mainstream news media organisations that identify fake news, including both instances of fake news and sources of fake news (Reference Asr and TaboadaAsr & Taboada 2018). This information is then used as a basis for compiling a corpus of fake news texts. These texts most commonly include passages from news articles (e.g. Reference Vlachos and RiedelVlachos & Riedel 2014; Reference WangWang 2017), social media posts (e.g. Reference Shu, Sliva, Wang, Tang and LiuShu et al. 2017, Reference Shu, Mahudeswaran, Wang, Lee and Liu2020; Reference WangWang 2017; Reference Santia and WilliamsSantia & Williams 2018), and complete news articles (e.g. Reference Rashkin, Choi, Jang, Volkova and ChoiRashkin et al. 2017; Reference Horne and AdaliHorne & Adali 2017; Reference Santia and WilliamsSantia & Williams 2018; Reference Castelo, Almeida and ElghafariCastelo et al. 2019; Reference Lin, Tremblay-Taylor, Mou, You and LeeLin et al. 2019; Reference Bonet-Jover, Piad-Morffis, Saquete, Martínez-Barco and García-CumbrerasBonet-Jover et al. 2021). Alternatively, some studies have used crowdsourcing to conduct their own fact-checking, having people rate the veracity of social media posts (e.g. Reference Mitra and GilbertMitra & Gilbert 2015) or news text (e.g. Reference Pérez-Rosas, Kleinberg, Lefevre and MihalceaPérez-Rosas et al. 2018). Crucially, these approaches not only make it possible for large collections of fake news to be compiled, but they maintain a certain degree of researcher objectivity, which is especially important in such a politicised domain. These collections of fake news are then generally contrasted with collections of true news, often collected from mainstream news media (e.g. Reference Horne and AdaliHorne & Adali 2017; Reference Rashkin, Choi, Jang, Volkova and ChoiRashkin et al. 2017; Reference Pérez-Rosas, Kleinberg, Lefevre and MihalceaPérez-Rosas et al. 2018; Reference Castelo, Almeida and ElghafariCastelo et al. 2019; Reference Bonet-Jover, Piad-Morffis, Saquete, Martínez-Barco and García-CumbrerasBonet-Jover et al. 2021).

In addition to the advantages of a veracity-based approach to the study of fake news, there are disadvantages. Most obviously, classifying news based on a true–false distinction inaccurately reduces the veracity of a news text down to a single binary variable. News texts, however, generally convey a large amount of information, some of which can be true and some of which can be false. Some fact-checking organisations and researchers have acknowledged this limitation, classifying fake news on a scale (e.g. Reference Rashkin, Choi, Jang, Volkova and ChoiRashkin et al. 2017; Reference WangWang 2017), although such an approach still assumes veracity can be reduced to a single quantitative variable.

Analysing the language of fake news based on veracity also requires that someone judges what qualifies as true and false news. Relying on external fact-checking services is a common and convenient solution to this problem, but while it may appear to increase the objectivity of a study, by limiting the involvement of the researcher at this stage of data collection, it actually immediately politicises the study, linking the research directly to the policies of the fact-checking organisation. Furthermore, these policies are not always consistent or accessible, making it difficult for the validity and biases of such research to be assessed, even by the researchers themselves. Fact-checking organisations may even be invested in the dissemination of fake news.

Finally, the veracity-based approach risks taking focus away from the type of deceptive fake news that is generally of greatest societal concern. People are not primarily worried about false news, as inaccuracies might be unintentional or inconsequential, but about deceptive news that is intended to manipulate readers, especially for establishing forms of political, social, and economic control (Reference GelfertGelfert 2018). It is important to consider the full range of practices encompassed by the term fake news, but we must not lose sight of those forms of fake news that are most problematic, even if they are inevitably more difficult to study. We certainly should not assume that natural language processing systems trained to identify false news can be used to identify deceptive news with similar levels of accuracy, especially as deceptive news is presumably the type of fake news that is most difficult for humans to identify.

The distinction between misinformation and disinformation is especially relevant to this discussion, because it helps us better understand the range of phenomena referred to as fake news, and because it helps us better understand what types of fake news should be of greatest concern. Specifically, misinformation is defined as false information, whereas disinformation is defined as deceptive information (Reference StahlStahl 2006; Reference RubinRubin 2019). The distinction between information and misinformation is therefore grounded in the concept of veracity, defined independently of the knowledge or intent of the individual, whereas the distinction between information and disinformation is grounded in the concept of deception, defined relative to the knowledge and intent of the individual. Essentially, this terminology reflects the difference between a falsehood and a lie.

It is important to acknowledge that the distinction between misinformation and disinformation pre-dates the rise of modern concerns about online fake news, and traditionally disinformation was recognised as the greater problem (Reference FallisFallis 2009). For example, Reference FetzerFetzer (2004: 231) wrote that

Fetzer was not referring to fake news directly but highlighting how deception is of far greater concern than inaccuracy. Nevertheless, the veracity-based approach to the study of fake news, which dominates current research, tends to focus implicitly on misinformation, at least in part because it is far easier to identify instances of misinformation than disinformation in the news media, for example, by drawing on the work of fact-checking organisations. This approach allows researchers to work at the scale required to train and test modern machine learning systems, but it has effectively moved focus away from the central problem of fake news – deceptive and dishonest news practices.

In most discussions of misinformation and disinformation, misinformation is presented as the larger category, including both information that is accidentally false and information that is purposely false (Reference StahlStahl 2006; Reference Tandoc, Lim and LingTandoc et al. 2018). In other words, disinformation is seen as a type of misinformation – purposeful misinformation. It is clearly necessary to draw a distinction between misinformation that qualifies and does not qualify as disinformation: people can inadvertently communicate falsehoods when they intend to share accurate information, and this should not be confused with lying. For example, a journalist might report false information obtained from a source, who the journalist believed was telling the truth. This qualifies as misinformation, but not disinformation, which only occurs when the journalist lies, reporting information they believe to be false.

It is wrong, however, to insist that all disinformation is misinformation, as people can also state the truth when they intend to deceive, if they are misinformed themselves. If Democritus, who believed the earth was flat, tried to convince his student that the earth was round (e.g. as part of a lesson), he would be deceiving his student, but inadvertently telling the truth. Situations such as these are presumably rare in journalism, although sustained disinformation campaigns would likely result in the production of some truthful disinformation over time. Perhaps a more common form of truthful disinformation in the news media is deception by omission, where the information contained in a news text is true, but important information is purposely excluded to manipulate the reader (Reference WenzelWenzel 2019). Most notably, this type of fake news would include selective and biassed reporting. For example, the American news media has been criticised for reporting mass shootings differently, excluding relevant information, depending on the race of the shooter (Reference Duxbury, Frizzell and LindsayDuxbury et al. 2018). Real news could even potentially contain purposeful falsehoods and contradictions, so as to allow true information to be communicated that could not otherwise be published openly (Reference StraussStrauss 1952).

The relationship between misinformation, disinformation, and fake news is illustrated in Figure 1, which shows the overlap between disinformation and misinformation along our two main dimensions of fake news: veracity and honesty (see also Reference Tandoc, Lim and LingTandoc et al. 2018). Prototypical real news is both honest and true: generally, the goal of journalists is to share information that is true and that they believe is true. However, there are three distinct ways news texts can diverge from this expectation, creating three broad categories of fake news. Type I Fake News is unintentionally false news, which occurs when journalists report information they believe to be true, but which is false. Type I Fake News therefore qualifies as misinformation but not disinformation. Alternatively, Type II Fake News is intentionally false news, which occurs when journalists purposely report information they believe to be false. Type II Fake News therefore qualifies as both misinformation and disinformation. Finally, Type III Fake News is news that is true but is nevertheless intended to deceive, including fake news by omission or selective reporting. Type III Fake News therefore qualifies as disinformation but not misinformation.

Figure 1 A typology of fake news

Our opinion is that research on fake news should primarily focus on disinformation, especially Type II Fake News – news that was intended to deceive its readership into believing information the journalist does not believe is true – as this is the type of fake news that we believe is of greatest societal concern. We also believe it is important not to conflate these different types of fake news: studies should either focus on one type of fake news, and compare it to real news, or distinguish between different types of fake news. In our analysis of the case of Jayson Blair and The New York Times, we therefore focus on comparing Type II Fake News, where Blair purposely published false information, to real news, where he purposely told the truth. Consequently, our framework and our study differ from most research on the language of fake news, which tends to focus on false news, effectively conflating Type I and Type II Fake News, obscuring the difference between misinformation and disinformation.

2.2 Language Variation and Fake News

The second major challenge in the study of the language of fake news is collecting comparable real and fake news texts for analysis. In general, research on the language of fake news, including fake news detection in natural language processing, is based on the comparison of patterns of language variation across texts that have been classified as real and fake. If this research is to identify actual differences between real and fake news, these texts must otherwise be as comparable as possible. We know, however, that there are many other factors that also naturally cause language variation across news texts, including variation in register, author, and dialect. Consequently, once fake news has been defined in a clear and practical manner, it is still necessary to identify real news for comparison that allows for these other sources of linguistic variation to be somehow controlled. The challenge is to build corpora that allow us to isolate variation between real and fake news.

Most importantly, to study the language of fake news, it is necessary to control for register variation, that is, variation based on communicative context and purpose (Reference Biber and ConradBiber & Conrad 2019). No matter how fake news is defined, we must contrast otherwise comparable registers of real and fake news. For example, if real news is collected from traditional newspapers and fake news is collected from blogs, it would be unclear if any observed differences in language use are related to differences between real and fake news or to differences between newspapers and blogs, which are associated with different communicative purposes and production circumstances, and which therefore show distinct patterns of grammatical variation (Reference BiberBiber 1988; Reference Grieve, Biber, Friginal, Nekrasova, Mehler, Sharoff and SantiniGrieve et al. 2010). Specifically, blog writing is more informal than newspaper writing for a wide range of reasons, independent of its status as real or fake news. We would also likely find clear topical differences between these two registers of news, especially, for example, if fake news from fringe outlets online is compared to real news from the mainstream press.

This type of register imbalance is common in fake news datasets used in natural language processing. For example, the LIAR dataset (Reference WangWang 2017), which has been used in many studies (e.g. Reference Shu, Sliva, Wang, Tang and LiuShu et al. 2017; Reference Aslam, Khan, Alotaibi, Aldaej and AldubaikilAslam et al. 2021), consists of news statements that were scored for veracity by a major fact-checking organisation. The register of these statements, however, varies substantially, including statements drawn from news reports, campaign speeches, and social media posts. No information is provided on how these registers were selected or whether the distribution of true and false statements is balanced across these categories, nor is register variation generally taken into account during analysis. Similarly, the Buzzfeed fake news dataset (Reference Silverman, Lytvynenko, Vo and Singer-VineSilverman et al. 2017), which has also been used in many studies (e.g. Reference Shu, Sliva, Wang, Tang and LiuShu et al. 2017; Reference Mangal and SharmaMangal & Sharma 2021), was designed to allow for real and fake news to be compared between mainstream news outlets and extremist left- and right-wing websites. The dataset, however, is accompanied by descriptive statistics showing that data from mainstream news is rated at over 90 per cent true, whereas data from extremist websites is rated at around 50 per cent true. Fake news identification systems trained on such datasets may achieve high levels of accuracy, but this classification would likely be driven primarily by broad register differences between the texts selected to represent real and fake news as opposed to whether a text was real or fake, especially when fake news is largely concentrated in one register.

A related issue is that much research on the language of fake news tends to focus only on part of the news media landscape, especially non-traditional online sources, often associated with extreme political viewpoints, as the Buzzfeed fake news dataset clearly illustrates. This type of fake news is convenient to study because these are the sources that most fact-checking organisations target, and because labelling news from such sources as fake is less likely to be questioned, at least within the scientific community. There is no reason to doubt that fake news, including genuine disinformation, originates from these sources, but it is wrong to assume that these are the only sources of fake news, as our case of Jayson Blair and The New York Times and many of the examples we have considered thus far illustrate. In many ways, these are not even the most serious sources of fake news, because texts from non-traditional outlets are often obviously fake to a large proportion of the public, and because their status as potential fake news can often be inferred based entirely on source. Mainstream news media is generally excluded from fake news research, except as the source of true news comparison data. As a result, most fake news research does not generalise to fake news published by the mainstream news media, as fake news detection systems are trained to treat all mainstream news as real, even though this is the type of fake news that has the potential to be far more significant and difficult to detect.

In addition to register variation, it is also important to control other forms of linguistic variation when collecting comparable examples of real and fake news for analysis. For example, research in stylometry has shown that each author has their own style, with different sets of linguistic features distinguishing between the unique varieties of language used by individuals (Reference GrieveGrieve 2007). Certain features used by one author to create real or fake news may therefore be different from other authors, introducing possible confounds in large-scale corpus-based research on fake news that does not attempt to control for individual differences.

Furthermore, we know the social background of authors more generally affects their writing style (Reference GrieveGrieve 2016), so even taking large random samples of real and fake news written by many authors will not necessarily allow for sociolinguistic variation to be controlled. For example, it seems likely that when fake news is drawn from fringe online sources, and real news is drawn from the mainstream press, there will be clear social differences between the two sets of authors, especially in terms of education level and socioeconomic status, which are well-known correlates of dialect variation (Reference TagliamonteTagliamonte 2006). The apparent differences observed between real and fake news in previous studies therefore may also reflect differences in the social backgrounds of the authors being compared, as opposed to the status of the news as real or fake. For example, if authors of real news are primarily professional authors, while authors of fake news are primarily amateurs, we would expect broad differences in the language in which they write, regardless of whether or not they are telling the truth. Some research has begun to address these types of issues at least obliquely. For example, Reference Potthast, Kiesel, Reinartz, Bevendorff and SteinPotthast et al. (2017) found that hyper-partisan left- and right-wing news appears to share a style associated with the language of extremism. However, controlling for social variation in the authors of real and fake news is a topic that has received remarkably little attention in the literature, and represents another major limitation of current research on the language of fake news.

2.3 A Framework for the Linguistic Analysis of Fake News

To summarise, there are two key limitations with previous research on the language of fake news: researchers do not generally analyse the most problematic forms of fake news, focusing on misinformation as opposed to disinformation; and researchers do not generally control for other forms of linguistic variation, including variation in register, authorship, and dialect. These issues stem in part from reliance on fact-checking services for identifying fake news, as well as from the lack of a clear and meaningful definition of fake news. Consequently, researchers often miss the central problem of fake news or fail to isolate the distinction between real and fake news. To address these issues, we propose a framework for the linguistic analysis of fake news that is grounded in theories of disinformation and register variation.

We believe the language of fake news can best be understood as a form of register variation between information and disinformation in the news media. If we define fake news based on the communicative purpose of the journalist, as our focus on disinformation demands, the theories and methods of register analysis provide us with a basis for analysing the language of fake news in a meaningful way. Research on register variation has repeatedly shown that the use of a wide range of grammatical features varies systematically across contexts depending on their function (Reference HallidayHalliday 1978; Reference Chafe and TannenChafe & Tannen 1987; Reference BiberBiber 1988; Reference Biber and ConradBiber & Conrad 2019). Because the communicative goals of people change across contexts, along with the inherent communicative affordances and constraints associated with those contexts, the structure of discourse varies systematically across contexts as well. Grammatical differences between registers are not arbitrary, but directly reflect how people vary their language for effective communication in different situations. As Reference HallidayHalliday (1978: 31–2) writes, ‘the notion of register is at once very simple and very powerful. It refers to the fact that the language we speak or write varies according to the type of situation’. For example, when people tell stories, they tend to use many past tense verbs, because they are recounting events that took place in the past (Reference BiberBiber 1988). These grammatical differences exist because the communicative needs of people vary across different contexts.

For this reason, we should expect that the language of real and fake news also varies systematically, if subtly, so long as real and fake news are defined based on the intent of the journalist, for instance, to inform or deceive. Register variation can therefore provide a basis for describing and explaining grammatical differences in the language of real and fake news, specifically between information and disinformation. This link between the concepts of disinformation and register variation is the basic theoretical insight that underlies our framework for the linguistic analysis of fake news.

Our framework draws most directly on the type of quantitative corpus-based register analysis developed by Douglas Biber and his colleagues, known as multidimensional analysis (Reference Biber and ConradBiber & Conrad 2019). Most notably, Reference BiberBiber (1988) presented an extensive analysis of linguistic variation across registers of written and spoken British English, identifying clear grammatical differences across registers, showing that these patterns derive from variation in communicative purpose and context. In total, six aggregated dimensions of register variation were extracted from the corpus based on a multivariate statistical analysis of sixty-seven grammatical features, whose relative frequencies were measured across each of the individual texts in the corpus. These dimensions were then interpreted functionally based on the most strongly associated features and texts. For example, the first dimension, which accounts for the largest amount of variance across the feature set, was interpreted as reflecting a distinction between more informational and more involved forms of communication. On the one hand, texts from registers like academic and newspaper writing were found to be relatively informational and formal, characterised by frequent use of features like nouns and noun modifiers, which are associated with a relatively dense style of communication. On the other hand, texts from registers like face-to-face and telephone conversations were found to be relatively involved and informal, characterised by frequent use of features like verbs, pronouns, and adverbs, which are associated with relatively casual and spontaneous discourse, and consequently lower levels of informational density. For example, conversations tend to contain many pronouns due to natural communicative constraints on the production of spontaneous speech, which greatly limit opportunities for individuals to compose complex noun phrases. Instead, individuals tend to repeatedly reference entities under discussion using pronouns, adding new information with each reference. Other dimensions of register variation identified in Reference BiberBiber (1988) are related to factors like narrativity, persuasion, and abstractness.

The extensive and interpretable set of grammatical features identified in Reference BiberBiber (1988) provides us with a basis for identifying meaningful differences between the linguistic structure of real and fake news that are directly related to variation in communicative function. These features have been validated through a long history of use for the empirical analysis of register variation in the English language, with features and clusters of features having been linked to a wide range of different communicative functions. These functional linguistic patterns can therefore help us understand the motivation for linguistic variation observed when comparing real and fake news. Notably, comparable feature sets have also been compiled for other languages, including Somali, Korean, and Tuvaluan (Reference BiberBiber 1995).

Drawing on insights from register analysis therefore allows us to address another basic limitation with current research on the language of fake news – reliance on relatively superficial feature sets, like the use of individual words and word sequences (i.e. n-grams), which are easy to count, but more difficult to explain, especially from a grammatical as opposed to a topical perspective. By describing differences between real and fake news based on an established set of grammatical features with clear functional relevance, we can overcome this limitation, and identify differences in the linguistic structure of real and fake news that are driven by differences in the communicative intent of journalists – to produce news texts that are intended to inform and news texts that are intended to deceive.

In addition to conceiving of variation between real and fake news as a form of register variation, our framework also acknowledges that there are other sources of linguistic variation in the news media, which might obscure differences between real and fake news if ignored. When compiling fake news corpora, it is especially important to control for other sources of register variation, as there are clear linguistic differences across news registers, reflecting variation in communicative purpose and context. For example, it would seem to be much easier for people to distinguish between different types of news texts – for example, newspaper articles and blog posts – than it is to identify whether the journalist is honest or dishonest. This is why fake news identification is such a challenging task. Controlling register variation can be achieved either by focusing on one news register or by factoring register variation into the analysis directly. Controlling register variation also helps control dialect variation, as the social background of people who produce different types of news inevitably differs, especially in terms of education and class. For example, it seems likely that writers at mainstream newspapers are generally more highly educated than social media posters. Such differences will have clear linguistic consequences that must somehow be controlled in any rigorous study of fake news. Finally, it is important to control directly for authorship to account for individual differences of style, ideally by considering the language of one or more authors who are known to have produced both real and fake news.

Overall, by treating disinformation in the news media as a form of register variation, by drawing on a rich and interpretable grammatical feature set, and by controlling for other important sources of linguistic variation, we believe it is possible not only to identify actual differences in the language of real and fake news but to offer explanations for why these differences exist, based on variation in the communicative goals of journalists, as well as the production circumstances in which they write. In this way, we address a final limitation with current research on the language of fake news – the lack of any underlying theory for why real and fake news should differ or any explanation for why observed differences in the language of real and fake news exist. Providing the basis for truly understanding the language of fake news is the primary goal of this Element. Furthermore, our framework should be of direct value to work on fake news detection in natural language processing, because it provides a theoretical basis for detection, which can help improve the performance of these tools, especially in the most challenging cases, and because it provides justification for the application of these tools in real-world settings.

Finally, it is important to acknowledge that while there is good reason to assume systematic linguistic differences exist between news media texts that are intended to inform and deceive, this assumption does not extend to news media texts that are true and false. This is because misinformation may be shared without the knowledge of the journalist, precluding the possibility that differences in language structure arise from differences in communicative purpose. In fact, from this perspective, it is unclear why we should expect systematic linguistic differences between true and false news to exist at all, further calling into question standard approaches to the analysis of fake news based on fact-checked data. Unfortunately, the study of disinformation is more challenging than the study of misinformation, as it requires knowledge of the intent of an author, which is often inaccessible to researchers. Nevertheless, as the case of Jayson Blair and The New York Times demonstrates, such cases do exist and can be identified if we take the time to review the history of the news media. In the next three chapters, we therefore focus on this one important case of fake news, not only to demonstrate the application of our framework but to begin to truly understand the language of fake news.

5.1 Quantitative Results

To compare patterns of grammatical variation in Blair’s real and fake news, we analysed our corpus using the Multidimensional Analysis Tagger (Reference Nini, Berber Sardinha and Veirano PintoNini 2019). The tagger computes normalised values for sixty-seven grammatical features across the texts in the corpus based on the feature set most commonly used in multidimensional register analysis (Reference BiberBiber 1988). As we are working with standard news texts, the tagger achieves over 90 per cent accuracy on our corpus. More information on the feature set and the tagger can be found in Reference BiberBiber (1988) and Reference Nini, Berber Sardinha and Veirano PintoNini (2019), including detailed descriptions of each feature. Although we use this tagset so that we can draw directly on insights from register analysis (especially Reference BiberBiber 1988), in practice, any sufficiently accurate part-of-speech tagger would allow for similar patterns to be broadly observed.

Out of these sixty-seven linguistic features, we excluded seventeen that occur very infrequently in our corpus (less than once per thousand words) because their low frequencies preclude the meaningful comparison of their relative frequencies across the texts in our corpus, which are mostly under 1,000 words long. We also excluded type-token ratio given the well-known issues with this measure (Reference BaayenBaayen 2002). This left us with forty-nine normalised linguistic features for further analysis all of which are measured per hundred words, aside from average word length, which is measured as the mean number of characters per word.

To identify differences in the values of these forty-nine features measured across each of the twenty-eight real articles and the thirty-six fake articles in our corpus, we computed Cliff’s delta (Reference CliffCliff 1993) using the effsize package in R (Reference TorchianoTorchiano 2020). Cliff’s delta is an ordinal measure of effect size that compares the difference in the values of a variable measured across two groups. Cliff’s delta is computed by taking all pairs of values across the two groups and computing the proportion of times that the value from one group is larger than the value from the other. Specifically, Cliff’s delta is calculated as

$\text{d}\text{=}\frac{\#(x_i>x_j)-\#(x_i<x_j)}{mn}$

where the variable x is measured across m observations for the first group and n observations for the second group, where #(x_i > x_j) is the number of times for all pairs of observations from the two groups that the value of the variable is higher for the first group, and where #(x_i < x_j) is the number of times for all pairs of observations from the two groups that the value of the variable is lower for the first group. A positive Cliff’s delta simply indicates that the first group is characterised by a higher proportion of larger values than the second group, whereas a negative Cliff’s delta indicates that the second group is characterised by a higher proportion of larger values than the first group.

We chose to use a non-parametric ordinal test of effect size because many of our features are positively skewed, with many texts having no occurrences of that feature. This makes more common measures of effect size like Cohen’s d inappropriate. Cliff’s delta has also been shown to be generally as powerful as Cohen’s d when these assumptions are met (Reference CliffCliff 1993), making it suitable for analysing our full feature set. We do not conduct inferential statistical analysis because our goal is only to describe variation in our corpus, which consists of the complete population of texts written by Blair that meet our criteria.

Table 2 presents the forty-nine linguistic features and the results of the main quantitative analysis conducted for this study, including median values for all texts in the corpus (with all variables aside from average word length computed as relative frequencies measured per hundred words), real texts, and fake texts, ranked by the strength of the difference between real and fake texts, as indicated by Cliff’s delta. Table 2 also includes qualitative interpretations of the magnitude of the effect for each feature based on general thresholds, classifying variables as showing large (|d| > 0.474), medium (|d| > 0.33), small (|d| > 0.147), or negligible effects (Reference Romano, Kromrey, Coraggio and SkowronekRomano et al. 2006). Based on these criteria, we find that twenty-eight of the forty-nine features show non-negligible differences between Blair’s real and fake articles, reflecting substantial linguistic differences in how Blair constructs real and fake news. Boxplots for these twenty-eight features are presented in Figures 3 to 5, sorted by effect size, with superimposed strip plots to show the full distribution of the variables across the two sets of texts. These plots show clear differences across real and fake texts.

Figure 3 Feature boxplots (Part 1)

Figure 4 Feature boxplots (Part 2)

Figure 5 Feature boxplots (Part 3)

In the remainder of this section, we examine each of these twenty-eight features in detail from a functional perspective to understand how and why Blair’s writing style varies between real and fake news. For each feature we consider its communicative purpose in the English language in general, drawing on previous research in register analysis, and in Blair’s writing in particular, through a close analysis of its use in the corpus, supported with numerous examples. Our goal is to identify sets of features that show systematic differences, reflecting variation in Blair’s specific communicative goals and the effect of the specific communicative contexts in which he wrote real and fake news. In this way, we can provide evidence that what we observe is not just random variation, but systematic patterns in grammatical variation, as commonly identified in register analysis.

In addition, Table 2 highlights the features that show non-negligible differences when we restrict our comparison to the twenty-five real and twenty fake articles on the D.C. Sniper Attacks (marked with an asterisk). We present this additional information so as to help us assess whether the general patterns we identify remain stable when we further control for variation in topic/register. We find that fourteen of these twenty-eight features show non-negligible differences in the Sniper articles, including nine of the top ten features. We also find four additional features that show substantial differences in only the D.C. Sniper texts. Notably, in all cases, all features trend in the same direction, even if the magnitude of the difference changes. Overall, the results are therefore largely consistent, especially considering the decrease in the number of observations. We discuss these differences in greater detail as we consider the overall patterns of stylistic variation we have identified.

5.2 Information Density

Of the twenty-eight features that show non-negligible differences between Blair’s real and fake news, twenty-three are associated with variation in information density. Blair’s real news contains more frequent use of a range of features that are characteristic of denser texts, like academic writing and newspaper reporting, where authors express large amounts of detailed information in a limited space and have time to edit their texts to achieve high levels of concision (Reference Chafe and TannenChafe 1982, Reference Chafe, Olson, Torrance and Hildyard1985; Reference BiberBiber 1988). Most commonly, this involves creating long and complex noun phrases to maximise the amount of information that can be packed into sentences. Alternatively, Blair’s fake news contains fewer nouns and noun modifiers, but more features associated with an interactive and spontaneous style of discourse, including pronouns, verbs, and adverbs. This is not to say that the structure of Blair’s real and fake news differs markedly. All Blair’s articles are written in a relatively dense style as is standard for newspaper articles, making them difficult to distinguish. We find, however, that Blair tends to shift towards a slightly less dense style in his fake news. We discuss this pattern in this section, considering each of these twenty-three features individually.

5.3 Conviction

Although twenty-three of the twenty-eight linguistic features we have considered that show non-negligible differences between Blair’s real and fake news can be linked to his use of a denser style of communication when he is telling the truth, five features do not conform to this pattern. In fact, these features trend in the opposite direction: WH Relatives are more common in his fake news but involve noun modification, whereas Suasive Verbs, Possibility Modals, By-Passives, and Public Verbs are more common in his real news but involve verbs. These features are therefore highly marked, going against Blair’s general tendency to use more nominal forms in real news and more verbal forms in fake news. All five of these features, however, can be associated with stance – how a speaker or an author expresses their assessments, attitudes, and opinions towards the information they are communicating (Reference Biber and FineganBiber & Finegan 1989; Reference Gray, Biber, Hyland and GuindaGray & Biber 2012). A number of these features can also be associated more specifically with evidentiality, which is how languages mark information about the source of an utterance and the reliability of the knowledge being conveyed (Reference AikhenvaldAikhenvald 2004). In some languages this is achieved through obligatory grammatical marking, for example, encoding information about whether or not an event was witnessed directly by the speaker. In English, however, this information is optional, communicated obliquely through a wide range of grammatical, lexical, and discursive strategies (Reference Chafe, Chafe and NicholsChafe 1986; Reference Biber and FineganBiber & Finegan 1989). We find that Blair’s real and fake news differs in terms of how these strategies are deployed. In particular, in our opinion, Blair writes real news with greater conviction than fake news, adopting a more persuasive and confident stance towards both the information he is reporting and his sources. In this section, we consider Blair’s use of these five features, as well as Agentless Passive, a feature that shows substantial differences when we restrict our analysis to the articles on the Sniper Attacks. We also revisit Blair’s use of the Perfect Aspect and Downtoners, as they are directly linked to the expression of evidentiality as well as information density.

5.3.1 Real News Features

The most distinctive feature that we have not yet discussed is Suasive Verbs (e.g. allow, decide, determine, propose), which are substantially more frequent in Blair’s real news. Suasive verbs are often used to describe an action that is intended to actuate some kind of change and are therefore often associated with persuasive texts (Reference BiberBiber 1988). As a reporter, however, Blair is not generally using these words to persuade the reader of the validity of his personal views; rather, he is conveying his assessment of the likelihood of the events he is covering resulting in change. For instance, consider examples 16 and 17:

16. Law enforcement officials said Attorney General John Ashcroft would most likely decide this week which jurisdiction should proceed with the first case against the two. (R08)

17. During the first day of a hearing that will determine whether Mr. Malvo, 17, faces charges that could bring a death sentence, he heard of the pain caused by the killing with which he is charged. (R18)

In both sentences, Blair uses a suasive verb (decide, determine) to describe a future action that will have legal consequences and to express his assessment of how likely that event is to take place. Notably, in example 16, Blair specifies that this information is coming from law enforcement officials. Through his frequent use of Suasive Verbs in real news, Blair is therefore conveying a level of confidence in the information he is reporting and in his sources: he is writing with conviction.

Similarly, Prediction Modals (will, would), which are associated with the expression of stance (Reference Biber and FineganBiber & Finegan 1989; Reference Gray, Biber, Hyland and GuindaGray & Biber 2012), also tend to occur more frequently in Blair’s real news. Even more directly than suasive verbs, these forms are used to refer to possible future events and outcomes with high levels of certainty. For instance, Blair uses these modals in examples 16 and 17 preceding the two suasive verbs previously highlighted, expressing a high level of confidence in the likely outcomes of the event he is covering. This type of language is especially notable in newspaper reporting, as the goal is generally to focus on describing newsworthy events that have recently taken place in the past; discussing the likely outcomes of these events in the future can therefore be seen as a more marked communicative choice that directly imbues an article with a higher level of conviction.

Blair’s use of the passive voice is also related to the expression of conviction. Passives are used to emphasise the entity being acted upon by the verb, making it the grammatical subject of the sentence, as opposed to the agent engaging in the action. Blair uses By Passives, which include the agent in a by-phrase, substantially more often in his real news. This construction is illustrated in example 18.

18. Mr. Malvo’s court-appointed guardian, Todd G. Petit, tried to halt the questioning, but was rebuffed by both the police and prosecutors. (R28)

In this sentence, the passive voice is used to highlight that Malvo’s guardian was rebuffed, but a by-phrase is also included that identifies transparently who was doing the rebuffing (the police and prosecutors). No information is therefore lost compared to the corresponding sentence in the active voice (e.g. The police and prosecutors rebuffed Petit). The use of By Passives can therefore be associated with the transparent identification of the people involved in the news being reported and the sources of this information.

Alternatively, Blair uses Agentless Passives, which omit the agent, and are often used to obscure the agent in newspaper writing for various reasons (Reference Blanco-Gómez and JuanaBlanco-Gómez 2002), more often in his fake news – substantially more often if we restrict our analysis to the articles on the sniper attacks. This construction is illustrated in example 19.

19. In Mr. Malvo’s case, other law enforcement officials said today that even if Mr. Malvo’s admissions were thrown out by a judge who determined that he had been denied his rights to counsel and a guardian at the time he was questioned, there is enough evidence, including computer files on a laptop that was recovered when the men were arrested, to convict him of killing Ms. Franklin. (F04)

In this sentence, four passives are used in one sentence (were thrown out, had been denied, was questioned, was recovered), including only one by-passive (by the judge). The other three examples are agentless passives, and it is not entirely clear who did the denying, questioning, or recovering, substantially increasing the imprecision of this sentence. The use of Agentless Passives, as opposed to By Passives, therefore increases the amount of uncertainty in the text.

Public Verbs are also substantially more common in Blair’s real news and can be directly related to the expression of conviction. This semantic class of verbs involves actions that are observed publicly, specifically speech acts (e.g. agree, report, suggest). By far the most common public verbs used across Blair’s news are forms of the verb to say, which accounts for 76 per cent of the 1,296 tokens of public verbs across our full corpus. Public verbs are regularly used by Blair to introduce direct and indirect quotations, as illustrated in examples 20 and 21.

20. “In essence,” Mr. Horan said, “they were saying: ‘Do you want us to stop killing people? Then give us the money.’” (R19)

21. Sgt. Joseph C. Gentile, a spokesman for the Washington police, said witnesses reported seeing a burgundy Toyota leaving the scene of that shooting. (R03)

In both examples, there are two public verbs, including three tokens of say and one token of report, with three tokens in the past tense. The first example is presented as quoted speech, including a quotation within a quotation, while the second example is presented as reported speech. The use of public verbs is one way journalists attribute the information they report to specific people, effectively marking information as second-hand knowledge. If a specific source is named, this information can then be checked for accuracy, as was done for The New York Times investigation. For all these reasons, using public verbs, especially when stating the source, can be interpreted as a sign of conviction – being transparent about the status of the information being reported.

The pattern, however, is much more complex. Although public verbs are more common overall in Blair’s real news, Blair uses more quotations in his fake news, both in terms of the rate of quotation and the overall amount of text in quotation. These results seem incommensurate. We therefore compared Blair’s use of individual public verbs in his real and fake news, as well as public verbs grouped by tense and aspect. We repeated the Cliff’s delta analysis for these features (relative frequencies per hundred words) and found two opposing trends hidden by the overall public verb counts. Table 3 presents the results for the eleven features that show non-negligible differences between Blair’s real and fake news, as well as the two quotation metrics, ranked by strength of the difference. Note that means for real and fake news are reported as opposed to medians, as presented in Table 3, because a number of these features are very infrequent, resulting in medians of 0 for real and fake news. Box plots for a selection of these features are also presented in Figure 6, where medians are marked.

Table 3 Quantitative results: Public verb analysis

Feature	Mean	Mean (Real)	Mean (Fake)	Q1 (Real)	Q1 (Fake)	Q3 (Real)	Q3 (Fake)	Cliff’s Delta	Direction	Magnitude
Third Person Present Tense Public Verbs	0.08	0.01	0.13	0	0	0	0.15	0.36	Fake	medium
Says	0.06	0.01	0.1	0	0	0	0.1	0.27	Fake	small
Quotations Mark Relative Frequency	0.94	0.85	1	0.51	0.76	1.18	1.22	0.24	Fake	small
Progressive Voice Public Verbs	0.12	0.15	0.09	0	0	0.2	0.13	−0.22	Real	small
Agree (All forms)	0.02	0.04	0	0	0	0	0	−0.21	Real	small
Agreed	0.02	0.04	0	0	0	0	0	−0.21	Real	small
Past Tense Public Verbs	1.88	2	1.78	1.57	1.39	2.29	2.28	−0.19	Real	small
All Public Verbs	2.35	2.45	2.27	1.98	1.91	2.87	2.63	−0.19	Real	small
Suggest (All forms)	0.03	0.01	0.05	0	0	0	0.08	0.18	Fake	small
Said	1.47	1.56	1.4	1.25	1.01	1.98	1.9	−0.17	Real	small
Say (All forms)	1.76	1.81	1.72	1.56	1.39	2.19	2.07	−0.15	Real	small
Percentage of Text in Quotations	13.33	12.5	13.97	6.95	8.52	16.53	18.63	0.15	Fake	small
Saying	0.04	0.05	0.03	0	0	0.12	0.01	−0.15	Real	small

Figure 6 Public verb boxplots

On the one hand, the two most distinctive features are the use of public verbs in the present tense and the use of says more specifically, but surprisingly these forms are more common in Blair’s fake news, going against the general trend for public verbs. Remarkably, Blair uses says fifty-six times in his fake news, generally to introduce quotations or reported speech, but only once in his real news, where it is used to quote a report as opposed to a person. In a sense, the use of says is therefore the most distinctive pattern we have identified. This use of says is illustrated in examples 22 and 23, for both quoted and reported speech.

22. “I am not sure what’s worse, the fear or the loneliness,” Ms. Thompson says of the recent nights without her husband, Cpl. Alan Thompson, a member of a Marine artillery battalion that was among the first to enter Iraq. (F33)

23. At moments, Ms. Anguiano says, she can picture her son in an Iraqi village, like the ones she has seen on television, surrounded by animals and the Iraqi people he has befriended. (F35)

The only other non-negligible feature that is more common in Blair’s fake news is use of the verb to suggest, which seems to be broadly in line with lower levels of conviction in his fake news. On the other hand, the other eight of these features are more common in his real news, driving the overall pattern for public verbs. Most notably, these features include said, which is used in the simple past tense in the vast majority of cases, as illustrated in examples 20 and 21, and which is by far the most frequent of all the individual forms of public verbs.

The main pattern for public verbs is therefore Blair’s more frequent use of past tense forms, especially said, in his real news, and Blair’s more frequent use of present tense forms, especially says, in his fake news. In general, Blair is using these public verbs to report information he has obtained from his sources, although we know that he has often fabricated this information in his fake news. Blair therefore appears to be unwittingly encoding evidentiality into his real and fake news through his inflection of the verb to say: he uses the present tense almost exclusively in articles where he is lying. This pattern is not without motivation. Past tense appears to be the standard in newspaper writing for quotations from human sources; it certainly is his standard when writing real news. Blair’s use of present tense says is therefore inherently marked. More generally, it appears that using the present tense to introduce reported information can be used in English to express doubt. Consider, for example, the sentences he said he will come and he says he will come, where the use of the present tense in the second example can imply that there is some doubt about the validity of the statement. The fact that Blair uses the present tense to introduce quotations only in his fake news therefore appears to inadvertently signal that he has less confidence in the validity of the quotes, subtly reflecting a lower level of conviction.

Overall, we can therefore see how Blair’s use of a range of different grammatical features marks stance in his real news. The use of both suasive verbs and prediction modals reflects Blair’s confidence that what he is covering is not only accurate but consequential, while his use of by-passives and public verbs allows Blair to explicitly identify and attribute information to his sources, which is often crucial information for establishing the validity of reporting. We believe that both these patterns, which are deployed substantially less often by Blair in his fake news, result in his real news conveying higher levels of conviction.