Editor’s Note: We’re putting this note outside of our subscription paywall, because we want as many people as possible to learn about our vision – to inoculate the world against the weaponized narratives of Big Tech, Big Media and Big Politics – and the concrete steps we are taking to achieve that goal. If you share those goals and would like to support our efforts … join us!
Several years ago, we introduced the concept of fiat news on these pages.
It is a simple idea. In the same way that money created by fiat debases real money, news created by fiat debases real news.
Although it misinforms, fiat news should not be understood as misinformation, at least in the colloquial sense. News which contains false information or distorted interpretations of facts can be better thought of as counterfeit news. Like counterfeit money, enough counterfeit news can debase the real thing, too. Yet even considering how widespread counterfeit news has become, fiat news exists on such a massive scale that its power to debase is in a different category. Nor is fiat news synonymous with bias. We think bias represents a causal explanation for a very specific kind of recurring fiat news.
Most “media watchdogs” are in the business of identifying one of those two things: misinformation or bias. The problem with these efforts, beyond that they do not capture the full scope of actions which debase the information content of news, is that it is practically impossible to report on misinformation or bias in a manner that is itself not colored by the opinions of the author, or else designed to shape how the reader interprets facts and events. While they may in some cases offer a useful service in the face of blatant lies published through politically invested news outlets, too often they become yet another source of fiat news.
Why? Because fiat news is the presentation of opinion as fact. Fiat news is news which is designed not to provide information for the reader to process, but to provide interpretations of information for the reader to adopt. Fiat news is the primary vector for nudging, shaping common knowledge, or what everybody knows everybody knows. Fiat news is how governments, parties, corporations and other institutions in a free and always connected society meticulously shape it – then tell us that it was our idea.
By design, fiat news isn’t always easy to spot. Outside of editorial pages, it is rare indeed that an expression of opinion as fact would include obvious phrases like “we believe” or “we think.” Instead, media outlets guide interpretations through more subtle means that may sometimes be as invisible to the author and editor as they are to the reader.
Several years ago, we also introduced what we called the narrative machine.
It is also a simple idea. We think that recurring patterns in language make it possible to identify narratives. We also think they make it possible to identify similar patterns indicative of the various types of fiat news. We think a revolution in mainstream natural language processing (NLP) software and techniques has made this feasible at a high level of detail. We think a revolution in the availability of low-cost compute power has made this feasible at scale.
Finally, some months ago we made a couple of quiet announcements. The first was that we would be devoting time and capital to develop tools to inoculate citizens against weaponized narrative, especially from media sources (in case you were wondering where I have been the last 6-7 months). Ben referred to the goal of this effort as the Narrative Early Warning System, or NEWS. The second was the announcement of an advisory partnership with Vanderbilt University, who through a major gift by long-time Epsilon Theory pack members Suzanne and Patrick McGee established the McGee Applied Research Center for Narrative Studies there.
This research is the first step toward our goals for both of these things. We are working to develop applications that will inform citizens about the fiat news content of their media consumption. Our plan is that these applications will do so in both real-time and, if users elect to make this data available, over longer periods of time.
We are also excited to work with the students and faculty attached to the Applied Research Center for Narrative Studies to find flaws with these methods and uncover better ones. Ultimately, we think that it is important that criticism of the current role of media be leavened with a passion for the indispensable function of the fourth estate in a democracy. We look forward to mutual accountability with our partners.
The early basis for both that research engagement and our NEWS application development will be the framework for identifying fiat news that we have been scaffolding in earnest over the last 18-24 months. Over the next several months, we will begin regularly publishing our insights about how topics, outlets and the general media environment are infected by attempts to tell you how to think about the news. In the months and years thereafter, we hope to be in a position to deliver tools to citizens that can be directed at any information consumed – not just what we decide to write about here at Epsilon Theory.
Allow me to be the first to introduce you to the Fifteen Faces of Fiat News.
|Appeals to Authority||Assumptions of Causality||Bogeymen|
|Confidence and Doubt||Content Context||Coverage Selection|
|Generalized Attribution||Interpretive Language||Missionary Statements|
|Missionary Warfare||Question Begging||Response Coverage|
|Rhetorical Questioning||Superlative Language||Unsourced Attribution|
Appeals to Authority
Now, there are a great many ways to tell the reader how they should interpret the facts of a story or event without actually telling them. Few are more effective and probably none more difficult to parse the ethical considerations of than telling them ‘here are the facts, and also you should know that someone smarter and more authoritative than you has this opinion about them‘.
In a wide range of news stories, of course, it is the fundamental role of the responsible journalist to provide supplemental facts from credible “expert” sources. It would be practically impossible to present useful information about the James Webb Space Telescope, Large Hadron Collider or new BA.5 variant of Covid-19, for example, without information provided by scientific experts. The same goes for an especially wonkish article about an upcoming piece of legislation, a new plan to build a light rail system in a metropolitan area, or a modified K-12 curriculum standard coming before state education commissioners. This is not what we mean by fiat news, and this is not what we have designed our model to capture.
In some cases, the information being provided by an individual of authority or influence may not necessarily lean upon their authority, influence or expertise, but be included simply because they are the primary source of the information. Think something like a chief financial officer reporting earnings or a chief executive officer discussing the company’s growth plans. The news is the thing they are saying, and in some cases the news is that they are saying it. This is also not what we mean by fiat news, and this is not what we have designed our model to capture.
To be fair, even though it is obviously sensible to reference authoritative sources for either of the reasons above, there remains a clear opportunity for the author to inject opinion. The journalist often has the power to select the authority, many of whom you may be surprised to learn do not always agree with one another. If you require exaggerated examples, imagine the expert on CRT selected for a segment by Fox News and the expert on Christian Nationalism selected by the New York Times for a feature piece. Beyond the opportunities to select the expert, the journalist may also select the specific statements used. These opportunities for the injection of opinion are very real. Based on our research to-date, we also think that these techniques are nearly always revealed by other faces of fiat news. Accordingly, our approach to seeding models of linguistic patterns associated with Appeals to Authority accepts that we may underreport injected opinion (i.e. false negatives). This will be a recurring theme.
Instead, our aim is more parsimonious. NEWS seeks to identify linguistically three recurring mechanisms for injecting opinion: (1) appeals to authority which are both generic and explicitly confidence-shaping by nature (e.g. “experts agree”), (2) author summarization of the consensus of authorities and (3) author summarization of the aggregated implications of “research” or “analysis” on a topic. In other words, our language model is designed to capture not experts or authorities reporting facts or their opinions, but rather patterns that would generally only arise in the course of an attempt by the journalist to dampen or emphasize a particular interpretation by the reader of the facts presented in the article.
Assumptions of Causality
For those in the financial industry – our largest universe of readers – this method for the injection of opinion would take all of ten seconds to explain. You see, on a daily basis, the content published by practically every financial ‘news’ outlet is dominated by explanations for the movements in the prices of stocks, indexes and markets more broadly. I don’t think it’s taken yet, so call it Rusty’s Law: For any event above a critical threshold of significance, there will be at least one ‘news’ article describing the stock market as declining/rising as a result of that event.
This is pure fiat news, and among the most easily manipulated by those with an interest in guiding how financial market events are perceived and framed. Thankfully, once a reader begins paying attention to it, it is also among the easiest faces of fiat news to detect.
However, Assumptions of Causality exist in all forms of news. Not all of them are so easy to detect as in news about markets. That’s because other social spheres generally lack the kind of daily scoreboard for which every news consumer demands an explanation and for which no one has one that could even vaguely be described as factual. But constructions suggesting implicitly or explicitly that ‘because X, then Y’ are widespread. They serve a useful function for narrative creation in that they connect facts and events to powerful memetic forces, that is, to other things that people care deeply, even innately about.
Assigning causal relationships can be used to malign or whitewash the intent of actors in events. It can be used to establish a logical chain intended to guide a reader to a desired conclusion or opinion without explicitly stating that conclusion. It can be used to amplify or relax the perceived importance of a topic or event.
As with all of these language models, there will nearly always be a baseline of causal assumptions, many of which are benign. Abnormal density of this language at a point in time, around a certain topic, or from a particular media outlet, therefore, becomes the operative focus of the analysis.
The Bogeyman face of fiat news means the knowing use of linguistic patterns in reference to institutions and individuals that everybody knows everybody knows have become pejorative. That does not mean that we want to flag every simple mention of “George Soros” or “the Koch Brothers.” These people do a lot of things, and many of those things are newsworthy.
And yet, if a journalist elects to violate Godwin’s law and make a comparison of some event or person to the Nazi Party and/or Hitler, more often than not we feel comfortable judging that to be an injection of opinion, even if the underlying comparison (e.g. “They were both failed aspiring artists!”) is technically factual. While this is an extreme example, there are hundreds of milder examples which incorporate varying degrees of negative affect. There are no ways to reference “shadowy cabals” or “corrupt politicians” that do not convey the author’s intent to affect the state of mind of the reader as they consume the information, friends.
Bear in mind with this category (and all others, frankly) that we aren’t deeply concerned about having some baseline result from our model that indicates the presence of fiat news even when the examples might seem to some readers to be innocent. It isn’t a grade intended to be applied to a single news article in isolation in which less is always better. Rather, it is intended to draw attention to changes in aggregate levels over time, comparative levels across outlets, comparative levels across topics, events and individuals references, and changes in levels experienced on the part of an individual reader’s news consumption habits.
Confidence and Doubt
Journalists have incredible breadth of options when it comes to affecting the reader’s perception of facts and events. Explicit and implicit assessments of confidence and doubt when presenting those facts or descriptions of events are among the most effective – and common.
Many of these pass by our eyes without notice. Yet news articles describe the implications of events as clear and obvious with regularity. Future consequences of events are described as probable, likely, unlikely, inevitable without hesitation. Facts are widely acknowledged and states of the world are widely understood. Should the journalist suffer the misfortune of a hawkeyed editor, then there is still the option for a strategically timed flip into the subjunctive.
As with the other faces, there will always be a baseline of innocent usage that will trigger any language model used to identify Confidence and Doubt. But abnormal rates of attempts to instill confidence in a fact, quoted speaker or conclusion on the one hand, and doubt in a fact, quoted speaker or conclusion on the other, absolutely debase the value of news.
We haven’t figured out how to model this one yet. We think we will, but it’s a tough nut to crack.
We are still including it here because we want to present a full picture of how fiat news happens, and that’s a story that cannot be told without Content Context. We’re also including it because it’s important enough that we’d get half a dozen comments telling us that we missed something important if we didn’t.
By Content Context we mean the intentional or negligent graying of lines between news and opinion content. In online and print media, this happens in many ways. Outlets publish a news piece but distribute it on social media with an opinion-loaded tweet. They publish a nominally news piece but attach an opinion-loaded, click-bait headline. They physically nestle opinion and “analysis” articles among news pieces covering a similar topic. They neglect to flag opinion and analysis articles in headers, sub-headers, categories or meta-data. They treat “explainers” as a form of news with no disclosure. They treat “feature” pieces, which are typically just long-form opinion pieces that feel like reportage, as a form of news with no disclosure.
Language models are no good for categorizing this face of fiat news, since some news content is so rife with language indicative of opinions that identifying an unidentified opinion piece and attempting to distinguish it algorithmically from news content is often…shall we say, problematic. Likewise, there is a temporal element to how outlets massage the perception of news through Content Context that is difficult to capture. That is, on the average news webpage or social media account, placement changes. Tweets are deleted. Headlines are modified, almost always without disclosures that would be present for a similar change in the article’s body text.
It is disappointing, of course, not to be able to model this behavior, not least because outlets use this technique to such great effect. As we continue to explore a solution, we are open to ideas.
Like Content Context, Coverage Selection is not so much a problem to be solved by the development of language models. Unlike Content Context, however, we think we have more than adequate tools to identify and measure Coverage Selection.
The idea here is simple. Media outlets influence how readers think about the news by choosing what they deem newsworthy enough to cover, and what they do not. Newsworthiness is unquestionably an opinion, if an unavoidable and necessary one. Yet it is an opinion that manifests less at the micro level of the individual article, and more at the meta layer of the overall landscape of news coverage.
Volume of coverage on topics, both at an outlet level and aggregated across news sources, informs perceptions of common knowledge, what people believe everyone else believes. High volume reinforces confidence in the underlying contentions, especially if they align. Low volume reinforces doubt in the underlying contentions, whether in their veracity or in their importance.
Unlike our other measures, there can be no baseline of a topic’s coverage volume. That is, it would be dishonest to claim the ability to say what the correct amount of coverage is, and then to measure deviations from that. What we can do is identify an outlet-level relative measure (i.e. how much more or less coverage is an outlet devoting to a topic than peer outlets) and an aggregate-level measure of multi-polarity in perceptions of newsworthiness (i.e. how much the volume of coverage differs, on average, among outlets). The former measure may be but isn’t always a sign of the dreaded “bias”, since some outlets naturally report on topics more than others. The New York Post is going to cover Yankees games more than the Miami Herald. The Hill will cover the minutiae of Capitol Hill goings-on more than Bloomberg News.
As a result of this, the outlet-level measure is a thing we consider independent of our language-model driven measures for the other faces of fiat news. The aggregate measure, too, while less prone to harmless false positives, is best as a standalone assessment of divisive perceptions of newsworthiness and how that might be affecting the conclusions of citizens with narrow media exposure (i.e. basically everyone).
A face of fiat news that journalism schools and style guides have been trying (unsuccessfully, I might add) to eradicate for decades, Generalized Attribution is the simple trick of attributing ideas, statements and assessments to conveniently amorphous entities. In 2022, major national publications, websites and news magazines still attribute all manner of judgments to “some”, “many” or “most” of…us? Americans? Citizens? It isn’t always clear.
And that’s the point.
What is clear is that, beyond being lazy, the Generalized Attribution technique can be used by the enterprising narrative weaponizer to great effect to establish the idea of common knowledge. It creates in the mind the reader that a deviating perception would make them an outlier.
Probably the second most on-the-nose version of what we mean by fiat news, Interpretive Language simply means the explicit language patterns you might associate with someone explaining the implications of a fact that they just provided or an event that they just described.
Interpretive Language is a more expansive face of fiat news than Assumption of Causality in that it does not apply narrowly to direct cause-and-effect relationships in the real world but to cause-and-effect relationships in the logical sense. That is, Interpretive Language tells you here’s why and here’s how.
Still, like Assumption of Causality and Missionary Statements (below), when Interpretive Language is present in a piece that is nominally news, it is nearly always an opinion-laden attempt by the author to shape how the reader thinks about the implications of facts and events.
Missionary Statements are the damn-the-torpedoes variant of both Interpretive Language and Assumptions of Causality. They do not imply that intelligent people would process the facts in a particular way. They do not imply a chain of logic that ought to be followed to come to the correct conclusion. Missionary Statements are explicit statements telling the reader this is how you should think about the facts and events they just heard about.
If it seems to you an odd thing to exist outside of explicit opinion articles, op-eds and letters to the editor, you would be correct. If you think that would make them rare in outright news coverage, you would be incorrect. In fact, over the last 15 years or so, the merger of explainer content with news content within our dataset has been inexorable.
If you were wondering, we call them Missionary Statements because of their relationship to one of the seminal stories of the intersection of game theory and information theory – the Green-Eyed Tribe. Epsilon Theory’s application of this story is nearly ten years old at this point, but still holds up.
A face of fiat news that has largely emerged in tandem with the development of a high-peaked, bi-modal political distribution in the United States, Missionary Warfare is the reference to the reportage of other media outlets for reasons other than sourcing. If you wanted to picture what our language model is trying to identify, you could do worse than imagining a Newsmax segment gleefully describing some dumb thing the unrepentant Marxists over at MSNBC said, or a condescending, tongue-in-cheek bit from a New York Times reporter about how the ignorant boobs at Fox News were reporting on a controversial topic.
Like Response Coverage (below), Missionary Warfare is fundamentally about news that treats the response to news as news. In other words, we believe that non-sourcing references to other outlets that treat their coverage as newsworthy in itself are generally likely to represent attempts to dampen the effect of, cast doubt on or effectively argue against alternative interpretations of facts and events. All of this would be perfectly acceptable, if a bit tacky, on op-ed pages. It’s the presence of this activity on news pages that hits our fiat news radar.
Capturing all question begging taking place in news articles – that is, statements which assume identifying a feature of a premise demonstrates the premise – would be extraordinarily complicated, especially doing systematically with an algorithm. You could argue some degree of circular logic or at least unproven premises in almost any statement of fact, especially those provided in a truncated fashion by design, as they are in most news content.
But the assumption of an undemonstrated premise is a fundamental feature of many kinds of fiat news. The Question Begging face of fiat news doesn’t present facts alone. It doesn’t even present facts and a suggested interpretation in order to start a chain of logic in the mind of the reader. It treats a fundamental premise as self-evident (and thus, not an opinion in need of being excised from news content), and proceeds from there.
We do think, however, that we have built a language model capable of capturing some of the most egregious examples while minimizing the risk of false positives. Make no mistake, this is our weakest model, with a very high rate of false negatives as a result of our parsimony and aversion to innocuous false positives; however, as with other faces of fiat news, we observe a measurable historical baseline, around which certain controversial topics and themes consistently demonstrate a measurable abnormal impulse of language indicative of Question Begging.
The generalized cousin of Missionary Warfare, Response Coverage is the face of fiat news that exists in the cherry picking of “responses” to events – largely from social media. In the most benign cases, it treats those responses as sources in an article about another topic. In the more egregious cases, those responses are the news being covered.
The risk of abuse of Response Coverage manifests in multiple ways. In all of them, however, the objective, whether intentional or unintentional, is to frame common knowledge, to denigrate Bad Interpretations and to celebrate the internet points scored by people with Good Interpretations. We simply think that any news value that may exist in a citizen’s or pundit’s activity on, say, Instagram will nearly always be dwarfed by what it is that induced the journalist to select that one to be their chosen source or, God forbid, the core topic of their piece.
Some public figures use social media as a primary mechanism for conveying official information and statements. For that reason, our model does not presume any social media reference or embeds are fiat news. They may represent an entirely appropriate way to reflect a public statement, even if the trend is somewhat lazy and off-putting to some readers. Our language model seeks to identify generalized references and characterizations of the responses on major social media platforms, alongside the major linguistic patterns that tend to accompany the “laundry list” model of including selected Good Interpretations or Bad Interpretations of news events from Facebook or Twitter.
Journalists and news outlets are supposed to ask questions. It’s what they do. There’s nothing wrong with that.
However, some journalists seem to lament that no one seems to want to ask them questions. After all, they are informed. They’ve done the research. It is a shame, and one empathizes. But the solution found by many of those journalists – to rhetorically ask a question in an article that they then go on to answer – is a prime entry point for fiat news.
There was a daily email some years ago – it might even still be around – published by the Wall Street Journal. It was a sort of news digest with a mild editorial bent. One of the best recurring segments was called, “Questions Nobody is Asking.” All it did, every day, is identify hilarious versions of the Rhetorical Questioning face of fiat news in the wild. And every day, they found three or four. And they were only looking for the most ridiculous examples, nearly all of which were found only in headlines.
The problem is much more widespread when you begin to examine the body text and don’t constrain yourself to the ridiculous or hilarious. When a journalist asks themselves a question in a news article, it suggests to the reader this is the kind of question you should be asking. It implies the existence of common knowledge, of this being the kind of question the journalist has become aware their readers are concerned about. Like other faces of fiat news which seek to establish in the mind of the reader a framing of the opinions of the masses, of experts and of Those Bad People who consume right-wing media/left-wing media, Rhetorical Questioning is nearly always a technique not for presenting facts, but for framing how those facts should be interpreted and to what inevitable conclusions they should lead the reader.
We use the term superlative somewhat loosely in our name for this face of fiat news. We don’t only mean literal superlatives. We mean heavy adverb usage, higher order adjectives and other loaded phraseology. And yes, literal superlatives.
Unlike some of our other language models which require a bit more sophistication, this is brute force simplicity. There are words which, if used in a news article, are almost instantly fiat news on their own merits, regardless of context. Tremendous, wondrous, marvelous, atrocious, abominable, unspeakable things exist, but such descriptions simply cannot be made in a news article without what we think is brazen intent on the part of the author to color how the reader thinks about the facts reported.
This is all true with the exception of people quoted, of course. This is as good a time as any to note another principle of our project: we exclude quotations from our analysis. Again, parsimony over preciousness about false negatives. Yes, the content of quotes used in news articles is an avenue for fiat news. Yes, the manipulation of those quotes is common; however, in all of the iterations of our analysis, we nearly always found that the false positive rate contributed by including quote sources in the analyzed text was unacceptably high.
A quoted individual describing a congressman’s behavior as reprehensible may be legitimate news. A journalist doing the same with their own words in a news article is practically always fiat news.
There is a time and place for anonymous sources. Maybe even vague references to “sources” who “said” something, without an explicit explanation for the need for their anonymity. We aren’t here to make the argument that protecting the identify of a sensitive source is never appropriate; indeed, in order to produce institutional change of the variety we often champion here at Epsilon Theory, it may at times be indispensable.
For this reason, perhaps more than any other face of fiat news, Unsourced Attribution must nearly always be thought of against some baseline. Its presence is not a Bad Thing. We have little interest in grading a single article as “Bad! Fiat News!” because it includes unsourced attribution. We have a great deal of interest in making consumers of news aware that a topic has reached an abnormally high density of unsourced attribution. We have a great deal of interest in making consumers of news aware that an outlet has taken to a steadily increasing diet of unsourced attribution in their news content. We have a great deal of interest in making consumers of news aware that the media in general have adopted a more aggressive posture with respect to unsourced attribution.
We think the skeptical news reader, who knows that journalistic integrity is widespread but not uniform, will see the potential for sharp changes in the use of such sources to shape the perception of facts and events.
Why am I reading this now?
Those who are familiar with our work on fiat news know that our most common advice when consuming news is to ask the question: “Why am I reading this now?” So why are you reading this now?
Because we think the prevalence of fiat news is rising rapidly.
Because fiat news seems to rise sharply during election cycles.
Because fiat news seems to rise sharply during uncertain events (e.g. the immediate post-COVID period registered the highest level of fiat news in the historical period reviewed in our analysis).
Because we think narrative is being weaponized.
Consider the below nearly 15-year analysis of one of our broadest news datasets. After peaking at twice the levels of 2007 during the COVID fiat news frenzy, our measure of Fiat News density today is still at 170% of those early levels, with a steady upward trend over most of that period.
You’re also reading this now because we think we’re now at a crossover point in our capacity to analyze this credibly and offer useful ways to think about and improve news consumption.
Over the next several months, we plan to tell you more about our project. We’ll discuss our milestones. We’ll ask for feedback on tools and features you would want to see. We’ll listen to feedback. We’ll discuss our datasets. We’ll discuss new datasets we want to acquire. On some of those, we may even ask for your help.
Over the next several years, I think we can work together to inoculate ourselves against weaponized narrative.