A Reasoned Approach to Dealing With Fake News

Memory Biases

One common source of bias is the human memory system. Research shows that memory does not work like a video recording of our life. Although we may briefly have a verbatim representation of the exact words of a text, these details are quickly lost without attention-demanding retention strategies, such as rehearsal (Sachs, 1967). We use our verbatim representation to create an interpretation of the author’s message by making connections to our prior knowledge and beliefs (Kintsch, 1998). In addition to representing the content, we can also represent the source of the communication (e.g., Alex Jones of InfoWars) and associate content with a source and even identify relationships among sources (e.g., Alex Jones and @DavidGoldbergNY agree; while Reddit owners disagree) (Perfetti, Rouet, & Britt, 1999). In this way, we can represent otherwise inconsistent information across texts or webpages.

The nature and extent of this processing will depend on why a person is reading. People read information on the web for a variety of reasons, including relieving boredom, keeping up-to-date on something, looking up something to do some task. In general, we read for a purpose or goal (Britt, Rouet, & Durik, 2018), and these goals can impact what we read (e.g., CNN, FOX, Facebook posts), how we read (e.g., read only headlines and leads, read deeply, skim, skip until locate a piece of information), how long we read, and so forth. We do not simply represent only what the text says. We interpret the text, guided by our goals and based on prior knowledge as well as our attitudes and beliefs (Albarracin & Shavitt, 2018; Falk & Scholz, 2018).

One set of biases originate with the quick loss of verbatim memory and potential overreliance on an interpreted representation of the communication. For example, in one study, college students who were asked to read short, two-sentence arguments (e.g., “The U.S. is right to intervene in other countries’ affairs because local events can catastrophically impact the entire world.”) were only about 75% accurate at recalling the claim of the argument immediately after judging their agreement with it (Britt, Kurby, Dandotkar, & Wolfe, 2008). Furthermore, the errors they made most often changed the meaning of the claim, although it may seem minor to the untrained eye (e.g., “The U.S. should intervene in other countries’ affairs”).

The problem of relying on an interpreted representation is illustrated in a well-known series of studies showing false memories can be created based on the situation at retrieval (false memory effect). Two groups of people watched a video of a car accident, and one group was asked how fast the car was going when it “smashed” into the other car. The other was asked how fast it was going when it “bumped” into another car (Loftus, 1979). People falsely remembered the car going faster in the first case, and they were very certain that they had an accurate memory. This false memory effect can be reduced (but not eliminated) when a “false memory warning” is given prior to encoding (Roediger, 1996). But, when the warning is given after encoding, people are still likely to “remember” the false information and use it as if were true (Roediger, 1996). The effect is amplified the more often we encounter and retrieve the information (Roediger, Jacoby, & McDermott, 1996).

People are also highly affected by the mere repetition of information. Researchers have found that statements presented several times across occasions, led to more confidence in the truth of those statements as compared with statements that were not repeated (illusory truth effect) (Polage, 2012). A related effect, availability cascade effect, shows how this illusory truth effect can snowball at a more global level. Kuran and Sunstein (1999) presented an analysis of the public attention and attitude toward three events (waste dumps in the Love Canal, the use of Alar chemical, and the TWA 800 crash) to show how repeated statements of beliefs can make claims seem truer and can be used to manipulate public opinion. It would seem inevitable that in an information environment such as the Internet, social media algorithms that identify and highlight “trending” stories and source filters that restrict alternative information would produce an availability cascade.

Finally, our memory system does not handle new discrepant information by simply replacing old information. That is, initially faulty or wrong information remains available and can continue to have an effect despite encoding new, correct information. This is called the continued influence effect (Lewandowsky, Ecker, Seifert, Schwarz, & Cook, 2012). For example, when people read a story about a warehouse fire that was initially reported to be caused by oil paint cans but later discounted, readers continued to think that the fire was caused by the paint cans (Johnson & Seifert, 1994). If you think of human memory like a collection of assertions networked together by associative links of varying strength, updating information is more akin to creating a new link that tags the information as wrong but doesn’t remove it. Nevertheless, such a tag might be limited in scope, applying to one specific assertion but not to all the other information in memory connected to that assertion. In this case, the updated information will not necessarily be stored with the later-found-erroneous information without systematic updating all of the information learned. This continued influence of misinformation effect is difficult to overcome (Lewandowsky et al., 2012) and can even backfire, as discussed below.

Prior Belief Biases

In addition to these memory related problems, several biases relate to those assertions categorized as beliefs or attitudes. In general, these biases affect what information we choose to attend to and how we process it, usually in the direction of maintaining the consistency of our attitudes and beliefs. Our beliefs can lead to biased search, selection, and interpretation—thereby reducing exposure to information that is not consistent with what we already believe.

We are more willing to accept consistent information and judge the quality of arguments more positively when they are consistent with our beliefs, as compared with information that opposes our beliefs (belief bias; Lord, Ross, & Lepper, 1979). The other half of the belief bias effect is that we are also less likely to accept information that goes against our beliefs and judge the arguments as being of a lower logical quality (Edwards & Smith, 1996). We are more likely to scrutinize or attempt to find holes in the information that contradicts our beliefs. In fact, in some cases, when we encounter disconfirming evidence, it may actually strengthen our evidence-opposing previous beliefs (backfire effect) or reject information because of the source it came from. All of these biases are exacerbated by our tendency to believe we are more unbiased reasoners than others (bias blind spot; Pronin, Lin, & Ross, 2002).

We also have a strong tendency to seek out and use information that is in line with what we already believe to be true (confirmation bias; Nickerson, 1998); that is, we seek out information that confirms our beliefs rather than seeking information that could potentially disconfirm them. We also have a myside bias toward generating belief-consistent reasons (Baron, 1995) and ignore evidence counter to our own side when performing tasks such as writing an argumentative essay (Wolfe & Britt, 2008). Finally, beliefs even affect comprehension: readers spend longer rereading text that is contrary to their beliefs (Maier, Richter, & Britt, 2018) but have better memory for information that is consistent with them (Maier & Richter, 2013).

These biases are challenging to overcome because the Internet is being designed to create “filter bubbles” which as Bill Gates notes, let “you go off with like-minded people, so you’re not mixing and sharing and understanding other points of view. It’s super important. It’s turned out to be more of a problem than I, or many others, would have expected” (Delaney, 2017, para. 5).

Assessing the Accuracy of Information

People typically assume what others say is truthful and accurate unless there is reason for doubt (Grice, 1975). However, our cognitive biases work against accuracy evaluation especially when we rely on our memory, as the previous section reviewed.

Accurate evaluation requires readers to set and monitor accuracy as a goal, employ strategies to achieve that goal, and value the time and effort-consuming systematic evaluation. Often when reading on the web, one has many competing goals; obtaining and verifying that the information is accurate is only one goal. Monitoring for accuracy may not have been a well-practiced cognitive activity in the past, but it must become one today. Now more than ever, readers have to assess the accuracy of information that can range from clearly true or false to indeterminable. Indeed, the amount of falsehood and deception on the web has given rise to a plethora of fact-checking sites (e.g., https://www.snopes.com/, http://www.politifact.com/, and https://www.factcheck.org/), and even Google and Facebook have begun efforts to label suspect stories.

Strategies for evaluating accuracy can target the content itself. Readers can actively question whether they find the information believable, whether it makes sense or is consistent with prior knowledge (Richter, Schroeder, & Wöhrmann, 2009). We have to monitor and protect against our belief bias because related beliefs are quickly available (Voss, Fincher-Kiefer, Wiley, & Silfies, 1993). This takes a deliberate effort of setting a goal to acquire accurate information. The content can also be evaluated for accuracy based on an assessment of the support provided. Some types of evidence are better (e.g., experiments, quotes, reasoning) than others (e.g., number of retweets). The evaluation of what makes “good” support, however, depends on several factors: (a) one’s epistemic beliefs about what knowledge is (e.g., tentative and complex), and (b) beliefs about how we get knowledge (e.g., gut feelings vs. authority vs. use of rules of inquiry) (Bråten, Britt, Strømsø, & Rouet, 2011) and (c) other knowledge, skills, and dispositions that lay readers often lack. Even undergraduate college students are not very skilled in detecting problems in descriptions of scientific studies (Kopp, Britt, Millis, & Graesser, 2012) and do not spontaneously set an explicit goal to obtain accurate information when reading search-result listings of global warming sources (Kopp, 2013). Instead, they are “looking for support,” though not necessarily accurate support.

Readers can also evaluate features of the source of the information. Given that we cannot erase or overwrite memory, evaluating the source information would be most beneficial prior to encoding the content so that the reader can decide whether the source is reliable and knowledgeable enough for the reader to continue.

To illustrate, searching a news aggregator site for “Betsy DeVos” returned links to very different sites (e.g., “The Washington Post,” “Politico,” and “Reason.com”) that vary in terms of features of the author (e.g., author’s level of knowledge, motives, or bias) and outlet (e.g., criteria for accuracy checking, checking tools employed). This is the easy case if one sets the goal to encode and evaluate the source (which is not often a reading goal). The reader also has to be able to understand that when a site uses a headline like “DeVos: Civil rights office will return to being a neutral agency,” it is stating that DeVos is saying “Civil rights office will return to being a neutral agency.” DeVos becomes the source of the statement, and the reader’s knowledge of and evaluation of DeVos can be used to interpret it. In this case, the source precedes the content which helps. But in many cases, the source of a statement comes after the content has been encoded, which is a problem because it is now in our memory before we know whether we trust the source. An even more challenging example is when all that is stated to entice a click is “Betsy DeVos isn’t ‘Enabling rape deniers’ by pushing for due process on college . . . ” Without clicking, it is not possible to know who is saying that Betsy DeVos is “enabling rape deniers.” This is challenging, because reading headlines without clicking to read the details is very common—leaving the reader with a memory of an assertion without being able to evaluate whether the source was knowledgeable or biased.

In the case of a news aggregator, some source information is actually presented. Many times, however, there is no information about key features of the author or outlet responsible to distribution. Even worse, the source information can actually be deceptive. For example, “ABCnews.com.co” or “Breaking-CNN.com” are sites that mimic the look, feel, and logo of actual news sites but may present erroneous information and malware. When we talk with people in real life, their “sourceness” is obvious. On the Internet, the author as an entity can be obscured or even falsified (Britt, Rouet, & Braasch, 2013).

Finally, readers can look to corroborate information across independent sources. However, it can be challenging to find truly independent sources. For example, it may be that only one source came to the conclusion that 2 to 3 million illegal immigrants voted in the U.S. election, but one could have encountered it on Facebook, in discussion with friends and family, and on the television news. This could appear as corroboration, but it is not, because the information originally came from a single source.

Assessing the Relevance of Support and Search Results

Relevance of information can be evaluated according to goals for search (e.g., Does this page, article, or even paragraph give me the information that I am looking for?) or coherence (e.g., Does the support provided for their claim actual provide support? or Does the effect follow from the cause?).

For brevity sake, consider judging claim-support relevance. Of course, persuasion techniques, propaganda, and logical fallacies (e.g., ad hominem, emotional appeals) are very prevalent on the web and can be persuasive, but they are beyond the scope of this article. Here we focus on an appeal to reason.

Again, our cognitive system can work against the use of accurate reasoning. Relying on an interpreted representation makes it difficult to accurately evaluate whether the support is relevant for the specific claim made (we discuss this in more detail below). Indeed, readers may set a goal to identify support but not necessarily “good” support (Kopp, 2013), and this may lead to accepting all support as “good” without much evaluation. As a result, a wider range of support may be seen as acceptable. As noted, our beliefs can also interfere. We spend longer reading belief-inconsistent information to scrutinize it and often judge those arguments as of lower quality, whereas we read belief-consistent information more quickly and typically judge it to be of higher quality. Judging the quality of an argument, independent of one’s acceptance of its truth, is difficult.

Even with arguments for which one does not have strong beliefs, evaluation can be challenging. College students experience difficulty evaluating arguments that are structurally flawed (i.e., the reason failed to support the claim or there was no reason presented) as compared with those that are structurally acceptable (i.e., the reason supported the claim; Larson, Britt, & Kurby, 2009). It is even more challenging to evaluate argument elements that are spread across time or an expanse of text. Doing so requires one to hold in working memory the precise claim, while trying to detect the support, sometimes needing to go back and reread the precise claim. This type of monitoring takes effort. Readers have difficulty noticing discrepancies in even short texts (Otero & Kintsch, 1992), although they are slightly better when the discrepant statements come from different sources (Braasch, Rouet, Vibert, & Britt, 2012; Stadtler, Scharrer, Brummernhenrich, & Bromme, 2013). Thus, lay readers may find it challenging to monitor and evaluate the logical claim–reason connection.

Assessing the Sufficiency of Support and Perspectives

A final criterion for argument support is whether it is sufficient, that is, complete to some degree and capable of justifying the claim. Professional writers and news reports typically consider more than a single perspective (Wolfe & Britt, 2008), and it is the hallmark of academic writing and the scientific method. For the rest of us, comparing perspectives can help us understand the broader situation and make an informed decision.

Biases are most prevalent for this criterion of support. As the belief bias shows, we are less likely to believe things that are inconsistent with our views. We also have poorer memory for belief-inconsistent information. College students tend to ignore other-side information when composing argumentative essays (Nussbaum, Kardash, & Graham, 2005; Wolfe, Britt, & Butler, 2009) and they often do not make alternative-based objections when evaluating arguments without task supports (Shaw, 1996). Readers often do not seek out alternative perspectives or disconfirming evidence, but when available, they do read some of it, although generally do not include it in their arguments (Wolfe et al., 2009).

As with the other two evaluations, working against the myside bias and confirmation bias takes effort and requires that readers set specific goals to seek out, comprehend, and consider other-side information. In its simplest form, it means trying to find articles on the web that argue for the other side of the claim. It can also mean trying to find and account for all available evidence, regardless of side of the claim it supports. It can also mean taking into account multiple sides or perspectives and therefore include or address the “alternative viewpoints” or even a range of political camps (e.g., fiscal conservative, liberals, social conservatives, and moderates). This is challenging when only one side is presented because it requires the search for alternative perspectives without knowing what those perspectives are or how to find out about them. It is also challenging with multiple texts (for instance a series of chat messages or Google page lists) because it requires reading more, comparing information, and possibly dealing with the emotions that come from reading perspectives outside one’s own. It is not something even college students take on readily (Kopp, 2013).

Epistemic Vigilance Needs to Be Taught and Not Just in School

In the world of InfoWars, the public needs to acquire the intellectual skills needed to critically assess the accuracy, soundness, and sufficiency of information, in other words, epistemic vigilance. This review points to the insufficiency of teaching simply rules (e.g., use “.gov” sites, look for an M.D.). Rules do not help because the principle behind the rule is lost, and technology advances occur too quickly for heuristics to be useful. Instead, we need to teach the public how our memory and comprehension processes work and how our biases can lead us to fall prey to manipulators. Ideally, a citizen should be aware that conclusions from reading a post may be affected by the post’s wording, context, the reader’s own prior beliefs, and the presence of similar posts. We also have to teach the public how to evaluate information for accuracy, relevance/soundness, and sufficiency. This includes understanding what constitutes reason-based argumentation and how that differs from persuasion and other forms of belief change. We also have to teach the importance of applying this knowledge, especially when emotions run high and when the information relates to their beliefs.

In schools, we need interdisciplinary experts (e.g., journalists, information-science experts, teachers, cognitive scientists) to work together to develop curricula that can support the development of transferable knowledge. We also need professional development to support teachers in implementing effective instructional approaches to epistemic vigilance.

But teaching epistemic vigilance should not be limited to school. Bad actors throughout the world recognize that social media is an effective way to influence the fabric of our democratic society (e.g., elections, beliefs, harmful actions). And they are correct. The problem is too important to be limited to a few lectures or activities during school hours. It is even possible that Welch would have acted differently if epistemic vigilance was in the zeitgeist. Learning cannot end with primary or secondary schooling. Presumably, new technologies will emerge that will need new forms of epistemic vigilance. For example, DeepFakes are realistic videos which show someone (politician, celebrity) saying something that they did not say. Because they are hard to detect, the public has a right to be educated about them and ways to detect them. A step toward this goal would be public service announcements on the potential dangers of having low epistemic vigilance, such as psychological (e.g., trolls), physical (e.g., swatting), and financial (e.g., phishing) harm.

We Need a Set of Regulations or Guidelines for Using Social Media, Based on Ethics and Awareness of Cognitive Biases

Another recommendation is to focus on knowing how posts can lead to harmful actions taken by others. We think that the initial “pizza gate” Facebook post is akin to shouting “fire” in a movie theater in its outcome. Indeed, recent evidence suggests that people contribute more to spreading false news than robots (Vosoughi, Roy, & Aral, 2018). Whether it is a single malevolent person who gets people to falsely run for their lives, or an inaccurate viral post that roils up Internet users, people are potentially harmed (e.g., Mele, 2016). We note that over the past 15 years, social media companies have gradually implemented rules of use, but we think they should be scrutinized in regard to ethics and factors that affect epistemic vigilance. For example, should a tweet be retweeted if there is clear evidence that the information is suspect? It is possible that pizzagate could have been prevented if more was done to prevent the number of inaccurate posts that resulted from the initial inaccurate one. It is true that Alex Jones was ultimately kicked off YouTube, Facebook, and Twitter, but it took a while. Public policies and regulations may be the next condition for the sustained development of public communication online. Governments have set up agencies dedicated to understanding, tracking, and educating citizens on potential harmful events (e.g., the Centers for Disease Control and Prevention; National Institutes of Health). Regulations have also been created and enforced to limit the negative impact of major innovations such as the automobile. With the growth of social media, news agencies, and the availability of the Internet, we believe, comes the responsibility to identify causes of harmful events (e.g., “pizzagates”, bullying, school shootings) and the proliferation of inaccurate information via the web. Of course, the United States guarantees free speech, but there are limits to free speech when it promotes illegal activity, obscenity, or pandemonium. In some countries, laws prohibit public statements of racial hatred or insults toward law enforcement officers. The European Union has recently begun to take steps for countering fake news (European Commission, 2018). Just like the automobile created a need for traffic lights and speed limits, the Internet calls for some sort of guidelines as to what is permissible to post given human inadequacies in regard to the search for truth. Clearly, one’s motivation to find “the truth” is not enough because Welch was in pursuit of the truth, but likely fell prey to the illusory truth effect.

More Funding Is Needed to Understand and Teach Epistemic Vigilance

In 2009, the U.S. Department of Education’s Institute of Education Sciences launched the “Reading for Understanding Research Initiative” to fund projects to extend reading research beyond decoding and simple coherence-based inferencing (Douglas & Albro, 2014). We need similar leadership to drive rigorous interdisciplinary research to examine developmental trajectories, longitudinally, to study argument comprehension, evaluation, and production as well as factors relating to epistemic vigilance. We need to identify suitable interventions for different populations of readers in K-12 and adulthood. We also need developmental data to understand what is appropriate (e.g., perspective taking, coordinating cognition, attention and inhibition, strategic processing, epistemic beliefs) for a given population.

This review points to the need to support a reasoned approach to addressing fake news as a society.