2.1. Bias

Two theoretical perspectives support the claim that fact-checking in the real world is likely to be biased. First, as Joseph Uscinski and Ryden Butler (Reference Uscinski and Butler2013) argue, fact-checkers’ political ideology is likely to induce them to view epistemically similar claims differently depending on their political alignment (especially when it comes to the selection challenge). As they put it, fact-checkers

On this view, judgments concerning the importance of some statement – judgments driving selection – will necessarily be colored by non-epistemic values. For people for whom political ideology is important, it will constitute a major source of such values.

But it needn’t be the only one. The second perspective that predicts fact-checker bias comes from Roger Koppl’s (Reference Koppl2018) more general “economic theory of experts” (whom he defines as people paid for their opinion – a definition that fits the purposes of this paper, especially as far as Meta’s fact-checkers are concerned, given that the company spends substantial amounts on grants for fact-checking organizations). Koppl’s theory of experts is a deliberate parallel to a more established public choice model of political behavior, applying the toolbox of economic analysis to understand how experts act.

The theory is based on the assumption that experts seek to maximize utility. The assumption is relatively straightforward, though Koppl is careful to stress that expert utility functions are quite capacious. They include not just the (selfless) pursuit of truth, and not just the (selfish) pursuit of material goods, but a variety of other motivations, such as identification with the profession, sympathy for the client, the desire not to antagonize others, and so on.

Combined with the observation that expert cognition is limited and so their error is inevitable, Koppl argues that expert errors (both honest mistakes and deliberate deception) will be skewed by their incentives, rather than distributed randomly. They will favor some parties and perspectives at the expense of others, irrespective of actual truthfulness.

To substantiate this claim, Koppl surveys a wide-ranging literature on cognitive biases, highlighting findings of observer effects (what people notice is a partial function of their expectations and hopes), role effects (what people notice is a partial function of whose perspective – role – they adopt when solving problems), conformity effects (what people notice is a partial function of what they believe others are noticing), and the better-known biases such as anchoring. He argues that such biases influence how the experts process information, leading them not only to make different decisions than what a purely impartial estimation of evidence would warrant but decisions that systematically favor certain viewpoints and perspectives more than others, regardless of epistemic merit.

For example, experts’ sympathy with their clients could lead them to view things from the client’s perspective, which could precipitate role effects and consequently cause the experts to miss, ignore, or underweigh aspects of the situation detrimental to the client’s interests. An expert’s identification with the standards and ethos of their profession could lead to conformity effects, which, in turn, could induce them to give less credence than is warranted to claims they think many of their peers would disagree with. An expert’s expectation that her political opponents are unscrupulous liars could lead her to interpret ambiguities in the opponents’ claims to their disfavor, while she would not do the same for people on her side.

Applying Koppl’s framework specifically to fact-checkers, we arrive at the following picture: fact-checkers, qua experts, are motivated not just by an uncompromising pursuit of truth, but also by a variety of more and less selfish motives; fact-checkers are fallible and make errors; their motivations can (consciously and unconsciously) affect which content they fact-check, how they fact-check it, and how they determine the final rating. The ratings thus depend not just on the epistemic credentials of the content (how well is it supported by the evidence?), but also on who the fact-checkers’ clients are (and what the experts believe their clients’ preferences are), who the fact-checkers identify with in terms of their professional and other associations, and who they seek approbation from.

Placing Uscinski and Ryder’s insights within Koppl’s framework amounts to adding the explicit consideration of ideology as a further component of the fact-checkers’ utility function. The ideology espoused by fact-checkers could then, by the mechanisms outlined by Koppl, skew the distribution of honest errors in fact-checking verdicts in a way that favors claims and perspectives ideologically aligned with the fact-checkers.

2.2. Monopoly and monopsony

If platforms employed a variety of ideologically diverse fact-checkers, we would have some reason to take their necessarily biased nature more lightly. Competition disciplines and fact-checkers with superior track records would grow in reputation and following. On the other hand, the concern with bias acquires more urgency when we realize that one feature of social media platforms’ attempts to solve their misinformation problem is (what one might call) the monopolization of epistemic authority. Only a small number of organizations (“epistemic monopolists”) are invested with the official power to determine whether any piece of content should be considered misinformation. They alone decide what merits the official sanctioning as false, misleading, etc., and they face no opposition equipped with similar power.

Even if, as in Meta’s case, the platform employs multiple organizations involved in fact-checking, they share the same client, the same or similar identity (as fact-checkers, journalists, etc.), and the same desires for approbation, especially within their peer group. They also seem ideologically similar, in that the majority of fact-checking organizations exhibit left-wing bias (AllSides, 2023). More importantly, they don’t seem to compete among themselves (e.g. by providing different fact-checks of the same information, leaving the decision on which determination to accept to the client; rather, they tend each to respond to different kinds of content, serving different languages and world regions). Consequently, it is still appropriate to label them epistemic monopolists, though admittedly it stretches the concept a little. (Perhaps they would be better described as an epistemic cartel).

As Koppl (Reference Koppl2018; Reference Koppl2023) finds, monopoly expertise, particularly in fields where evidence is often unequivocal, is especially susceptible to influences that extend beyond the analysis of the evidential basis for any particular claim, including the expert’s non-epistemic preferences and biases. In a monopoly, such biases will not be counteracted by any competing experts; thus, the costs of indulging them are much reduced. This is likely to lead to the information ecosystem of the whole platform being skewed in directions favorable to the fact-checkers’ non-epistemic values. Thus, fact-checker bias, under conditions of cartelization or monopoly, results in the entire within-platform information environment being biased.

A further problem for Meta’s solution to SMIP is that the company is the only buyer of fact-checker services for its platforms. Thus, in addition to problems arising from the epistemic monopoly, we must also be wary of Meta’s monopsony power. As Koppl puts it,

Similarly, fact-checking organizations used by Meta are likely to be reluctant to issue verdicts they will see as opposed to what their client wants and expects, regardless of the purely epistemic merits of any given claim. After all, they depend on the company for a significant portion of their funding. As is the case with other major tech firms, Meta’s employees (especially its top-level executives) tend overwhelmingly to support Democratic candidates (Levy, Reference Levy2020). Since this is public knowledge, fact-checkers employed by Meta might be unwilling to treat pro- and anti-DemocraticFootnote ⁷ perspectives in the same way, given that they know which direction the buyer of the services is biased in.

Though I have struggled to find relevant hard data, the predictions generated from Uscinski & Butler’s and Koppl’s models bear out insofar as plentiful anecdotal evidence is concerned (for first- and secondhand accounts see e.g; Magness & Yang, Reference Magness and Yang2022; Fritts & Cabrera, Reference Fritts and Cabrera2022; Friedman, Reference Friedman2023; Houghton, Reference Houghton2024 ; for a general discussion, see Gable et al., Reference Gable, Brechter, DePoy, Mastrine and Weinzierl2022).

In line with Koppl’s complaints about expert monopoly and client monopsony in general, the lack of competition among fact-checkers, and their all working for Meta, creates an institutional structure where bias is harder to counteract and more influential on political discussions on the entire platform. The prevalence of such whole-of-platform skew, absent contrary influences, has serious consequences.

For example, if (as is reasonable to suppose) viewpoint diversity has epistemic benefits (see McBrayer Reference McBrayer2024), then the institution of epistemic monopolists and the consequent imposition of some form of viewpoint uniformity (or at least a move towards more viewpoint uniformity on the margin) will prevent (or make more difficult on the margin) the acquisition of these benefits. In other words, since an expert monopoly will exclude some perspectives for non-epistemic reasons, it will hinder the search for truth.

Consequently, it is highly probable that real-world anti-misinformation strategies instituted by some of the largest social media platforms will fail in systematic ways, yielding to the epistemic monopolists’ (and monopsonists’) biases in assessing accuracy of various claims, generating epistemic costs.

Overall, the above considerations make worries about fact-checker bias, especially under monopoly/monopsony conditions, eminently reasonable.

2.3. Ineffectiveness

Substantial empirical research suggests that labeling content as false or misleading – the main strategy in fact pursued by Meta to combat misinformation – can boast limited real-world effectiveness. The intervention appears largely unable to change people’s minds.

A meta-analysis of research on fact-checking offers this sobering summary:

In fairness, other researchers find robust positive effects of fact-checking on belief in false information (Wood & Porter, Reference Wood and Porter2019). And, to complicate matters, some studies seem to show that interventions such as fact-checks decrease belief in false and true information alike (Hoes et al., Reference Hoes, Aitken, Zhang, Gackowski and Wojcieszak2024). Yet, though the jury is still out, the balance of the evidence, for now, tilts toward ineffectiveness (insofar as meta-analyses should count for more than individual papers).

Thus, it is reasonable to conclude that interventions that stop short of removing offending content (i.e. interventions such as fact-checking and labeling) seem hardly able to reduce belief in misinformation.

Nor do outright bans and content removals significantly help with reducing access to misinformation. This is true for several reasons: first, in general, when the supply of a good is artificially restricted while demand remains unchanged, consumers will tend to seek alternative ways to obtain the good, for instance by patronizing less scrupulous competitors. By the same token, when the quantity of misinformation is restricted, but demand for it doesn’t change,Footnote ⁹ people can be expected to turn to other websites, with less stringent moderation policies, for the missing part of their information diets. Absent some totalitarian control over all information on the internet, such sites will be fairly easy to find.

Nevertheless, transaction costs of seeking out content that has been banned on mainstream social media are not negligible (e.g. it takes some time and effort to set up a new account, familiarize oneself with the new user interface, seek out the relevant profiles, etc.), and it’s possible to find substitute goods on the dominant platforms instead (e.g. those accounts that push the boundary of controversial discourse without crossing the threshold for bannable offenses). Still, while not everyone will actually seek out alternative information sources on other platforms, we can expect at least some users to do so.

Michael Huemer (Reference Huemer2021) makes a similar point. In his estimation, attempts to suppress ideas, when carried out by “factions” (ideological groups) that lack the complete control over discourse, are likely to be futile:

A telling illustration of this phenomenon – again, anecdotal – is that Twitter’s and Facebook’s bans on sharing and linking to anything related to the New York Post’s reporting on Hunter Biden’s laptop in the run-up to the 2020 US presidential election, which appear to have increased people’s interest in the story and its protagonist. As columnist Philip Bump puts it,

The point about the ineffectiveness of social media content bans is strengthened when we look at governmental efforts to censor information as a comparison. In Eric Berkowitz’s blunt summary of his history of the topic, government “censorship doesn’t work. The ideas animating suppressed speech remain in circulation and, in the end, can become more effective for being forbidden. Will Durant’s conclusion about the First Emperor’s destruction of unapproved texts applies to most censorship: ‘The only permanent result was to lend an aroma of sanctity to the proscribed literature’” (Reference Berkowitz2021, 8, emphasis in original).

Berkowitz’s points are echoed in other works examining the history of censorship (see e.g. Strossen, Reference Strossen2018; Mchangama, Reference Mchangama2022; Kosseff, Reference Kosseff2023).

If state censorship does not succeed in actually suppressing ideas, its corporate facsimiles, such as content removals seem to stand even less of a chance of doing so. The reason is simple: corporate suppression lacks both the scope and the means that states can deploy to stamp out information. The penalties for violations are typically much more severe for flouting government rather than corporate censors (Messina, Reference Messina2023). It is therefore no surprise that it’s more difficult to find information that’s illegal to distribute than it is to find information that’s banned on Facebook.

2.4. Why the interventions don’t work

Empirical psychology supplies a credible explanation for the ineffectiveness of the existing solutions to SMIP. We know from countless studies that people don’t reason about politics in truth-conducive ways; specifically, they tend not to change their minds in proportion to new evidence they encounter but assess the credibility of the evidence through the lens of their political ideology (if it’s congruent with the ideology, it’s accepted; if incongruent, it’s dismissed and/or underweighted). Fact-checks and labels operate by simply supplying new evidence; so, it is difficult to expect that they will significantly move people to change their beliefs. The reason for this epistemically irrational (but instrumentally rational) behavior lies in people’s incentives. Hrishikesh Joshi puts the point well:

There is no reason to think social media improves upon this dynamic in any meaningful way. The opportunity to discuss politics online doesn’t seem to affect the pre-existing incentive structure for most people, at least not in the desired direction. Instead, it may exacerbate incentives for partisanship: partisan signaling might be rewarded more in online contexts than in offline discussions. So, to the extent that political misinformation flatters and truth hurts, we can expect that people will sacrifice accuracy for partisan identity in political debates on social media.Footnote ¹⁰

However, in experimental conditions, manipulating people’s incentives by financial means reliably improves their ability to sort truth from falsehood. Paying people money for getting things right significantly reduces partisan bias (Bullock et al., Reference Bullock, Gerber, Hill and Huber2015; Prior et al. Reference Prior, Sood and Khanna2015) and significantly increases the ability to evaluate the accuracy of a news item (Panizza et al., Reference Panizza, Ronzani, Martini, Mattavelli, Morisseau and Motterlini2022; Ronzani et al., Reference Ronzani, Panizza, Morisseau, Mattavelli and Martini2024). The aim of political belief seems malleable. People can be incentivized to seek out the truth, even in politics.

As observed by Adam Gibbons (Reference Gibbons2023a), solutions such as fact-checks and many other methods seeking to improve the accuracy of political belief, do not target incentives (at least not strongly enough). They seek instead to prevent and to dissuade. This is why they fail.

3.1. The marketplace of ideas needs more markets

In what follows, I develop a modest proposal for an alternative solution to SMIP. I argue that it would be desirable to explore, in practice (for example, through small-scale experiments), whether replacing fact-checks sourced from monopolistic experts with user-driven prediction markets can improve upon the status quo. The proposal is modest because I am not claiming that we should, all things considered, implement this solution across the board right now. Rather, I argue that the evidence justifies thinking that it’s a promising approach with a significant chance of improving the relevant outcomes. As a result, it is worth trying.

Empirical evidence mentioned above (and common sense) shows that when people are adequately rewarded for being accurate, they will strive to be (and succeed in being) more accurate. This is why monetary incentives reduce partisanship and increase informed-ness. Consequently, using monetary incentives that reward accuracy, in order to improve discourse on social media, would make people less partisan and better informed.

A “solution” that would consist of simply giving people money for, say, sharing accurate news stories is obviously a nonstarter. The system can be abused by users (not to mention bots) at will, relentlessly sharing the stories officially declared accurate, merely in order to gain the rewards. That aside, the question remains of who would have the authority of official determinations of accuracy for the purposes of issuing the rewards. Reposing this authority in any single corporation (or a coterie of fact-checking organizations) replicates the problems associated with the epistemic monopolists and monopsonists I already discussed, and possibly exacerbates them.

However, there is a different way of bringing monetary incentives into the marketplace of ideas that appears to avoid these problems. What I have in mind is an institution enabling individual users to create and participate in betting markets concerning an essentially unrestricted array of topics. I call this the prediction market-based solution (PMBS) to the problem of social media misinformation. If individual users are empowered to take part in bets concerning the future (or indeed even the present or the past), where they obtain financial rewards for getting the facts right, they will ipso facto be incentivized to become better informed and aim for accuracy rather than partisan acclaim.

Per Robin Hanson (Reference Hanson2013), prediction markets are

How, specifically, do prediction markets translate financial incentives into probability estimates? Hanson offers one example:

In existing digital prediction markets, such as the online platform Manifold, users can “[b]et on politics, tech, sports, and more [o]r create [their] own play-money betting market on any topic [they] care about” (Manifold, nd., np.). Topics are unrestricted (within limits of the law) and no real money is involved (users bet non-redeemable virtual currency instead). My proposal is neutral on what the rewards should be: it seems that using real money would be the most effective way of affecting incentives, whereas virtual non-redeemable currencies raise the fewest legal issues.

The basic idea behind PMBS is to enable (perhaps encourage) social media users to engage in betting market creation and betting market participation without leaving the social media platform. They could also have the option of challenging others to a bet, on the basis of the statements others make (perhaps, in addition to “like” and “repost” buttons, platforms could add a “wanna bet?” button). This would introduce a new set of incentives to the social media ecosystem, rewarding accuracy and precision, and punishing error and exaggeration, in ways that traditional fact-checking does not.

Empirically, prediction markets boast a superior track record over other methods of knowledge aggregation with respect to accuracy. Specifically, they are superior to the kinds of information exchanges associated with an idealized conception of social media as a public square (such as deliberation). Consequently, it is reasonable to suspect that a social media platform inclusive of an integrated prediction markets functionality would outperform epistemically monopolized information ecosystems on the relevant benchmarks (such as political knowledge and belief in misinformation), just as prediction markets outperform both expert predictions (Gardner, Reference Gardner2010) and deliberation (Sunstein, Reference Sunstein2006) on group decision quality (provided, of course, that a significant number of users take advantage of this new functionality).

The same mechanisms that make prediction markets outperform more deliberative and expert-driven methods of information aggregation also explain why they would improve political discourse on social media to a greater degree than fact-checking does.

3.2. PMBS reduces biases and improves accuracy

Users creating a betting market alone decide what the winning conditions of the bet are. In many cases, established “official judgments of the fact in question” would be available and their sources uncontroversial (who wins the Premier League; what the GDP numbers are for a given year; how many female directors are among the Academy Awards nominees).

In cases where official confirmation is more difficult to come by (was the inauguration of president X better attended than that of president Y?), most bet conditions would presumably include the sources according to which the result will be adjudicated. Any other user can accept the bet, and thus, implicitly, accept the epistemic authority of such sources. However, there would be no externally imposed epistemic monopolist granted the ultimate decision-making power to declare what the truth is.

If betting market creators were rewarded, e.g. by the platform, in proportion to the number of people that accept the bet, they would be incentivized to select those conditions for settling bets that would be acceptable to as many other users as possible; hence, they would tend toward choosing the suppliers of judgments with widely endorsed epistemic credentials. Since bets are, in a sense, adversarial, traders with very different expectations as to the final outcome would have to agree to the bet-settling conditions, including the procedure by which to determine what facts are obtained. Consequently, reputation for impartiality and objectivity would be an important consideration in selecting the sources on whose determinations the winning conditions rest, even on highly controversial topics. Market creators would tend to rely on sources that come closest to commanding universal respect in matters epistemic because that would secure the biggest possible pool of traders.

So, one way prediction markets have of reducing bias is by incentivizing the search for impartial sources – epistemic authorities without monopoly – to settle individual questions. A good proxy for source impartiality is the willingness (the consensus) of adversaries to voluntarily abide by the source’s determinations when it comes to settling bets.

Secondly, and more importantly, in speculative markets, traders are straightforwardly penalized for their biases, while people who are perceptive enough to suspect biased estimates and set aside their own wishful thinking stand to profit (recall the influence of monetary incentives on increased judgment accuracy).

Strong partisans and ideologues are more likely to allow their biases to cloud their judgment and, consequently, forgo a dispassionate analysis of the available data. This can lead to poor decision-making and, ultimately, financial losses. Unlike in typical political discussions, partisan bias does not pay when forecasting the future.

Instead, prediction markets incentivize traders to seek out the most accurate and up-to-date information, since this is the most reliable method of earning profits. This motivation can help overcome personal biases, as traders prioritize gains over retaining incorrect beliefs; however comforting. The competitive nature of prediction markets disciplines traders to continuously update their beliefs in response to new data. Those who make biased decisions are more likely to lose their bets.

Theoretical models of speculative markets predict that both attempts at market manipulation and the presence of biased traders encourage informed participation by others. Empirical data seem to bear this out (Hanson & Oprea, Reference Hanson and Oprea2009; Lott, Reference Lott2022).

It is reasonable to expect that, for a large number of people, rewards for partisan signaling of the sort described earlier by Joshi will be worth less than the monetary rewards for getting facts right. So, PMBS incentivizes good epistemic behavior better than the systems relying on monopolist fact-checkers.

Further, awareness of potential exposure to a betting challenge could encourage precision and care in formulating judgments. Knowing that my assertions can be subjected to a quick reality test where people stand to gain/lose real assets may impose a greater degree of rigor and moderation on what I say than simply knowing that I might get fact-checked.

As Bryan Caplan puts it,

In contrast, under the present-day incentive structure, there is little (internalized) cost, and much reward, for ignoring evidence at the expense of political points-scoring. Being wrong on the facts is worth it, provided our political allies applaud what we say. Indeed, users may treat fact-checks on their posts and removals of their content “as badges of honor, earned by bravely standing up to the biased verdicts of overly powerful and influential social media organizations” (Gibbons Reference Gibbons2023a, np.). It is much harder to portray a betting loss in a positive light.

Betting markets, in virtue of their built-in incentives, will thus likely impose more rigor, and more moderation, on what people say. In our polarized age, this is a considerable benefit.

In sum, PMBS has the potential to improve on the fact-checking-based solutions to SMIP. The hypothesis that PMBS will be an improvement generates testable empirical predictions. An experiment comparing people’s misinformation susceptibility and political knowledge after exposure to traditional fact-checks versus PMBS does seem feasible. If I’m right, people exposed to PMBS would tend to score higher on these metrics.

4.1. Participation

One could raise some worries. On the whole, people don’t seem especially interested in prediction markets. Manifold has 1,700 daily active users. That’s less than 0.001% of Facebook’s numbers (2,000,000,000 daily active users). Why think that substantial numbers of social media users would want to bother with betting if they aren’t interested already?

This is a very serious challenge facing my proposal; of course, it is an empirical question of whether platform-based prediction markets will attract users, and most people might remain indifferent to this sort of activity. There is some reason to think, however, that prediction markets could generate significant interest.

For starters, people respond to incentives, and monetary rewards are well-known to be effective motivators. If the rewards offered are high enough to outweigh whatever “psychic profit” one earns from partisan acclaim, people will be likely to at least try their hand at prediction markets.

Second, social media platforms have the potential ability to design user interfaces in such a way as to make betting a seamless and attractive part of the user experience, thus eliminating, as much as possible, the transaction costs of participating in speculative markets. Outstanding user interface design is, arguably, one of the main reasons for the success of social media platforms. If anyone can make online forecasting a mainstream pastime, it’s them. In general, we should take seriously the idea that making participation not just easy, but also fun, could help induce greater popularity of prediction markets among ordinary users.

Total success is not necessary. Even if only a substantial minority took advantage of the new features, some epistemic improvements over the status quo could be expected. These are not to be scoffed at. After all, small improvements to the effectiveness of other methods of fighting misinformation are widely celebrated.

4.2. Belief

But would people rely on the information provided by prediction markets in their deliberations and decision-making? This question, too, ultimately, admits of empirical determination. I will, however, venture to offer some armchair speculation as to why it’s not unreasonable to think that at least some people would.

As far as active traders are concerned, it seems to me that it would in general be psychologically much easier for them to ignore (and take steps to avoid) testimonial evidence from outside sources (e.g. in the form of fact-checks) than it would be for them to ignore evidence they already believe, especially when it comes to the evidence that was gathered with what they believe to be epistemically sound methods. Successful participation in prediction markets requires acquiring evidence in an epistemically respectable way. Thus, for active traders, that kind of evidence would be that much harder to disregard than the evidence they encounter when looking at the fact-checking labels. Consequently, betting encourages good epistemic practices as well as belief in, and practical commitment to, their results. Even if such results cut against our ideological priors, we can’t ignore them as easily as we can ignore inconvenient testimony.

On the other hand, when it comes to passive users who, despite platforms’ efforts, would not partake in prediction markets (while remaining aware of their outcomes), PMBS does not offer a straightforward path to epistemic improvement. Such people are likely to retain their old habits of thought, processing the outputs of prediction markets in the same epistemically deficient way that they process the content of the fact-checks they encounter (rather like they process the outputs of existent prediction markets under the status quo). As a result, it appears that PMBS’s potential effectiveness is likely to be proportional to the number of people becoming active traders.