Platforms, Polarization and Democracy

So Cosma Shalizi and I have an article (messy pre-print) coming out Real Soon in Communications of the ACM on democracy, polarization and social media. And Nate Matias, who I’m friends with, has forceful objections. I’ve promised him a response – which is below – but am doing it as a blogpost, since I think that the disagreement could be turned into something more broadly useful.

Cosma and I wrote the article to push back against one version of the common claim that we can blame everything that is wrong and toxic with social media (and by extension, American democracy – this is a U.S. centric piece) on engagement maximizing algorithms and their cousins. Specifically, we don’t think that we can fully blame these algorithms for the kinds of belief polarization that we see online: people’s willingness, for example, to concoct elaborate justifications for their belief that Trump Really Won in 2020.

We do this by engaging in a kind of thought experiment. Would we see similar polarization of beliefs if we lived in a world where Facebook, Twitter et al. hadn’t started using these algorithms after 2012 or so? Our rough answer is that plausibly, yes: we would see lots of polarization. Following Mercier and Sperber, we assume that people are motivated reasoners – they more often look for evidence to support what they want to believe than to challenge their assumptions. And all they need to do this is a combination of simple search (Google like it used to be) and social media 2.0.

Search enables them to find evidence that will support their priors, while social media enables them to link to, comment, elaborate on and otherwise amplify this evidence. Because of how simple search works (it treats web links and activity as proxies for quality), the more that people link to and comment on stuff on social media, the easier it is to find, and the easier it is to find, the more that they will link to it and comment on it. We (for values of ‘we’ that actually mean ‘Cosma’) construct a simple model, which suggests that this feedback loop leads to a world where there are a few vast glob-like communities of mutually reinforcing beliefs surrounded by a myriad of smaller, and less consequential communities. In statistical terms, the sizes of different communities fall along a rough power law distribution.

Nate – who is no more a fan of the algorithmic polarization consensus than we are – takes exception to what we say. First – he doesn’t like that we use the word ‘toxicity,’ because it has no agreed-on definition, and is sometimes used as “a way to obfuscate and sidestep precision in order to avoid hard debates about democratic governance.” Second, he doesn’t like what he takes to be our “assumption that an Internet with many smaller groups would have greater toxicity and a world with fewer, larger groups would be less toxic,” and that “[s]ince [our] simulation can imagine an Internet with many small groups, [we] conclude that a toxic internet does not depend on social media algorithms.” Third, he thinks, following a recent article by Dan Kreiss and Shannon McGregor, that arguments about polarization and group size too are ways to “avoid talking about racism, sexism, and inequality.” Finally, he thinks that our model is misleading – while it is “compellingly simple,” it needs to look at “more empirical work beyond just famous papers published in Science and Nature,” and should ideally be grounded in community science.

It’s not going to surprise anyone that we, in turn, disagree with most of these criticisms (the exception is that we could certainly have been more specific in how we used the term “toxicity” and will do what we can to mitigate at this point in the production process). And perhaps there’s a more interesting and productive disagreement than you might think from academics having at it online, even in a reasonably friendly way, about what each said and meant. But to get there, we likely need to clear up some misunderstandings.

The most straightforward one is that Nate isn’t quite right about what our model says and what we argue. We don’t actually think that a world with many smaller groups would be less toxic, and a world with more big ones less so. Our argument is just the opposite of that. We think that even just with simple search and social media, the Internet creates a world in which deranged beliefs can scale more easily than they used to. Before the Internet, it was harder for people to find and glom onto mistaken beliefs that pushed against the common wisdom. This meant either that they were likely to go with that wisdom (which of course was itself usually dubious) or invent their own idiosyncratic dubious alternatives, pushing out in a myriad different directions, which to some extent canceled each other out. In our counterfactual, even simple Internet technologies of search and Web 2.0 would allow them to construct their own alternative realities, collectively, and at scale.

That counterfactual isn’t necessarily worse than the recent past, where there was high public consensus around ideas and beliefs that were often pernicious. Also, it isn’t obviously better than the world that we actually live in, which is the comparison that we are actually looking to make. Our simple model suggests that both with post-2012 algorithms and without, we end up with much the same outcome – a world where there are big agglomerations of people with fundamentally discordant political beliefs.

So what does this initially counterintuitive comparison get us? I recently read a book by the philosopher Cailin O’Connor, who independently adopted a very similar mode of counterfactual argument (she gives Liam Kofi Bright partial credit). O’Connor is interested in figuring out the causes of racism and gender discrimination. As she points out, many people attribute racism and gender discrimination to psychological biases, such as stereotype threat. She doesn’t want to discount these explanations. But she wants to investigate whether we would still see large scale gender and racial discrimination in a world where human beings weren’t biased in these ways.

Obviously, O’Connor can’t directly observe a counterfactual world in which individuals were perfectly rational and not subject to psychological bias. So, like us, she constructs a simple model, in which people are rational. She shows that under plausible assumptions and conditions:

A modeling perspective can show us that the conditions necessary to generate pernicious inequity in human societies are extremely minimal. Under these minimal conditions, cultural evolutionary pathways will robustly march towards inequitable systems. These models do not prove that real world systems of inequity have, in fact, evolved via these simple cultural evolutionary pathways, but they tell us that they could. In particular, they show that even if many of the most pernicious psychological facts about humans are removed or mitigated, inequitable conventions of the second sort are still expected to emerge.

In other words: we might see racial and gender bias continue at the system level (e.g. Black people and women consistently being discriminated against), even if we somehow, magically, got rid of all the psychological biases have at the individual level about Black people and women. This is a really valuable finding. Indeed, my only significant objection to O’Connor’s book is that she doesn’t make nearly as much as she might of it. Economists like Gary Becker and Milton Friedman were extremely fond of arguing that racism and sexism were irrational and would disappear if only markets were allowed to work their magic. O’Connor uses economic models to demonstrate the contrary: why racism and sexism may continue to thrive under conditions of rational exchange, and I would love it if she was just a little blunter in sticking it to the Becker/Friedman complex. Bright, Gabriel, O’Connor and Taiwo have used similar modeling techniques to build a model of enduring racial capitalism: perhaps this will provide a platform from which some lively and useful future polemic will be launched.

I’m not saying that what Cosma and I have done is nearly as valuable, but its approach is very similar. Like O’Connor, we use our model not to represent reality as it is, but to build a counterfactual, suggesting that if we had not invented the panoply of modern social media algorithms in the first place, we would likely have ended up in much the same place. This does not say that these algorithms didn’t contribute, any more than O’Connor’s arguments absolve the psychological bases of racism and sexism. It does strongly imply (assuming that our model is not utterly mistaken) that the problems would still exist even if the algorithms did not.

But like all theoretical frameworks, our counterfactual has its implied politics – and here is where I think there is scope for a more useful and specific disagreement. As mentioned, Nate links to a very new piece by Kreiss and McGregor, which argues that much of the literature on polarization is not only misconstrued but actively misleading. In their words, “Our foundational claim is that polarization might not be bad for democracy—it might in fact be a necessary outgrowth of efforts to achieve democracy.”

Kreiss and McGregor go on to detail the various ways in which the literature on polarization and platforms harks back to an imagined pre-polarization America (which enjoyed an apparent consensus only because Black people and others who disagreed were suppressed). They argue that we should pay attention to inequality rather than polarization when we look to assess the health of democracy. More bluntly: we should understand that much of today’s apparent polarization is the result of people’s efforts to redress the inequality that has been part and parcel of America’s purportedly democratic system for decades. The struggle to actively achieve American democracy inevitably involves contention – and not all sides are equal. Those who are pressing for more equality and justice – especially but not exclusively racial equality and justice – have a very different status than those who are trying to defend unjust relations. And the focus on polarization tends to push those important questions to the sidelines.

So Cosma and I largely agree both with this diagnosis of the literature and with the understanding of democracy that propels it. We don’t talk about this in the article, except indirectly in side comments – we were bounded by both sharp word count constraints and a fifteen citations limit (as an aside, we don’t just cite to “famous” Nature and Science pieces – we cite just one article from either Nature or Science and two more from the Nature/Science Extended Universe ™). We used one of our precious citations to point to a previous article where we set out our account of democracy, which (a) emphasizes that democracy involves rowdy struggle, and (b) stresses that “a commitment to democratic improvements is a commitment to making power relations more equal.” If we’d had more room (extensive self-citation is especially egregious when you’re cramped within the confines of a tiny bibliography) we’d likely have cited arguments elsewhere e.g. about the value of democratic instability in tearing up old racial and gender norms, and how “strong gatekeeping” media systems in the pre-Internet era subordinated Black voices and perpetuated myths about Black people.

In other words, our arguments start from a place that is broadly located within the equality-centric understanding of democracy that Kreiss and McGregor are looking for (of course: there may be aspects of our understanding that they and others might still very reasonably dispute). More broadly speaking, one of our major intellectual projects, with Danielle Allen is to try to build a model of democracy, explaining how its central commitment to equality provides it with dynamical advantages (this relates in important ways to the O’Connor book described earlier – it also relates in different ways to Danielle’s fantastic recent book).

All this said, our article is very explicitly a piece about polarization and democratic stability. Its underlying intuition is that if beliefs become too polarized, democracy will become unstable. And that is not an inherently stupid or biased argument. As Kreiss and McGregor summarize a broader literature:

at some fundamental level the groups that exist within a pluralistic society must accept one another as legitimate, even though they may have opposing values, interests, and ends. Groups must tolerate one another, accepting each other’s right to exist and to advance their interests in private and public spheres. This tolerance is essential given that groups often define themselves through drawing boundaries with others (Smith, 2003). It is often socially and politically powerful to create and draw hard edges around a shared identity, conjure a clear opposition, and define competing interests, especially through media spheres that support building, maintaining, and contesting political power (Squires, 2002). As such, some level of polarization is an endemic feature of social and political life. Polarization becomes problematic, however, when it is so extreme as to erode the legitimacy of opposing groups, the tolerance that democratic co-existence is premised upon and faith among partisans that the other side will continue to engage in free and fair elections (Haggard and Kaufman, 2021),

I think – though I am not entirely certain – that Kreiss and McGregor endorse this understanding. Their article is explicitly a “provocation” and a polemic against the tendencies of the polarization literature that they rightly detest. Still, they acknowledge that it only applies to “some” of the literature, that polarization can be “dangerous,” and that some share of the people who worry about polarization (including my Hopkins colleague Lily Mason) have an approach to polarization that doesn’t suffer from these flaws, and that polarization has risks too.

So what are the specific risks of the belief polarization that we talk about? Again, there’s writing elsewhere that we weren’t able to cite to, which emphasize that even under a minimalist account of democracy, we need shared (and justified) beliefs in the electoral process, and in the willingness of government parties and officials to give up office when they lose an election. That is a foundation of democratic stability, even if we embrace contention and equality as core elements of democracy.

As Kreiss and McGregor say, we should not embrace concerns about polarization “at all costs.” But we shouldn’t completely exclude these concerns either. Some opponents of polarization seem to think that to heal democracy, we all just need to start liking each other. That isn’t a particularly serious claim. But the claim that we need to figure out ways to live together in some minimal consensus, however grudging, is, I think, one that ought be taken very seriously indeed. Hence our argument, which stems from the claim that this consensus is democratically possible even under a realistic and moderately pessimistic account of human psychology (here, we implicitly push back against some prominent recent anti-democratic arguments). If those psychological microfoundations are right, we even have some general clues as to the foundations of a better and more stable democracy.

So this is the disagreement that I think is worth taking up. If Kreiss, McGregor, and for that matter Nate, don’t think that polarization is a problem at all, then it would be good to know this. But I really don’t think that they believe this. If, alternatively, they think that polarization is a problem, but one that has been misused by people who idealize a largely imaginary peaceful American past, then there isn’t any disagreement in principle between them, Cosma, and myself. Of course, there may be, and almost certainly are, practical disagreements, and articulating these disagreements and thrashing them out would be potentially very useful.

More broadly – I think we are all committed to an understanding of democracy that is both (a) more just and egalitarian, and (b) stable against urgent threats, which do include polarization. But figuring out how to reconcile justice with democratic stability is extremely difficult, both in the particular and the general. And it requires the bringing together of different kinds of knowledge. When Nate suggests that the framework that Cosma and I use is empirically unfounded, he’s wrong. We’re building on a large body of research in human psychology. But if he were to make the (mildly modified) statement that our framework is severely empirically limited, he would be absolutely right. It sketches out the landscape of one important problem, but doesn’t say much at all about how to solve it, or reconcile possible solutions with other major problems that we face.  I think that the “community science” that Nate favors is one enormously important – even crucial – source of ideas about how to do this, as part of a broader “translational” approach to building democracy, which helps address the deficiencies of big theories, but this post is already very long, so I’ll leave it there.