Democratized Speech and the Spread of Social Ills

How the architecture of online platforms turns ordinary disagreement into something more contagious—and why the virus metaphor only gets us halfway there.

Almost everything we know, we learned from someone else’s words. Speech is the oil of social cooperation, the medium through which we coordinate, trust, learn, and disagree. We became the dominant species on this planet not because we are strong or fast, but because we can tell each other things. 

Yet the same capacity that built our institutions seems to be eroding them. Trust in governments, journalists, scientists, neighbors, and strangers has frayed. Public discourse fragments into hostile tribes. Conspiracy and grievance circulate at high speed. Ordinary people, not just bad-faith provocateurs, find themselves saying things online they would never say in a room with another human being.

How did the medium that made us cooperative become a medium for collective cruelty?

The familiar answer points to social media, and it is correct as far as it goes. But it usually mistakes the mechanism. The popular version is essentially viral: bad ideas, like germs, jump from one infected user to another, and platforms simply provide more efficient transmission. A particularly nasty meme is like a particularly contagious cold, and the right response is some combination of quarantine, hand-washing (media literacy), and, for the worst pathogens, removal from circulation. 

This picture is intuitive. It is also, as I argue elsewhere, insufficient. To understand what online platforms are doing to public discourse, we need a different model of contagion altogether. Once we have it, we can see that the worry is not really about speech spreading. It is about platforms unusually well-engineered to convert ordinary speech into something more durable, more identity-forming, and harder to dislodge.

What is wrong with the viral picture

The viral model, or what epidemiologists call “simple contagion,” has one defining feature: a single exposure is, in principle, enough. You stand next to someone with measles; the virus enters your respiratory tract; you may now carry it. What matters is the probability of infection per contact, and a disease can sweep through a network rapidly because the threshold for adoption is, essentially, one.

Hate and conspiracy do not work this way. Almost nobody becomes a committed misogynist or ethnonationalist after a single tweet. Most of us have encountered any number of vile claims online without internalizing them. When someone does shift, the process is cumulative: a slow build-up of exposure, framed by repeated social reinforcement, until certain attitudes harden into beliefs and certain beliefs harden into identity. The “infected” person eventually becomes a generator of the very content that infected them.

Network scientists call this complex contagion. Unlike simple contagion, it requires multiple, reinforcing sources before adoption. The classic examples are not diseases but behaviors: joining a protest, changing diet, adopting a risky new technology. People are far more likely to adopt these when several different members of their social network do, and when those adoptions converge in a way that signals: ‘this is what people like me do.’ Empirical work on Twitter and other platforms has shown that contested information online actually does spread in a complex rather than a simple pattern.

The distinction changes what we should be looking at when we look at platforms. If the relevant model were simple contagion, the right intervention would be to reduce the number of nasty messages a user is exposed to: a crude quarantine. If the relevant model is complex contagion, the interventions that matter are the ones that disrupt reinforcement: the conditions under which the same idea, repeatedly endorsed by ostensibly varied sources within a community, comes to feel like consensus. The single nasty post is not the unit of harm. The unit of harm is the environment in which that post is one of many, all pointing the same way, all signalling that this is what one’s people now believe.

Democratization and its discontents

Why are online platforms such effective as engines of complex contagion? Three features of the online environment, taken together, distinguish it from the public spheres we had before.

The first is the democratization of content production. Until recently, broadcasting was capital-intensive. Reaching a national audience required a printing press, a transmitter, or a newspaper desk. Those gatekeeping institutions had many failures, but they also imposed friction: editorial standards, defamation liability, professional norms, reputational stakes—these filtered some (not all) of the worst material. Today the marginal cost of broadcasting is essentially zero, and the marginal cost of amplification depends mostly on whether you can capture attention. Anyone can speak; the people best rewarded for speaking are not necessarily the people we would want to hear. This is not a complaint about democracy. The old gatekeepers excluded much that should have been heard. The point is that gatekeeping does work, and removing it without replacing it with anything changes what reaches the top of a feed.

The second is the reshaping of communities. Geography no longer determines who you encounter. People with niche interests can find each other across continents: a teenager with a rare condition, a queer person in a hostile town, a hobbyist whose interest is too specialist for any single city. Each can now find their people in ways no previous generation could. But the same affordances let communities organized around grievance, conspiracy, or hatred to find each other just as efficiently. Anonymity weakens the social cost of cruelty. Group formation no longer requires shared physical space or even shared language. And once such communities form, they do something epidemiologically distinctive: they reinforce each other. The geographically dispersed misogynist is no longer the village oddity surrounded by neighbors who would correct him. He is part of a community of thousands who confirm his view of the world hourly.

The third is the structure of incentives. Platforms make money from attention, and the algorithms that shape what users see are trained to maximize it. Content that provokes a strong reaction—outrage, fear, partisan identification—performs better than content that does not. This is not a moral failing of platform design; it is the predictable result of optimizing for engagement on systems where engagement is what is measured. The consequence is that the items most likely to appear in front of you tend to be the items most likely to make you feel something sharp. Algorithms are not deciding to amplify hate. They are deciding to amplify whatever holds attention, and given how attention works on a feed, hostile content holds the eye more reliably than almost anything else.

Together, these features create an environment that does not merely transmit ideas. It selects for the ideas most capable of self-reinforcement and gives them the social conditions in which complex contagion thrives. The platform is not the pathogen. It is what oncologists call a promoter—not the original cause of the disease, but the environment that lets it grow. The mutation is one thing; the conditions under which the mutated cell proliferates rather than dies off are another, and the second is where platforms come in. This vocabulary helps disambiguate a debate that has often been muddled. Social media did not invent hate. Bigotry is older than the internet. What social media provides are the conditions under which a small cluster of bigoted users can become a self-sustaining ecology that recruits, retains, and exports.

Inside the ecology: policing and escalation

The complex contagion picture comes into sharpest focus when we look at what happens inside online communities organized around shared identity. Once a community has formed, two mechanisms make hostile content self-sustaining: policing and rhetorical escalation.

Policing enforces conformity to the community’s norms. Status, in an online community, is allocated by how convincingly you speak its language—its slang, its references, its loathings. Dissent is punished, sometimes by abuse, more often by ostracism: the withdrawal of the small dopaminergic rewards of likes and replies and recognition. Members learn quickly which moves earn approval and which do not, and they adapt. The mechanism is not new—every subculture polices its members—but the platforms have given it instruments of unusual precision. The like count is a real-time signal of community approval; its absence is a real-time signal of withdrawal. The correction is visible, in numbers, beneath every post.

Escalation follows. If status is granted by adherence to the group’s norms, and the norms are antagonistic to some outgroup, status can be gained by adopting more extreme versions of those norms. The user who is just a little more cutting, a little more dehumanizing, a little more contemptuous than her peers will tend to be rewarded. Over time this produces an arms race in which the rhetorical center of the community drifts further from the mainstream, not because anyone planned it, but because everyone, individually, is responding to local incentives. We have all watched some online subculture undergo this drift. Often it is invisible to participants until they look back at where the conversation was five years earlier and find they cannot quite explain how they got here.

What makes the picture distinctively troubling is that policing and escalation transform a community of users into something more like a self-sustaining ecology of infection, capable of growth, capable of seeding adjacent communities, capable of pushing fragments of its rhetoric into broader public conversation through memes, in-jokes, dog-whistles, and viral confrontations. Mainstream culture then has to absorb material that originated in environments most of us would never knowingly enter. The lexicon of manosphere incel forums, for instance, is now legible to people who have never read an incel post in their lives, because fragments of it have traveled. That travel is what the ecology was built to do.

That process moves through stages. Exposure: a user, scrolling, encounters the community’s content. Engagement: a like, a reply, a click that the algorithm registers. Community contact: the feed begins surfacing more, and the user finds themselves among others who share the response. Participation: the user begins posting in the community’s idiom. Commitment: what began as borrowed vocabulary has become the user’s own framework for understanding the world, and challenges to it are felt as challenges to who they are. None of these steps is irreversible. But each makes the next more probable, and the architecture of the platform smooths every transitions.

What this implies

Once the problem is framed in these terms, several familiar arguments about online speech begin to look less persuasive.

The first is the argument that online hate is essentially the offline problem with a faster delivery mechanism. The complex contagion picture suggests something stronger: the medium reshapes the speech act itself. A racist remark muttered in a pub dies in the air. The same remark on a platform is preserved, searchable, addressed to an audience selected by algorithm for receptiveness, embedded in a status economy that rewards more of the same. These are different acts. The equivalence only holds if we idealize what offline speech was already doing. 

The second is the framing of platforms as neutral conduits. If the platform is a promoter—if its architecture systematically converts ambient hostility into self-sustaining communities of infection—then questions about responsibility cannot be answered by pointing to the individual user who pressed post. The user is the immediate cause; the platform is the environment that made the harm reliably reproducible. We do not excuse tobacco companies because the smoker lit the cigarette. The analogy is imperfect, as analogies are, but the basic point holds: where an agent designs and profits from conditions that systematically produce a harm, that agent is part of the causal story whether or not they pulled the final trigger.

The third objection comes from a libertarian direction: restricting online speech means restricting freedom of expression, and that is what the principle rules out. The objection deserves a careful answer. The history of state speech regulation gives the libertarian real reasons for caution. But the objection still does not work, because it relies on an ambiguity in what “freedom of expression” means. 

Mill gave the most influential defense of the principle, and his argument depends on a particular picture of how speech produces its goods. Many speakers contribute to a shared conversation on roughly equal terms. Errors get exposed. Claims get tested. Audiences are in a position to weigh competing arguments and reach their own conclusions. Speech is valuable, on this picture, because speech under those conditions makes collective reasoning possible. The presumption against interference is grounded in those goods. Remove the conditions and you remove the grounds for the presumption.

The architecture of contemporary platforms does not look like Mill’s picture. Algorithms choose which speakers reach which audiences. Engagement is rewarded above accuracy. Coordinated hostility can prevent some speakers from being heard at all. These are not surface features; they affect the very mechanisms—open competition, fair uptake, comparable footing—that Mill’s argument relies on.

This puts the libertarian in a bind. “Freedom of expression” can be read in two ways, and neither rescues the objection. On the first reading, freedom of expression is the negative liberty to speak without state interference. Regulating platforms restricts that liberty, so the libertarian objects. But a coordinated pile-on that silences a speaker is also a restriction on her liberty to speak, and the libertarian has given no reason to count one as a restriction and not the other. Read this thinly, the principle condemns both, or neither. On the second reading, freedom of expression names the substantive value Mill was actually defending: speech as the mechanism through which a society reasons together. Read this way, the present platform architecture is already restricting it. Reforming the architecture would not damage free expression; it would restore the conditions free expression needs. 

Either reading collapses the libertarian’s central move. Defending the platforms as currently designed and defending free expression turn out to be different projects. None of this tells us what to regulate or how, and the worries about who decides remain real. What the argument does establish is more limited but still important: the libertarian objection cannot, on its own, settle the question against intervention. It depends on background conditions that the phenomenon we are discussing has already taken away.

The argument is not for any particular policy. I have gestured elsewhere towards mitigation strategies, including tax-based instruments that internalize some of the social cost of disseminated harmful content rather than criminalizing individual speech. The idea is that a platform profiting from amplifying harm should bear some of the cost of the harm amplified, and that the threat of financial liability is a sharper lever on corporate behavior than the prospect of individual prosecution. Whether such instruments would work in practice is an empirical question. The conceptual point is prior to it: until we have an accurate model of what online platforms are doing to public discourse, we will keep reaching for interventions calibrated to the wrong disease.

Speech remains the oil of social cooperation. But the machines through which most of us now do most of our speaking were not engineered with cooperation in mind. They were engineered to capture attention, and attention is captured most reliably by what divides us. If we want a public sphere that helps us understand each other, we will need to design, and demand, something different from the one we have. The first step is to see clearly what kind of contagion we are actually dealing with.