TL;DR
Holmes’s “free trade in ideas” makes a testable claim: let ideas compete and the better ones win, so counterspeech beats censorship.1 The first premise fails. Across ~126,000 stories tweeted by ~3 million people, falsehood diffused significantly farther, faster, deeper and more broadly than truth in every category.2 But the popular explanation fails too: robots accelerated true and false news at the same rate, meaning humans, not bots, are why false news travels.2 And the flood is not a flood. In the 2016 US election, 1% of individuals accounted for 80% of fake-news source exposures and 0.1% accounted for nearly 80% of the sharing, while across the political spectrum most political news exposure still came from mainstream outlets.3 A separate study concluded that speculation about the prevalence of exposure to untrustworthy websites “has been overstated.”4 The marketplace is failing. It is failing in a small, identifiable place, and most regulation is aimed somewhere else.
A stack of newspapers. The old marketplace had a small number of large sellers; the new one has a very large number of very small ones, and the argument has not caught up. Photo: AbsolutVision on Unsplash.5
A Model That Can Be Checked
The marketplace of ideas descends from Mill and enters law through Holmes’s dissent in Abrams v. United States (1919), where he argued for a free trade in ideas and held that the test of truth is a thought’s power to get itself accepted in competition.1
It is worth noticing that this is not merely a moral position about liberty. It contains an empirical prediction: in open competition, true claims will outperform false ones. That prediction licenses the whole policy structure built on top of it — a preference for counterspeech over suppression, a suspicion of state arbitration, a tolerance for a great deal of garbage on the grounds that it will lose.
If the prediction is wrong, the policy does not automatically flip. But it stops being a description of how things work and becomes something you have to defend on other grounds.
The Prediction Is Wrong
Vosoughi, Roy and Aral examined the diffusion of every verified true and false story distributed on Twitter from 2006 to 2017 — about 126,000 stories, tweeted by roughly 3 million people more than 4.5 million times, classified using six independent fact-checking organisations that agreed with each other 95–98% of the time.2
Falsehood diffused significantly farther, faster, deeper and more broadly than the truth in all categories of information, and the effect was strongest for political news.2 Their proposed mechanism is not that people are stupid: false news was more novel than true news, and people are more likely to share novel information.2 Falsehood has an advantage in the competition because it is unconstrained by what happened, and novelty is the currency.
So the marketplace does not reliably select for truth. On this evidence it partially selects against it.
But Not for the Reason You’ve Been Told
Here is the finding that gets dropped when this study is cited, and it is the most consequential one in the paper.
Robots accelerated the spread of true and false news at the same rate.2 The authors’ own inference: false news spreads more than the truth because humans, not robots, are more likely to spread it.2
That single result cuts the legs from under the dominant public account of the problem. If the story is foreign troll farms and automated amplification, then the fix is detection, takedown and platform enforcement, and the population is a victim of manipulation. If the story is that ordinary people prefer novel things and falsehood is more novel, then the mechanism is demand-side, sits inside normal human behaviour, and is not something you can moderate away by finding the bots.
Bots exist and do real work. They are not why this happens.
And the Flood Is Not a Flood
The second half of the popular picture is that we are all swimming in this stuff. That claim has been measured too, and it does not survive.
Grinberg and colleagues matched Twitter accounts to registered voters for the 2016 US election and found engagement with fake-news sources was extremely concentrated. Only 1% of individuals accounted for 80% of fake-news source exposures. Only 0.1% accounted for nearly 80% of the fake-news sources shared. The people most likely to engage were conservative-leaning, older, and highly engaged with political news — and crucially, for people across the political spectrum, most political news exposure still came from mainstream media outlets.3
Guess, Nyhan and Reifler reached a compatible conclusion by a different route, using survey and web-traffic data. Supporters of one candidate were most likely to visit factually dubious websites, but those sites made up a small share of people’s information diets on average and were largely consumed by a subset of Americans with strong preferences for information that already agreed with them. Their summary is blunt: the widespread speculation about the prevalence of exposure to untrustworthy websites has been overstated.4
What the Two Findings Mean Together
Put them side by side and you get a picture that neither side of the public argument is describing.
The alarmed version says the information environment is drowning in falsehood and democracy is being dissolved by it. The complacent version says the marketplace is working and the panic is a moral panic about new media. The data supports a third thing.
Falsehood genuinely outcompetes truth in open circulation — the marketplace’s core premise is empirically false. And the resulting exposure is concentrated into a very small, identifiable, self-selecting population, most of whose members were already committed to the direction the falsehood points. The problem is real and it is small and dense rather than broad and thin.
That matters because it changes the target. A regulatory regime designed on the assumption that the average citizen is being deceived — mass fact-checking labels, broad takedown mandates, media-literacy campaigns for everyone — is aimed at a population that is mostly reading mainstream outlets. Meanwhile the 0.1% doing most of the sharing are the least likely people in the country to be moved by a fact-check.
It also means counterspeech is not obviously the answer. Counterspeech works when the audience is undecided and listening. Its weakest case is precisely the concentrated, high-engagement, strongly pro-attitudinal group where the exposure actually lives.
Where I’d Hold This Loosely
Four limits, and two of them are substantial.
The concentration studies are 2016, US, and largely Twitter and desktop web traffic. That predates TikTok’s rise, the collapse of Twitter’s research API, and generative AI making fabricated material cheap and personalised. Whether 1%-of-users-80%-of-exposure still holds is genuinely unknown, and I would not assume it does.
Second, concentration is not the same as harmlessness, and this is the caveat I’d most want kept. A small group can matter enormously if it is politically active, well-organised, or violent. “Only 1% were heavily exposed” is a statement about distribution, not about consequences, and the two get conflated in exactly the direction that flatters complacency.
Third, Vosoughi’s true/false classification depends on stories that fact-checkers chose to investigate. That is not a random sample of claims — it is skewed toward the contested and the viral, which is the population where you would most expect the effect. The finding is robust within that set; generalising it to all information is a step beyond the data.
Fourth, none of this settles the regulatory question. Showing that Holmes’s empirical premise is false does not establish that a state arbiter would do better, and the case against government adjudication of truth never rested mainly on the marketplace working. It rested on what happens when it doesn’t and someone has to decide. That argument is untouched by any of this — which is roughly the same shape as the observation that agreement is weak evidence: knowing a mechanism is unreliable tells you to stop trusting it, not who to trust instead.
