Media, Polarisation and Misinformation
The standard account is that recommendation algorithms sort people into echo chambers and polarisation follows. The research is more equivocal than that, and the parts that hold up are more specific.
This is a field where the popular narrative and the empirical literature have drifted apart, in both directions. Some widely repeated claims are weakly supported. Some effects that receive less attention are robust. Both are worth stating precisely.
Echo chambers: smaller than advertised
Measured directly, most people's online information diets are more diverse than their offline ones. Social media users encounter more cross-cutting political content than non-users, largely because weak ties supply material that close friends and family would not. The finding is uncomfortable for the standard story and it has replicated across several designs.
Two qualifications keep it from being reassuring. First, the averages conceal a tail: a small, highly active minority does inhabit genuinely closed information environments, and that minority is disproportionately influential in what gets amplified. Second, exposure is not the same as engagement. Encountering opposing content while primed for conflict can harden a position rather than soften it, and several experiments find exactly that.
Polarisation: which kind
Separating the two forms of polarisation resolves much of the apparent conflict in the literature. Issue polarisation - the distance between the policy positions people hold - has moved comparatively little in most measured populations. Affective polarisation - how negatively people feel about the other side - has increased substantially.
The mechanism is not primarily about which arguments people see. It is about which people they see. Platforms surface the most extreme and least representative members of an opposing group, because that content generates engagement. The resulting impression of the other side is systematically distorted, and the distortion is a straightforward consequence of ranking by engagement rather than any intent to polarise.
How false information actually spreads
The robust findings about misinformation are narrower than the coverage suggests:
- Volume is concentrated. A very small number of accounts produce most of it, and a small number of users consume most of it. It is not evenly distributed across a population.
- Novelty drives sharing. False stories spread faster largely because they are more surprising, and surprise is what people forward. This is a property of the content, not of the recipients.
- Correction works but decays. Corrections reduce belief in the false claim, and the effect fades over weeks. Backfire effects, where correction strengthens the original belief, are much rarer than early work suggested and largely failed to replicate.
- Prompting attention to accuracy helps. Simply asking people to consider whether a headline is accurate before sharing reduces sharing of false content measurably. Much misinformation sharing appears to be inattention rather than conviction.
The algorithmic contribution, honestly stated
Field experiments that altered real users' feeds - replacing algorithmic ranking with chronological ordering, or reducing exposure to like-minded content - found smaller effects on political attitudes than expected, over the months they ran. That is genuine evidence and it constrains strong claims about algorithms causing polarisation.
It also has limits worth naming. These experiments ran for months in systems shaped by years of algorithmic curation, so they measure the marginal effect of a change, not the cumulative effect of the system. And attitudes are not the only outcome: what gets produced in the first place responds to what gets distributed, and that supply-side effect is largely outside the design of these studies.
Why the network view helps
Platform curation changes the effective network. It does not remove ties, it reweights them, deciding which of your existing contacts you actually encounter. In the language of opinion dynamics, ranking by engagement narrows the effective confidence bound - it preferentially shows you content you will react to - and narrowing that bound is precisely what fragments a population into non-communicating clusters in bounded-confidence models. That is a specific, testable mechanism, which is more useful than the general claim that algorithms cause polarisation. The measurement problems involved are the ones described in digital traces and privacy.