Platforms, Algorithms and the Public Sphere

The findings of the last lesson were established under conditions that have dissolved: a handful of channels, a shared audience, professional gatekeepers, and a clear separation between those who produced content and those who consumed it.

All four are gone. And the confident claims about what replaced them — filter bubbles, echo chambers, algorithmic radicalisation, an epidemic of misinformation swinging elections — have run so far ahead of the evidence that separating the two is the whole task of this lesson.

The honest summary, given first: the structural changes are real and large; the psychological and political effects claimed for them are, on the best current evidence, considerably smaller and more conditional than the public discussion assumes.

The experiment that turned off the algorithm.

A recurring belief holds that social media algorithms trap people in filter bubbles — feeds curated to show only what confirms existing views, sealing users into information worlds that never challenge them, deepening division and detaching people from a shared reality.

It is a plausible mechanism, it is enormously influential, and it has been tested.

In the run-up to a major election, researchers partnered with a large platform — with independent oversight and pre-registered analysis (see 7.4.2, 7.8.2) — to run genuine randomised experiments on the algorithm itself. Consenting users were randomly assigned to have their algorithmically ranked feed replaced with a simple reverse-chronological feed — everything from the accounts they followed, in time order, with no algorithmic curation — for three months.

If the algorithm were the engine of polarisation and misinformation, removing it should have measurably reduced both.

It did not.

Turning off the algorithm changed what people saw — they saw more content from untrustworthy sources, in fact, and less from moderate sources, on the chronological feed — but it did not detectably change their political attitudes, their affective polarisation, their factual beliefs, or their political participation over the study period.

Related experiments in the same collaboration found the same pattern. Reducing exposure to like-minded content did not move attitudes. Removing reshared content changed what people saw and knew about current events but not their broader political views.

Read this carefully, and read it with appropriate caution in both directions.

The caution against over-reading it as vindication of the platforms: three months is a short window; the studies cannot capture effects that accumulate over years, or the effect of the environment existing at all rather than of a change within it; and the platform was a partner, which constrains what could be studied.

The caution against dismissing it: these are the best-designed studies ever conducted on the question, they were pre-registered and independently overseen, and they directly tested the central mechanism of the dominant theory — and found the mechanism did not produce the effect.

The proportionate conclusion is the one this whole Topic keeps reaching. The confident story — algorithm causes bubble causes polarisation — is not supported at the strength it is asserted. Something is happening; it is not that.

What actually changed, structurally, is not in doubt. The dispute is about effects.

Five structural changes are real, large, and not seriously contested.

The collapse of gatekeeping. Anyone can publish. The professional editorial filter that decided what reached a mass audience is gone for most content most of the time.

Disintermediation and re-intermediation. The old intermediaries (editors, broadcasters) were removed — and replaced by new ones (platforms and their ranking systems) that are far less visible and operate by different criteria (see 9.3.3 on algorithmic management; this is the same mechanism applied to information).

The collapse of the bundle. News was once bundled — you bought a newspaper and got the front page, the sport and the crossword together, cross-subsidising the expensive journalism. Unbundling destroyed the business model that funded reporting, which is why local news has collapsed across many countries.

The measurement of everything. Every interaction is recorded, and content is optimised against engagement metrics (see 7.2.2 on Goodhart — the metric becomes the target, and the target is attention, not truth or value).

And the reversal of the attention economy. When information was scarce, attention was cheap; now information is unlimited and attention is the scarce resource being competed for. This is the structural fact underneath everything else.

None of that is disputed. What is disputed is what these changes do to what people believe, how divided they are, and whether democracy can function — and that is where the evidence and the assertions part company.**

The claims, tested

One — echo chambers and filter bubbles: much weaker than assumed.

The distinction matters. A filter bubble is imposed by an algorithm; an echo chamber is chosen by the user. They are different mechanisms and the evidence differs.

On filter bubbles: the experimental evidence above, plus a body of observational work, finds that algorithmic curation does not seal most people into homogeneous information worlds. The counterintuitive finding is that social media users, on average, encounter more diverse news and more cross-cutting views than non-users — because the weak ties in a large network (see 8.3.3) expose people to more variety than their offline social world does.

On echo chambers: the picture is more mixed and depends heavily on measurement. By some measures most people have politically diverse information diets; by others, a minority — the most politically engaged — do construct highly homogeneous ones. The echo chamber is real for a committed minority and overstated as a description of the general public.

And the crucial finding that inverts the intuition: some studies find that exposure to opposing views, on social media, can increase rather than decrease polarisation — because encountering the other side in a hostile, decontextualised, worst-example form hardens rather than softens views. The problem, where there is one, may be too much exposure to the other side in the wrong form, not too little.

Two — misinformation: real, and smaller and more concentrated than the panic suggests.

Deliberately false content exists, spreads, and is a genuine problem. And the measured scale is consistently smaller than the discourse implies.

Studies measuring actual consumption find that false news is a small fraction of people's overall information diet — for most people, a tiny one. Its consumption and sharing is heavily concentrated in a small, distinctive group: older, more politically extreme, more engaged users share the overwhelming majority of it. The average person encounters little and shares almost none.

This does not mean it is harmless — concentrated effects on a committed minority can matter, and a small share of a vast total is still a large absolute quantity. It means the model of a credulous mass swept along by falsehood is wrong, and the reality — a small, identifiable, motivated group producing and consuming most of it — implies different remedies.

And the finding that most complicates the standard account: the best evidence suggests people believe and share misinformation less because they are deceived and more because it is congenial — it fits and flatters their existing political identity, and they share it as an expression of that identity rather than as a report of fact (see 7.4.1 on considerations; this is identity-expressive, not belief-forming). Which means "correcting the facts" addresses the wrong mechanism, and explains why fact-checking has such limited effects.

Three — polarisation: rising, real, and not obviously caused by the technology.

Affective polarisation — dislike and distrust of the other side, as distinct from disagreement on issues — has risen substantially in several countries, and it is one of the most consequential political facts of the period.

But the causal attribution to social media is weak.

It rose fastest among the demographic that uses social media least — older people — which is difficult to reconcile with a primarily technological cause. It rose in some countries and not others despite similar technology. And it began rising before social media in several cases. The most careful comparative work concludes that the technology is, at most, one contributor among several, and probably not the principal one — with partisan media, political elites and party realignment stronger candidates.

This is 7.6.2 and 7.6.4 operating together: a real and important trend, a plausible technological cause, and a body of evidence showing the cause cannot carry the weight assigned to it.

Radicalisation, recommendation, and a genuinely open question.

The strongest version of the technological claim is that recommendation systems actively radicalise — that autoplay and "up next" algorithms lead users progressively toward more extreme content in a documented pathway.

The evidence is genuinely mixed and this is the area where honest uncertainty is most warranted.

Some studies auditing recommendation systems found that following recommendations from a neutral start does not reliably lead to extreme content, and that consumption of extreme content is better predicted by users seeking it than by algorithms pushing it. On this evidence, the algorithm follows demand more than it creates it.

Other work, and a good deal of qualitative and journalistic evidence, documents real pathways in which recommendation, community and content combined to move specific individuals toward extreme views — and argues that the audit studies, run from clean accounts, miss the effect on already-vulnerable users with existing histories.

Both can be true. The algorithm may not radicalise the average user from a neutral start, while playing a real role for a vulnerable minority with existing susceptibility — which is exactly the shape of finding this Topic keeps producing: small average effects concealing concentrated effects on an identifiable minority (see 7.6.4 on averages that describe nobody).

The honest position is that this is unresolved, that the strongest public claims outrun the evidence, and that the concentrated-minority possibility is the one most worth taking seriously and hardest to study.

The public sphere

Habermas, the ideal, and what the platforms did to it.

The public sphere (see 5.4.2) is the space between the state and private life where private people come together as a public to discuss matters of common concern, ideally through reasoned argument in which the better argument prevails rather than the more powerful speaker.

Habermas's own account was historical and critical: he argued the bourgeois public sphere of the eighteenth century — coffee houses, journals, salons — was progressively hollowed out by mass media and commercial pressure, its rational-critical debate replaced by consumption and manipulation. It was never an achieved ideal; it was a standard against which to measure decline.

And the platform era can be read against that standard in two directions, both of which have evidence.

The optimistic reading: the barriers to participation collapsed. Voices excluded from the old public sphere — which was, as feminist and postcolonial critics noted, restricted by property, gender and race — can now speak, organise and be heard. Movements have been built, abuses documented, and publics assembled that the gatekept media would never have admitted (see 6.7.1, 8.4).

The pessimistic reading: the conditions for rational-critical debate have been degraded. Attention is captured by outrage because outrage is engaging; the shared factual basis for debate has fragmented; anonymity and scale enable harassment that silences; and the economic model rewards heat over light. The space expanded and the quality of what happens in it deteriorated.

Both are supported, and the resolution is not a compromise between them but a recognition that they describe different users and different moments. The same infrastructure that lets a marginalised movement organise lets a coordinated campaign harass; the same openness that admits excluded voices admits bad-faith ones. The technology is not directional — which is the same conclusion 9.4.2 reached about religion and 8.4.3 about organisational resources.

And the concentration underneath it all, which the effects debate can obscure.

Whatever the effects on individuals, a structural fact is not in dispute: a very small number of platforms now mediate a very large share of the world's public communication, and they do so as private companies, optimising against commercial objectives, with ranking systems that are proprietary, changeable without notice, and accountable to no public process (see 9.3.3, 9.5.2 on algorithmic authority).

This is a concentration of infrastructural power over the public sphere with few historical parallels, and it is separable from the question of persuasion effects. Even if the platforms changed no individual's mind, the fact that a handful of firms set the rules of visibility for global public discourse is a sociological fact of the first order — and it is the one that survives every deflation of the effects claims.

The right question is therefore not only "what do the algorithms do to what people believe?" — where the answer is "less than claimed" — but "who controls the infrastructure of public attention, and by what right?" — where the answer is "a few private firms, by none in particular." The second question does not depend on the first, and it is the more durable.

Because the media is where nearly everything else in this course is learned about, and the confident claims about it are wrong in a consistent and identifiable way.

The pattern across both lessons of this Topic is the same. Strong effects are asserted; careful measurement finds them smaller, more conditional, and concentrated in identifiable minorities rather than distributed across a credulous mass; and the deflation is itself resisted, because a story of powerful media serves many interests (those selling advertising, those seeking to regulate, those explaining political outcomes they dislike).

What the evidence supports.

The structural changes are real and large — gatekeeping, bundling, intermediation, measurement and the attention economy have all transformed.

The psychological and political effects are smaller and more conditional than claimed — filter bubbles do not seal most people in, misinformation is concentrated in a small motivated group, polarisation's technological cause is weak, and turning off the algorithm changed little over three months.

And effects, where they exist, are concentrated — on the committed, the extreme, the already-susceptible — which averages conceal (see 7.6.4).

Four questions for any claim about the new media environment.

Is the claim about structure or about effect? The structural claims are usually solid; the effect claims usually outrun the evidence.

Was the effect measured, and on whom? Behaviour or a survey; the average or a concentrated minority.

Does the causal story survive the comparisons? Polarisation rising fastest among the least-online is the model case.

And who benefits from the claim of media power? The oldest question in the Topic, and still the most useful.

One closing observation that holds the two questions apart. The evidence deflates the claim that the platforms are reprogramming individual minds. It does not deflate — it sharpens — the observation that a handful of private firms now control the infrastructure through which societies attend to themselves. The first is a question about persuasion, and the answer is "less than you think". The second is a question about power, and the answer is "more than any institution in the history of communication". Keeping those two apart is the whole of thinking clearly about this.

The conditions that produced the last lesson's findings — few channels, shared audience, professional gatekeepers — have dissolved, and the confident claims about what replaced them outrun the evidence.

Randomised experiments on a major platform replaced the algorithmic feed with a chronological one for three months and found it changed what people saw but not their attitudes, polarisation, beliefs or participation — the best-designed test of the filter-bubble mechanism, finding the mechanism did not produce the claimed effect, with appropriate cautions about the short window and the platform partnership.

Five structural changes are real and uncontested: the collapse of gatekeeping, dis- and re-intermediation, the collapse of the news bundle (and with it local journalism), the measurement of everything against engagement, and the reversal of the attention economy.

The tested claims deflate. Filter bubbles: algorithmic curation does not seal most people in, and social media users encounter more diverse news than non-users through weak ties. Echo chambers: real for an engaged minority, overstated for the public — and exposure to opposing views can increase polarisation when it comes in hostile form. Misinformation: a small fraction of the information diet, concentrated in a small older, extreme, engaged group, shared as identity-expression rather than deception — which is why fact-checking has limited effect. Polarisation: real and rising, but rose fastest among the least-online, in some countries not others, and before social media in several — so the technological cause is weak.

Radicalisation by recommendation is genuinely unresolved: audit studies find the algorithm follows demand more than it creates it, while other evidence documents real pathways for vulnerable minorities — small average effects possibly concealing concentrated ones.

The public sphere can be read optimistically (excluded voices admitted, movements built) and pessimistically (rational-critical debate degraded by outrage, fragmentation and harassment), and both are supported because the technology is not directional.

And the durable fact, separable from all the effects debates, is concentration: a few private firms now control the infrastructure of global public attention, by no public right — a question about power that does not depend on the questions about persuasion, and outlasts every deflation of them.

Filter bubble — algorithmically imposed information homogeneity; weakly supported by evidence.

Echo chamber — user-chosen information homogeneity; real for an engaged minority, overstated for the public.

Disintermediation / re-intermediation — the removal of old gatekeepers and their replacement by less visible algorithmic ones.

Unbundling — the separation of news components that destroyed the cross-subsidy funding journalism.

Engagement optimisation — ranking content by predicted interaction, making attention the target (Goodhart).

Attention economy — the competition for the scarce resource of attention when information is unlimited.

Congenial misinformation — false content shared as identity-expression rather than from deception.

Affective polarisation — dislike and distrust of the other side, distinct from issue disagreement.

Concentrated effects — real effects on an identifiable minority concealed by small averages.

Public sphere — the space of reasoned public debate; a critical standard, degraded and expanded at once.

Infrastructural concentration — private control of the systems mediating public communication; the durable fact.

One — audit your own diet. For one week, note where you encounter news and how diverse the sources are. Compare it with what you assumed.

Two — separate structure from effect. Take a claim about social media and mark whether it is about how the system is built or about what it does to people. The first is usually better supported.

Three — find the concentrated minority. For any claim about misinformation or radicalisation, ask what share of people it actually applies to. The honest answer is usually small and identifiable.

Four — test the causal story. Take the claim that a technology caused a social change. Find one group that uses the technology least, and check whether the change is absent there. If it rose fastest among the least-exposed, the story is in trouble.

Five — hold the two questions apart. Write one sentence about what platforms do to individual belief (answer: less than claimed) and one about who controls the infrastructure of attention (answer: a few firms, by no public right). Notice that the second does not depend on the first.