What social media actually shows is which desires an algorithm can amplify

A companion claim travels with the first and needs correcting here. Recent psychoanalytic research is said to have shown how ranking and personalisation systems reorganise desire and enjoyment. There is a literature: Luca M. Possati's The Algorithmic Unconscious (Routledge, 2021), and Jacob Johanssen's Psychoanalysis and Digital Culture (Routledge, 2019), which argues through Freud and Didier Anzieu that digital media shape subjectivity at unconscious levels. Both are interpretive, and neither establishes nor claims an empirical result about what ranking algorithms do to desire. Adjacent questions run through our coverage of autonomy and automation.

None of which makes the recurrence unreal. Certain narrative shapes do keep surfacing, across platforms and languages, and the question is what explains it. Jung's answer, that they well up from an inherited substrate, is the least tractable available. Two older sociological frameworks do better.

What Durkheim and Castoriadis buy you that Jung does not#

Émile Durkheim's collective representations are not buried, and that is the point of them. In the introduction to The Elementary Forms of the Religious Life, in Joseph Ward Swain's translation, he writes that "collective representations are the result of an immense co-operation, which stretches out not only into space but into time as well; to make them, a multitude of minds have associated, united and combined their ideas and sentiments; for them, long generations have accumulated their experience and their knowledge" (pp. 15–16). A few lines later: "Society is a reality sui generis; it has its own peculiar characteristics, which are not found elsewhere and which are not met with again in the same form in all the rest of the universe."

The methodological gain is immediate. Collective representations are public, external and countable; they live in rituals, symbols, calendars and speech, not in a hypothesised psychic basement. A meme is a collective representation in exactly Durkheim's sense: a compressed symbol whose function is to mark membership. You can date its emergence, measure its spread and watch it die. You cannot do any of that with an archetype.

Cornelius Castoriadis pushes further, in the direction the platform case requires. In The Imaginary Institution of Society, published in French in 1975 and translated by Kathleen Blamey for MIT Press in 1987, he insists that the imaginary "is the unceasing and essentially undetermined (social-historical and psychical) creation of figures/forms/images, on the basis of which alone there can ever be a question of 'something'" (p. 3). The stress falls on creation. The social imaginary does not picture a prior reality; it institutes reality. A society does not harbour significations waiting to be uncovered. It makes them, and then lives inside them as though they were given.

Apply that to a feed and the question changes shape. You stop asking what the platform discloses and start asking what it institutes: which significations it makes cheap to produce and which expensive to sustain. Castoriadis, who spent a working life arguing against the idea that history has an underlying logic to be read off, would have found "content moderation" a peculiar name for what is going on. The interesting power is the making, not the deleting.

The best recent evidence is about amplification, and it is quite specific#

In October 2025, Jacopo D'Ignazi, Andreas Kaltenbrunner, Gaël Le Mens, Fabrizio Germano and Vicenç Gómez posted "Rewarding Engagement and Personalization in Popularity-Based Rankings Amplifies Extremism and Polarization", later published by the 32nd ACM SIGKDD conference. They combine a formal model of the user-algorithm loop, simulations calibrated on behavioural data, and experiments with human subjects facing rankings that update as they click.

Their mechanism rests on four separately documented assumptions: "(i) users exhibit position bias and tend to prefer items displayed higher in the ranking, (ii) users prefer like-minded content, (iii) users with more extreme views are more likely to engage actively, and (iv) ranking algorithms are popularity-based, assigning higher positions to items that attract more clicks." The conclusion is stated flatly: "when platforms additionally reward active engagement and implement personalized rankings, users are inevitably driven toward more extremist and polarized news consumption."

Notice what that is not saying. It does not say people secretly want extremity. It says people with extreme views click more, that a system counting clicks over-samples them, and that personalisation compounds the error by giving each group its own version of the over-sample. The asymmetry sits in the instrument, not the population. Here is the modest claim at full strength: the archive records a society's desires in proportion to how vigorously they are expressed, not how widely they are held, and those are different quantities.

The echo-chamber evidence is messier than either camp admits#

Anyone invoking filter bubbles as settled fact has not read the 2023 results, which surprised the people who ran them. That July, four papers from the US 2020 Facebook and Instagram Election Study appeared simultaneously in Science and Nature, produced by academics working with Meta data under a pre-registered protocol.

Brendan Nyhan, Jaime Settle, Emily Thorson, Magdalena Wojcieszak and colleagues, in Nature 620, 137–144, reduced exposure to like-minded sources by about a third for 23,377 Facebook users between 24 September and 23 December 2020. Like-minded content had been a median 50.4 per cent of what users saw, against 14.7 per cent cross-cutting. The intervention worked as designed: cross-cutting exposure rose, uncivil language fell. The attitudes did not move. The paper's summary is unhedged: the changes "had no measurable effects on eight preregistered attitudinal measures such as affective polarization, ideological extremity, candidate evaluations and belief in false claims."

Andrew Guess, Neil Malhotra and Jennifer Pan, in Science 381, replaced the algorithmic feed with a reverse-chronological one for three months. Users saw considerably more from untrustworthy sources, up 68.8 per cent on Facebook, and spent markedly less time on the platform. Polarisation was unchanged. A companion paper removing reshares cut exposure to political news by more than half without shifting attitudes either. A fourth, led by Sandra González-Bailón and David Lazer, found a genuine asymmetry: 97 per cent of URLs rated false by fact-checkers reached more conservative than liberal audiences.

Then the complication. On 27 September 2024, Chhandak Bagchi, Filippo Menczer, Przemyslaw Grabowicz and colleagues published an eLetter in Science, "Social media algorithms can curb misinformation, but do they?", noting that the experiment ran while Meta had emergency "break glass" measures against voter-fraud misinformation switched on. Those measures, they argue, "systematically reduced the amount of misinformation in the control group of the study, which was using the news feed algorithm." The control condition was the algorithm on its best behaviour. Guess and colleagues replied that this may bear on how far the results generalise but not on their validity, and both sides ended by asking for better access to platform data.

Three months of feed adjustment during one election in one country, with a control group the platform was quietly tuning, tells us little about whether a medium reshapes a psyche over a decade. It does tell us the crude filter-bubble thesis, in which exposure drives attitude, is not holding up. We have traced that argument in our reporting on the limits of the polarisation debate.

The adolescent mental-health dispute is a lesson in how little is known#

Jonathan Haidt's The Anxious Generation: How the Great Rewiring of Childhood Is Causing an Epidemic of Mental Illness appeared from Penguin Press on 26 March 2024 and has sold over two million copies. Three days later, Nature ran Candice Odgers's review, at 628, 29–30. It is not a polite disagreement. "The book's repeated suggestion that digital technologies are rewiring our children's brains and causing an epidemic of mental illness is not supported by science," Odgers writes, adding that "hundreds of researchers, myself included, have searched for the kind of large effects suggested by Haidt. Our efforts have produced a mix of no, small and mixed associations."

Her formulation is worth keeping: "Two things can be independently true about social media. First, that there is no evidence that using these platforms is rewiring children's brains or driving an epidemic of mental illness." The second is that the platforms need reform anyway, on grounds that do not require the epidemiological claim.

On the sceptical side sits the largest dataset anyone has brought to bear. Matti Vuorre and Andrew Przybylski, in Clinical Psychological Science in November 2023, examined well-being data on roughly two million people across 168 countries between 2005 and 2022, alongside internet and mobile broadband adoption. "We looked very hard for a 'smoking gun' linking technology and well-being and we didn't find it," Przybylski said on publication. No subgroup, including adolescent girls, behaved differently.

That is not where the matter rests. In 2026, Liam Sigaud, Zachary Rausch, Alec McClean and Haidt published a commentary in the same journal, "Why Three Studies by Vuorre and Przybylski Should Not Be Used to Assess the Impact of Social Media on Adolescent Mental Health". Their objection concerns aggregation: the studies, they argue, rely on "blending together groups and technologies while dismissing consistent evidence of harms when those groups and technologies are unblended", and Meta was involved in the research design of one. It has force. A population-level null is compatible with substantial harm concentrated in a minority, and lumping broadband adoption in with Instagram use is a coarse instrument for a fine question.

Haidt's causal story runs ahead of its evidence, and Odgers is right to say so. The sceptics' aggregate nulls are weaker refutations than they are taken to be, and Sigaud and colleagues are right about that. Anyone reporting this literature as settled in either direction is reporting something other than the literature.

The like is a ballot, and that is the problem with treating it as one#

The most useful recent account of what this does to public life comes not from a philosopher but from a political sociologist. Paolo Gerbaudo, in "Theorizing Reactive Democracy", Democratic Theory 9(2), 120–138, argues that the online public sphere is "a 'plebeian public sphere' whose functioning is markedly different from the traditional public sphere, described by Jürgen Habermas". It is dominated "by online crowds that come together in virtual gatherings made visible by a variety of social media reactions and metrics that measure their presence", producing "a plebiscitary form of democracy in which reactions are understood as an implicit vote indicating the mood of public opinion on a variety of issues."

His sharpest line is his plainest: "Given that the number of people reacting to content becomes an implicit measurement of its 'popularity', every reaction becomes a ballot of sorts: one 'like' equals one vote." He means this descriptively, and he is right that this is how reactions are read, by journalists and politicians as much as by the platforms. He is also describing a franchise with no roll, no returning officer and a counting method kept as a trade secret.

Jürgen Habermas revisited his own 1962 argument for these conditions in Theory, Culture & Society 39(4), 145–171, in 2022. His diagnosis is structural rather than moralising: digitalisation is "leading to the expansion and fragmentation of the public sphere and is turning all participants into potential authors", with the threat falling on "the traditional role of the public sphere in discursive opinion and will formation in democracies". Not everyone finds Gerbaudo's framing adequate. Antoine Sander, in Constellations in 2025, credits him with recovering counterpublics but faults him for "limiting himself to noting a 'resemblance' between the plebs of ancient Rome and the crowds gathered online" without supplying a mechanism.

Debord's spectators now do the producing, and Zuboff says who owns the output#

Guy Debord published La société du spectacle in 1967. Thesis 4, in the Black and Red translation, supplies the correction always needed: "The Spectacle is not a collection of images, but a social relation among people, mediated by images." Thesis 30 states the cost: "The more [the spectator] contemplates the less he lives; the more he accepts recognizing himself in the dominant images of need, the less he understands his own existence and desires."

Debord's spectator watched. The change worth naming is that the spectator now also supplies the material, unpaid, and competes for position in a ranking that decides whether anyone sees it. Here the spectacle acquires a balance sheet; the account of how activity comes to stand over against its producer is in Marx on alienation.

Shoshana Zuboff supplies the accounting. The definition at the front of The Age of Surveillance Capitalism (2019) opens by calling it "a new economic order that claims human experience as free raw material for hidden commercial practices of extraction, prediction, and sales", and its second clause names "a parasitic economic logic in which the production of goods and services is subordinated to a new global architecture of behavioral modification". In the introduction she sets out the operation: "Although some of these data are applied to product or service improvement, the rest are declared as a proprietary behavioral surplus, fed into advanced manufacturing processes known as 'machine intelligence,' and fabricated into prediction products that anticipate what you will do now, soon, and later."

By the standard applied to everyone else here, "behavioral modification" claims more than has been demonstrated, and the Meta experiments are, awkwardly for Zuboff, among the better evidence that the modification is weak at the level of political attitude. They do not touch her real point, which concerns ownership: whatever the predictive apparatus is worth, it was built from material nobody was asked to surrender and nobody was paid for.

The strongest case for the platforms is a case about visibility#

The version of this argument that circulates most widely skips the part where social media did something that could not be done before, and the omission weakens it. On 25 May 2020, Darnella Frazier, then seventeen, stood on a Minneapolis pavement and filmed the murder of George Floyd on a telephone. In June 2021 the Pulitzer board gave her a special citation. With a camera, a network and no editor between the two, she overrode the official account, a police statement describing a man who died after a medical incident during an interaction with officers.

Tarana Burke had been using the phrase "me too" since 2006. When Alyssa Milano posted "If you've been sexually harassed or assaulted write 'me too' as a reply to this tweet" on 15 October 2017, Twitter told CBS News nine days later that the hashtag had appeared in more than 1.7 million tweets, with 85 countries registering at least a thousand each. Whatever one thinks of what followed, the aggregation was new: a self-executing census of an experience institutions had spent a century declining to count.

Jacques Rancière gives this its sharpest formulation. In The Politics of Aesthetics, translated by Gabriel Rockhill for Continuum in 2004, he writes: "I call the distribution of the sensible the system of self-evident facts of sense perception that simultaneously discloses the existence of something in common and the delimitations that define the respective parts and positions within it" (p. 12). Politics, on this account, is a quarrel over who registers as a speaking being at all. The platforms genuinely redistributed the sensible. They put a seventeen-year-old's telephone footage in front of a jury and a hundred million strangers.

The redistribution runs on the same mechanism as everything else, which is why the González-Bailón asymmetry and the D'Ignazi findings matter. A system that raises whatever engages will raise Frazier's video and a fabricated atrocity with equal indifference, because it is evaluating neither. It is counting. Our reporting on what voters actually see takes up the electoral version.

What survives#

Nothing in the record supports the claim that a society's unconscious has become legible. What it supports is narrower and harder to dismiss. A ranking function that rewards engagement is a selection device, and it selects for intensity of expression rather than breadth of holding. Fed through it, a minority position held fiercely will outrun a majority position held calmly, and the output gets reported as the mood of the public, including by people who should know better.

The question, then, is not what the unconscious is. It is who operates the selection, on what parameters, and whether anyone outside the company may inspect them. The answer to the last is no, which is why the 2023 studies needed Meta's cooperation to exist at all and why their control condition turned out to have been altered mid-experiment without the researchers knowing. A society that reads its own amplified echo as its voice has mistaken an instrument reading for a self-portrait, and it cannot see the instrument.

Martin Heidegger, Michel Foucault