The screen lights up before you ask it to. Not aggressively, not urgently, just there. Your hand moves without instruction; the thumb already knows the motion. The first image feels familiar. The second irritates. The third confirms something you already thought yesterday. Nothing tells you what to believe. Nothing asks for your agreement. There is only sequence, and after a few minutes the world feels thinner, complexity becomes tiring, and silence feels wrong. You close the app, not because you are finished but because nothing else appears.
That sensation is not a side effect. It is the product, and it is worth pausing on, because almost everything that follows is an attempt to explain why a machine would be built to leave you feeling exactly that, and who benefits when it does. And it is the clearest evidence of a shift that this piece exists to make visible: modern propaganda does not persuade. It arranges perception. The determining variable of information power is no longer the message that is sent but the environment in which all messages appear, stall, escalate, or vanish. Control what is visible when a person thinks, and you no longer need to control what they think. You have moved upstream of the thought.
Propaganda lost its voice and gained an environment
Propaganda once had a voice. Posters, slogans, broadcasts, declared enemies. It announced itself, which meant you could recognise it and, recognising it, reject it. That model is now largely obsolete, not because anyone defeated it but because something more durable replaced it. Contemporary propaganda does not require belief. It requires attention. Not what you think, but what appears when you think. Not truth, but visibility.
Facebook, Google, X, and TikTok did not invent a new ideology. They built the environment in which ideas circulate, and an environment does not argue, it conditions. This is the distinction the whole analysis turns on, so it is worth fixing precisely. Influence operates on a person who is still free to disagree. Infrastructure operates on the conditions under which disagreement is even formed. The platforms are not a louder voice in the old debate. They are the room the debate now happens in, and whoever owns the room sets the acoustics.
There is a lineage here worth naming, because the mechanism is older than the technology. A century ago Walter Lippmann argued that the public's picture of the world is necessarily secondhand, assembled from what reaches it, and that the management of that picture, the manufacture of consent, was already the quiet work of modern government. Decades later Edward Herman and Noam Chomsky formalised it into a propaganda model: in a free society with no censor, the news that reaches people is shaped not by a ministry but by a set of filters, ownership, advertising, sourcing, that determine what becomes sayable without anyone issuing an order. The platforms did not break from that tradition. They automated it, accelerated it, and personalised it to the individual scroll, so that the filtering Herman and Chomsky described at the level of the newspaper now happens, continuously and invisibly, at the level of the feed.
The myth of the autonomous algorithm
The most useful fiction the platforms tell about themselves is that the system runs itself. The word "algorithm" is deployed to suggest an autonomous force, a neutral mathematics that no one is responsible for. This is where the fiction collapses, because what platforms call the algorithm is policy translated into code. Every ranking, every recommendation, every quiet suppression is the outcome of human decisions, not made once at the design stage but made continuously, daily, operationally: what is slowed, what is protected, what is labelled sensitive, what is allowed to surge. These choices are not made by machines. They are made in meetings, and code does not decide so much as execute interests, and interests are never neutral.
This is not a theoretical claim. In 2018 Facebook changed its ranking system to favour what it called "meaningful social interactions," an engagement-based model that boosted content drawing reactions, comments, and shares. The change was framed as a way to bring people closer together. Its own internal research said the opposite. By 2019 a team of Facebook's own data scientists had documented, in memos later disclosed by the whistleblower Frances Haugen, "strong evidence" that the change was producing unhealthy effects on politics and news, rewarding the most divisive, outraged, and misleading content because outrage was what the metric measured. One internal study conceded that the financial incentives the algorithms created were not aligned with the company's stated mission. The point is not that Facebook is uniquely cynical. The point is that a documented, human decision about what to rank produced a more polarised public, the company knew, and the ranking stayed, because the ranking worked, for the metric it was built to serve. The algorithm did not drift into outrage. It was tuned to engagement, and outrage is engagement.
The shield that makes it possible
Behind the ranking sits a legal arrangement that is rarely discussed alongside it but does much of the quiet work, because it answers the question of who is responsible and answers it, in effect, no one. In the United States, a single clause written in 1996, Section 230 of the Communications Decency Act, grants platforms immunity from liability for what their users post while also protecting their right to moderate that content as they see fit. The clause was written for a web of message boards and was meant to let fledgling services host conversation without being sued into oblivion. Three decades later it underwrites the largest information systems in history, and it produces a peculiar and consequential status: the platforms are treated as neutral conduits for the purpose of liability and as active editors for the purpose of control. They get the legal protection of the pipe and the operational power of the publisher, and they are accountable as neither.
Europe took the opposite formal route to a structurally similar place. Rather than immunity, the Digital Services Act imposes duties, but those duties run to the management of "systemic risk" rather than to the accuracy of what circulates, which means the law institutionalises content management without ever making truth the standard. One regime shields the platforms from responsibility and the other conscripts them into a public risk-management role, and both leave the determining variable, the power to arrange what is visible, in private hands operating under public pressure. The legal scaffolding differs across the Atlantic. The outcome converges: an infrastructure of perception that is governed, but not accountable in the way a publisher or a broadcaster or a government would be.
One architecture, four surfaces
The same architecture appears on each major platform, expressed through whatever that platform is built to do.
Facebook demonstrates it through social normality. What remains visible becomes normal; what repeats becomes acceptable; what provokes emotion spreads faster than nuance ever can. This is not accidental but measured: polarising content holds attention, calm content releases it, and a system optimised for retention will, without any operator wishing it, find the emotional register that keeps people scrolling. Facebook did not need to manipulate its users. It only needed to rank them, and social reality shifts through repetition far more than through argument.
Google demonstrates the same logic through knowledge itself. Search is not neutral; search is selection. What appears first is read; what does not appear effectively does not exist. Google does not need to ban information, which would be visible and contestable. It needs only to optimise for relevance, authority, and trust, words that sound objective and conceal a thousand choices about whose authority counts. The proof is in how completely the second page of results has ceased to exist in practice: a link that ranks tenth is read by a sliver of searchers, and a link that ranks fiftieth, however true, is functionally invisible, so the difference between suppression and a low ranking dissolves into the same outcome by gentler means. The features built on top of search deepen it. Autocomplete proposes the question before you finish asking it, nudging the query itself. The boxed answer at the top resolves the question without a click, which means a single source, chosen by the system, becomes the answer for millions who never see the alternatives beneath it. Censorship announces itself and can be fought. Invisibility is final and cannot, because you cannot protest the absence of something you never knew was there. When truth depends on discoverability, knowledge becomes governable, and a search engine becomes epistemic infrastructure: not the library, but the librarian who decides which shelf you will ever reach.
X applies the architecture through speed. It is less a mass platform than an elite interface, the place where journalists, politicians, analysts, and officials react in real time, and where reaction routinely precedes understanding. Trending becomes urgent, urgent becomes real, real demands a response. Speed destroys context, and contextlessness is manageable in a way that careful argument never is. A claim that detonates in an hour and is corrected in a week has already done its work, because the correction reaches a fraction of the audience and arrives in a different emotional weather, when the outrage has moved on to its successor. TikTok pushes the same architecture below cognition entirely, operating beneath language through rhythm, repetition, and imitation. It does not form opinions; it forms reactivity, until complexity feels physically uncomfortable and depth feels suspicious. Four different surfaces, one underlying structure: the management of attention as the management of reality.
Visibility as political power
This architecture becomes unmistakable the moment it touches power, and three domains show it under increasing pressure.
Elections are no longer decided only at the ballot box; they are prepared in the feed. Timing determines momentum, momentum determines the appearance of legitimacy, and large accounts and institutional voices stay reliably visible while deviating narratives quietly lose reach, not banned but ignored, which is more effective because it leaves no martyr and no evidence. Voters still choose freely. They do not see freely, and the difference between the two is the space in which this kind of power lives. It is what the political theorist Steven Lukes called the third face of power: not winning the argument, and not even keeping an issue off the agenda, but shaping the field so that certain questions never form in the first place.
War shows the same pattern under higher pressure. Images circulate, context collapses, emotion replaces analysis, and platforms stabilise a dominant narrative not because one side is always right but because instability itself threatens the political and economic interests the platforms are embedded in. War does not require truth from its information environment. It requires coherence, and coherence can be enforced through ranking long before it is enforced through law.
The pandemic made the mechanism explicit in a way that is now documented and, in places, conceded. Reach was aligned with what was designated responsible behaviour; visibility was aligned with prevailing policy. Some of what was suppressed was indeed false, and some of what was amplified was correct, but that was not the operative criterion, and the proof is in the reversals: the claim that the virus was man-made, which Facebook removed as misinformation in early 2021, was explicitly re-permitted in May 2021 once the official posture toward a possible laboratory origin shifted. The content had not changed. The policy had, and visibility followed the policy. In an emergency, information stops behaving like a right and starts behaving like an instrument, and the criterion that governs it quietly becomes compliance.
How a thing becomes true enough to act on
There is a deeper layer beneath visibility, and it is the one the older propaganda analysis could not quite reach, because it is not about which message wins but about how anything comes to count as real in the first place. A claim does not need to be proven to govern behaviour. It needs only to become credible, and credibility, in this environment, is manufactured by the same machinery that manages visibility. Repetition supplies it: a thing seen often enough acquires the texture of the known. Proximity supplies it: a claim surrounded by trusted sources borrows their authority. Velocity supplies it: a claim that arrives before its refutation owns the ground the refutation must later try to recapture, and the refutation always arrives second, to a smaller audience, because corrections are not engaging.
This is the production of truth as a process rather than a fact, and it is the missing step between what is visible and what people will act on. The platform does not have to tell you that something is true. It has only to make it ambient, to surround you with it until disbelief would require effort you do not have, and effortful disbelief is exactly what an exhausting environment is designed to deny you. A society's sense of what is real is, in the end, a function of what it repeatedly encounters, and the architecture has taken control of what is repeatedly encountered. That is a more fundamental power than persuasion, because persuasion still grants you the dignity of a choice. This operates one level below the choice, on the raw material from which the choice is made.
The discipline to keep here, the one that separates this from the ordinary complaint that "the media lies," is that the mechanism requires no lie. The individual posts can each be true. The selection, the ordering, the relative volume, the timing, that is where reality is shaped, and none of it shows up as a falsehood that could be fact-checked. You cannot debunk a ranking. There is nothing in it to be false. That is precisely why it is the more durable form of power: it has moved the lever from the content, where it could be challenged, to the arrangement, where it cannot.
Decision, not drift
Nothing described so far happened by accident, and this is the claim the system most wants to obscure, because drift cannot be held responsible and decisions can. These outcomes were not stumbled upon. They were chosen, repeatedly. Someone approved the thresholds. Someone signed off on the exceptions. Someone decided which risks were acceptable and which were not, and then decided again the next quarter. The system does not persist because no one controls it. It persists because too many interests depend on it, and at that point continuation stops being negligence and becomes policy by default.
It must be said carefully, because the careless version of this is a conspiracy theory and the careful version is sturdier. There is no master plan and no single controller. There is something more durable than a plan: a set of aligned incentives, held by different actors who never need to coordinate, that consistently reward control and punish openness. That is the determining variable restated. You do not need a conspiracy when the architecture, left to run, produces the managed outcome on its own.
The proof that these are decisions and not weather sits, conveniently, in the companies' own files. In 2021 a Facebook product manager named Frances Haugen walked out of the building with tens of thousands of pages of internal research and handed them to regulators and journalists. What those pages showed was not a smoke-filled room but something more banal and more damning: teams of the company's own researchers, in sober internal memos, documenting that the ranking changes were harming the public sphere, recommending fixes, and watching those fixes lose to the metric. The harm was not hidden from the company. It was studied by the company, written up by the company, and then declined by the company, because the version of the product that was worse for the public was better for engagement. That is the entire mechanism caught in a single disclosure: not malice, but a documented choice to keep a system its own scientists had flagged, because the incentive pointed one way and the conscience pointed the other and the incentive won, quarter after quarter, in meetings no outsider ever saw. The memos are the receipts. They exist because, for once, someone carried them out.
The powers that converge
Three forces meet in this architecture, and none of them has to conspire with the others, because their incentives already point the same way.
The first is the state, which is not an external regulator of the system so much as an embedded user of it. Law is slow and platforms are immediate, so normalisation precedes legislation, and policy ends up following what already feels normal. Coordination replaces command; labels replace debate. Once an issue is framed as a matter of security, safety, integrity, or health, a different set of rules quietly applies, because these labels are not descriptions. They are switches. And the formal machinery is now visible: the European Union's Digital Services Act, in force from late 2022 with its duties on the largest platforms applying from 2023, requires those platforms to assess and mitigate "systemic risks" on pain of fines reaching six percent of global turnover, which means a public authority now sets the terms within which private platforms decide what may circulate. Whatever one thinks of its aims, its structure is the point: it institutionalises the management of visibility as a matter of law.
The second force is the advertiser, and it may be the most decisive, because it funds everything. Platforms are not paid for by their users; they are paid for by advertisers, and advertisers do not demand a particular ideology. They demand predictability. They do not care what people believe; they care whether belief destabilises markets, brands, or consumption. This is not abstract. In 2020, under the banner Stop Hate for Profit, more than a thousand advertisers, over four hundred of them major brands including Coca-Cola, Ford, Adidas, and Unilever, paused their Facebook advertising to force tighter content rules, and Facebook moved. In 2023, after ads appeared beside extremist content on X, Disney, Apple, IBM, Comcast and others pulled their spending, costing the platform a reported fortune. In both cases the lesson for every platform was identical: brand safety is an economic veto. Topics associated with instability, systemic critique, or institutional distrust are rarely removed outright. They are deprioritised, because what cannot be monetised safely cannot circulate freely, and the advertiser enforces that boundary without ever having to name it.
The third force is security doctrine, which now treats information as a battlespace, perception as terrain, and narrative as advantage. Military and intelligence establishments across the world have stood up dedicated information-operations capabilities and openly describe the "cognitive domain" as a theatre of conflict alongside land, sea, air, space, and cyber. Victory in this frame no longer requires occupying ground; it requires perceptual advantage, and perceptual advantage is exactly what control of visibility delivers. None of the three forces commands the others. The platform wants engagement, the advertiser wants safety, the state and its security apparatus want a stable and steerable narrative, and these are not the same goal. They simply overlap, at the level of architecture, on a single shared preference for an environment that is managed rather than open, and a managed environment is what their convergence reliably produces, without any of them ever having to sit in a room with the others.
The consequence: fatigue as governance
Modern power, arranged this way, no longer seeks your consent. It seeks your participation. Scrolling is enough. Reacting is enough. Legitimacy is inferred from engagement, and engagement does not require that you approve, only that you do not look away. The system does not aim to convince you forever. It aims to exhaust you. Outrage shortens the attention span; memory collapses under the next refresh; and what disappears is not any particular fact but continuity itself, the ability to hold an idea long enough to let it mature, clash with another, and resolve. Power does not fear ideas. It fears sustained attention, and the architecture is, among other things, a machine for preventing it. Fatigue is governance.
It is worth, once, leaving the architecture and standing where it lands, because a piece about systems that never shows a person has missed the cost. Picture a sixteen-year-old who has never known a world without the feed, whose attention has been shaped since childhood by a system optimised to hold it. She is not misinformed in the old sense; she could pass a fact-check. What has been taken from her is subtler and harder to name: the capacity to stay with a difficult thing long enough for it to become hers, to read past the first irritation, to tolerate the discomfort of an idea that does not resolve in fifteen seconds. The architecture did not lie to her. It trained her, gently and continuously and from a very young age, out of the one faculty that self-government actually requires, which is sustained attention. A citizen who cannot hold an idea long enough to weigh it is not a citizen who has been deceived. She is a citizen who has been disarmed, and she will defend the device that did it, because it is also where her friends are, and the architecture has made sure that those are the same place.
That is the consequence that matters more than any single suppressed story. A public can recover from a lie; it has done so many times. It is far less clear that a public can recover from the erosion of its own ability to attend, because the faculty you would need in order to notice the loss is the faculty being lost. The managed environment does not have to win every argument. It only has to keep the population too scattered, too fatigued, and too perpetually refreshed to sustain the kind of collective attention that has, historically, been the only thing capable of forcing power to answer.
There is no easy road back to an unmanaged visibility, and not because it could not be built. It is because building it would weaken every aligned interest at once: the platform that monetises attention, the advertiser that needs a safe environment, the state that finds normalisation more convenient than legislation, the security apparatus that treats perception as terrain. Visibility without management threatens all of them, so management persists, quietly, procedurally, and by default permanently.
The strongest case for the defence
The objection to all of this is serious and deserves its full weight, because in its best form it is not a defence of manipulation but a description of how these systems actually work, and of the genuine problems they face. Engagement-based ranking, the argument runs, is not manipulation but responsiveness: the algorithm amplifies outrage because users reliably click on outrage, which is a fact about human nature, not a corporate plot, and a system that gave people what they said they wanted rather than what they actually clicked would simply lose to one that did not. Content moderation is not censorship but a necessary response to real harms, incitement, fraud, coordinated abuse, the flood of genuine disinformation, and the same critics who decry suppression would be the first to demand removal of the content that targets them. The advertiser boycotts cut both ways: they are the market disciplining platforms toward safety, which is what critics of "hate speech" asked for. And the platforms are not a state; they are private companies with no obligation to carry any particular speech, and treating their editorial choices as a sinister architecture of control mistakes ordinary commercial and legal pressure for a coordinated assault on the public mind.
This objection is strong, and several of its parts are simply correct. Engagement ranking is partly responsiveness; real harms do exist and do require moderation; the platforms are private actors under real legal and commercial pressure. If the claim of this piece were that a cabal sat in a room deciding what the public may think, the objection would demolish it. But that is not the claim, and the objection, for all its force, does not reach the actual argument. Grant that the ranking is responsive to clicks: the point is that a private interest then chose to optimise for those clicks knowing, by its own internal research, what optimising for them did to the public sphere, and kept doing it. Grant that moderation is necessary: the point is not that lines are drawn but that the criterion for drawing them shifts with policy and with advertiser comfort rather than with truth, as the pandemic reversals showed. Grant that the platforms are private: that is precisely the problem, because it means the most consequential infrastructure of public perception is governed by commercial incentive and shielded from democratic accountability at the same time. The defence establishes that no one is twirling a moustache. It does not establish that the environment is neutral, and neutrality, not malice, is the thing being denied.
There is a second objection, gentler and in some ways harder, and it should be met on its own terms. People are not passive dupes, it runs. They mock the feed, they organise on it, they use the same platforms to expose the very powers this piece worries about; the architecture that supposedly manages them is also the one that let a Facebook whistleblower reach the world and let dissidents topple governments. To describe users as conditioned, attention-disarmed, and pre-selected is itself a kind of contempt, a theory that flatters the analyst and patronises everyone else. Human beings are stubborn, resourceful, and quite capable of seeing through manipulation, and they route around control as fast as control is built.
This is true, and it is the most important corrective in the whole debate, so it must be granted without flinching: people do resist, the same tools do serve exposure and organisation, and any account that reduces the public to programmed reflex is both wrong and ugly. But notice what the objection quietly concedes. The resistance it celebrates is real precisely because it is effortful, and effort is the scarce resource the architecture is built to drain. People can see through the feed, and seeing through it costs them attention, energy, and time that the environment is simultaneously working to exhaust. The dissident who uses the platform against power is swimming against the current the platform generates, and the fact that strong swimmers exist does not mean the current is not there or does not, on average, carry the rest downstream. The claim was never that the architecture is omnipotent or that individuals are helpless. It is that it sets the slope, and a slope does not have to stop the determined few to shape where the many end up. The exceptions are real. They are also, by construction, exceptions, and a system that can be beaten only by exceptional effort has already won against everyone not making it.
What would falsify the reading offered here? If platform visibility tracked verifiable accuracy rather than engagement and policy, if suppressed and amplified content sorted by whether it was true rather than by whether it was compliant or safe for advertisers, then the claim that the architecture manages perception rather than informs it would collapse, and the responsive-neutral-pipe account would win. The mechanism predicts the opposite: that across elections, war, and emergency, reach will track alignment, not accuracy, and that the criterion will move when the policy moves. Watch the reversals. The thesis lives or dies on whether visibility follows truth or follows interest.
The imprint
You open the same apps. The same feeds appear. Nothing looks different, except that now, perhaps, you know this: what appeared was pre-selected, and what did not appear was filtered out, not by fate and not by technology, but by people acting within a system that rewards control and punishes openness. Propaganda no longer needs a message. It is the environment. And as long as that environment is privately owned, politically useful, commercially disciplined, and structurally shielded from accountability, power will not bother to persuade you. It will pre-select the reality you are persuading yourself about.
The way out of that, if there is one, does not begin with a better feed. It begins where the architecture is weakest, which is exactly the place it works hardest to keep dark: the simple recognition that the order of what you see is not the natural order of the world, that someone chose it, that the choosing can be named, and that an environment which has been arranged can, in principle, be rearranged. The architecture's deepest protection is that it looks like weather. The first act of resistance is to remember that it is, in every ranking and every suppression, a decision, and that decisions have authors, even when the authors have arranged never to be found.
This is, in the end, why the analysis insists so stubbornly on the word decision against the word drift. Drift asks nothing of anyone; it is no one's fault and therefore no one's job to fix. Decision restores the only thing the architecture has worked to erase, which is the existence of a chooser. The Facebook memos had authors. The threshold that lets one narrative surge and another sink was set by a person in a meeting. The clause that shields the whole arrangement from liability was written by a legislature and can be rewritten by one. None of this is fate, and the moment it stops being experienced as fate, it becomes, for the first time, contestable. That is not a solution. It is the precondition for there ever being one, and it is the single thing this piece is for: to convert what feels like the weather back into what it actually is, a structure, built by people, serving interests, and therefore, like every structure people have ever built, answerable, if enough sustained attention is ever brought to bear on it at once. The architecture is betting that it will not be. That bet is the whole game.
Evidence Map
Facts, interpretations, forecasts, and disconfirming signals.
Core claim. The determining variable of modern information power is no longer the message but the environment of visibility, what appears when a person thinks, governed by ranking choices that are policy translated into code. Three forces, the platform's attention economy, the advertiser's demand for brand safety, and the state's preference for normalisation, converge to produce a managed environment without any of them needing to coordinate. The claim concerns aligned incentives and documented decisions, not a central plan.
Evidence level. Facts (high): Facebook's 2018 "meaningful social interactions" engagement-ranking change and the internal research, disclosed by Frances Haugen in 2021, showing it amplified divisive content; the 2020 Stop Hate for Profit advertiser boycott (500+ companies) and the 2023 advertiser exodus from X, both moving platform behaviour; the EU Digital Services Act (in force 2022) imposing systemic-risk duties on very large platforms with fines up to 6% of global turnover; documented pandemic-era content reversals (e.g. the lab-origin hypothesis). Interpretation (medium, marked): the synthesis of these into a single "architecture that manages perception" and the reading of fatigue as a mode of governance are analytical conclusions, not documented intentions; "no master plan, only aligned incentives" is explicitly the position taken.
What would confirm this. Platform reach continuing to track engagement and policy alignment rather than accuracy; criteria for amplification and suppression shifting with official posture and advertiser comfort.
What would disprove this. Evidence that visibility tracks verifiable accuracy rather than engagement and compliance, that suppressed and amplified content sorts by whether it is true rather than by whether it is safe or aligned. A truth-tracking pattern would defeat the managed-environment reading.
Watchlist. How the DSA's systemic-risk regime is used in practice; whether moderation criteria move with policy in the next emergency; advertiser pressure as a driver of what stays sayable.
Jerry van der Laan writes The Manifest Archive, where he traces the structures beneath the headlines. He traces the structures beneath them.
