Open your phone to check one thing, and an hour later you surface from a stream of outrage clips and arguments you never went looking for, a little more wound up than when you started. Ask yourself honestly whether that was what you wanted, and then ask why it keeps happening anyway. Nobody censored the calm, useful version of the day; it is still right there if you go dig for it. It just lost, quietly, at every turn, to whatever kept you scrolling. What follows is the machinery behind that, and about its most uncomfortable feature: there is no villain to blame here, no censor to overthrow. Just an arrangement that pays better when the noise wins.
Everything so far has been machinery: how ideas get compressed, which gates they pass, why the medium decides what can win. Machinery has no interests. But the moment you say a gate works by criteria, and criteria have to be set, you have to ask who sets them and who benefits. Chapter 5b called selection the mechanism with a steering wheel and pointed here for whose hands are on it. Chapter 8 found the modern collapse asymmetric, preservation mostly holding while training hollowed, and pointed here for which interest that asymmetry serves. That is the question here.
Selection has an owner because it is criterial
Transport scrambles with no preferred outcome. You can make it cheaper (a share button, an LLM that decompresses on demand) or dearer (a character limit, a paywall), but a cost has no will: once it is in place it does not favor one variant over another. Choosing which cost to impose is a design move with an aim, and I will come back to that. At runtime, though, there is nothing to own.
Selection is different in kind. A gate ranks candidates against criteria, and a criterion is the sort of thing that has to be set. Even when no one sets it consciously, even when it emerges from the shape of an interface or the drift of an algorithm trained on clicks, something is doing the setting, and that something has interests that are not yours. A mechanism that operates by criteria has, by construction, an entity that sets the criteria; calling that entity the medium’s owner is a definition, not an accusation.
The frozen-selection point sharpens it. The medium is selection performed once, in advance, and baked in until it stops looking like a choice, so the owner who matters is not someone tuning weights at runtime but the one who wrote the option space before any gate fired. *Amusing Ourselves to Death* (p.84) puts it bluntly: “every technology has an inherent bias. It has within its physical form a predisposition toward being used in certain ways and not others.” That predisposition has a designer, and the designer’s interest is part of what got designed in. Pope Leo XIV’s *Magnifica Humanitas* (2026), which Chapter 11 takes up in full, names the result at the political layer: tech firms “define conditions for access, rules of visibility” (para. 71), a private authority over the selection surfaces that has displaced the public kind rather than merely joined it. So who are these owners, what do they optimize for, and why those optima? For the medium this book is actually about, the engagement-optimized internet, the next sections answer.
What engagement actually measures
The dominant medium’s gates are tuned for engagement, a word repetition has drained of meaning. It does not measure how much a piece of content helps you, or teaches you, or anything you would reflectively want. It measures how much time you spend inside the medium as a result of seeing it: clicks, dwell, scroll depth, comments, return visits. That time is what gets sold to advertisers. So engagement is a proxy for revenue, dressed up as a proxy for user benefit, and the fit between the metric and any honest reading of “what people want” is an accident, not a design goal.
The evidence that the fit is bad is sharper than I expected. *Psychology of Virality* calls it the paradox of virality (p.9): people know that “negative and high-arousal content goes ‘viral’,” and “the vast majority of people across the political spectrum reported that they do not want this to be the case,” preferring “more positive, nuanced, and accurate content.” When asked, people want different content than the gates deliver, and the gates deliver it anyway.
That looks like it should break any selection story: if selection tracked what people wanted, the unwanted thing would lose. It doesn’t lose, because the gate doing the choosing was never yours. It is the medium’s, tuned by someone whose interest is not yours. O’Connor and Weatherall give the live mechanism: fake news “is amplified by algorithms that social media sites use to identify which stories are highly engaging so as to maximize their viewership” (*The Misinformation Age*, p.173). Not a side effect. The point of the system. So the political-economic core, plainly: the dominant medium’s gates maximize a metric whose alignment with truth, or with anything you would on reflection prefer, is structurally accidental.
The Orwell mode and the Huxley mode
Postman, in the foreword to Amusing Ourselves to Death (setting Orwell’s 1984 against Huxley’s Brave New World), draws the distinction needed here. There are two opposite ways to control a society’s relation to information. The Orwell mode is control by inflicted pain: censorship, surveillance, dissidents suppressed. The Huxley mode is control by inflicted pleasure: truth not hidden but drowned in a sea of engineered irrelevance, dissidents made irrelevant rather than silenced. The West braced for Orwell all century and walked into Huxley while no one was looking.
The modern failure mode is Huxley, not Orwell. The medium is not censoring the well-sourced account of anything that matters; that account is still out there, still returned by search, still in the journals. What the medium does is select against it. At every gate it loses to the engagement-tuned alternative, and your attention is spent before you ever reach it.
That is harder to fight than censorship, for one exact reason. Censorship has an enemy: a law, a censor, someone to point at and organize against. Engagement-optimized drowning has none, because nothing is being suppressed. You can organize against a censor. You cannot organize against having gotten distracted. And because nothing is suppressed, no one can be charged with censorship; the system wears the perfect cover of formal neutrality. Every variant is technically available; the engagement-tuned one just keeps winning. Some states now run both modes at once, engagement feeds for the Huxley work and coercion for the Orwell work, each reinforcing the other. The two-mode frame is the right starting point; the composite case is a real complication it has not fully absorbed.
Cost-shifting as the structural move
The Democratization Paradox, an earlier essay of mine, named the move that makes all of this quiet: cost-shifting. When the barriers to creating content fall, the cost of quality assessment does not disappear. It transfers, from producers (who used to bear it as the price of being published at all) onto the reader, who now has to do the filtering with no more time than before. In the old world a publisher paid to vet a book before it reached you, and that vetting was baked into the price of publishing at all. Now anyone can post anything for free, and the job of working out whether it is true has moved onto you, to be done in the same twenty-four hours you always had. The work didn’t vanish. It got handed to the one person in the chain with no budget for it.
This is what lets the political economy stay invisible while doing enormous work. The platform never has to defund a journal or censor a careful account. It just makes production free, which drops the assessment burden onto a receiver budget that was already full. With no time to spare and no line item for “quality,” the reader defaults to the cheapest heuristic: brand, in-group cue, familiar source. *The Misinformation Age*‘s account of credibility-weighted updating is this, receivers weighting evidence by source trust because they cannot check everything first-hand. The platform’s tuned gate and the cost-shifted reader converge on one criterion neither designed: trust by familiarity. There is no villain to point at, only an externality that landed on someone with no balance sheet to absorb it. And that is why the book’s prescription has to do more than name better media: any curation layer asks someone to pay the assessment cost, so unless that cost is funded directly, the free-and-engaging default out-competes it every time. Chapter 9’s curation layer needs this resourcing argument to stand up.
Two kinds of capture
The dilution loop left three outcomes for an institution repairing a diluted myth, and Chapter 8 named the stable bad one: the captured cell, an effective institution distributing a tuned key. Applied to the platform, capture turns out to come in two kinds that are not the same object. (The book’s unified treatment, which surface is captured and how hard recovery is, is the capture taxonomy; this is its by-whom axis.)
External capture is the kind described implicitly so far. An adversary (a propaganda operation, a state actor, an ad-tech firm) learns the criteria the gates run on and crafts variants that win at them. *The Misinformation Age* (p.175) calls it an “asymmetric arms race”: “whatever barriers we erect against the forces of propaganda will immediately become targets for these sources to overcome.” Hostile, but legible: there is someone to fight, regulate, expose.
Self-capture is the form the argument has been moving toward without naming. Engagement optimization is itself a tuning of the gates against truth, in the precise sense of manipulation: tuning a gate’s criteria against what the gate was meant to serve. Turning attention into ad revenue is the captured selection-tuner. No outside actor does it; the institution’s own operating logic does. The medium is captured by its own monetization.
And that is the harder problem, because it has no enemy. Self-capture is more stable than external capture, and that is what makes it the central problem. An outside attacker can be named, sued, regulated, exposed; a business model cannot, because the captured state and the normal state are the same state. You can fight a captor; you cannot fight an equilibrium. This is why reforming platforms from within keeps underperforming: it aims at an adversary that isn’t there. The only moves with teeth are to displace the equilibrium with a selection arrangement that operates outside it, or to dismantle it by changing the resource flows that hold it in place.
The two kinds are not separate in practice; self-capture builds the gradient external actors ride down. A propaganda operation need not fight engagement-tuned gates, only align with them, because they were already tuned toward the high-arousal, identity-flagged content it wants to inject. *Psychology of Virality* (p.8) names the substrate: “Our attraction to high-arousal negative information likely has deep evolutionary origins… Social media algorithms appear to take advantage of our attraction to threat.” The algorithms tuned themselves to a pre-existing vulnerability; the attackers then rode the tuned algorithms.
One more cut, along Chapter 2’s three-realities seam, because everything above concerned objective content, truth lost in transit while the facts wait outside. The intersubjective-truth note works out why the same equilibrium does something categorically worse to intersubjective content (currencies, nations, legitimacy). For that cargo the platform is not distorting reports about a reality; it is hosting part of the reality’s generator, because those truths are made by the network’s agreement and the platform is where the network now agrees. A captured platform therefore does not just shape what its users believe. It shapes which we exists: which identities can compose, which legitimacies hold. And here the capture ships with camouflage the objective case lacks. Reification, agreement wearing the costume of nature (“the algorithm decided,” “the market spoke”), files the generator’s outputs under physics, and a generator nobody can see is one nobody audits.
The same equilibrium, run on money
Be suspicious of a structural claim that only ever produces one instance. Self-capture was argued as a property of the arrangement, not of the ad business in particular, so it should appear wherever its conditions hold. It does, and the cleanest second case is public blockchains, where the machinery runs again with money in place of attention. A disclosure first: I work on this problem professionally, on private-by-default versions of these rails, so I have a stake in where this section lands. Read it with that discount. The argument needs no particular project to succeed and will not argue for one.
Blockchains made an intersubjective truth’s constitution machine-readable and self-enforcing, with the fork kept affordable and the rules holding because anyone can check they held. The design-time move that bought this was to publish the network’s entire objective shadow, the ordinary facts about an agreement’s participants: not “this token has value” (the truth) but who paid whom, how much, when, from where (the shadow). A bank publishes none of its shadow; a public blockchain publishes all of it, to everyone, forever. That was frozen selection in the exact sense, a criterial choice made once and baked into the substrate. The designers were buying verifiability, and in 2009 surveillance was nobody’s business model. It came bundled anyway, because the option space had already decided.
Then the equilibrium condensed, the way ad-tech condensed around attention, and an industry formed at every gate the medium concentrated. The query services most wallets route through can see which address asks what, from where. Pending transactions wait in a public staging area that firms archive, keeping a permanent record of intentions including the ones that never executed. The assembly of transactions into blocks has consolidated to where a couple of firms build most of them. Downstream, an analytics industry clusters addresses into identity graphs and sells them to exchanges and states. (The Anatomy of Exposure walks the full lifecycle leak by leak, with the receipts.) None of it is an attack; every piece is the default doing what defaults do, making one business model cheaper than the rest.
Both capture flavors replay. The self-capture is the ordering gate: whoever assembles a block chooses the sequence, visible pending trades are worth money to whoever can trade around them, and so the gate is tuned to extraction, at a scale where a single thirty-day window logged seventy-two thousand instances of the signature move (buy just before a victim’s pending trade, sell just after, so their order pays the profit). No adversary did this; it is the equilibrium of visible order flow plus a gate with an owner. External capture then rides the self-captured gradient: when the United States sanctioned a privacy tool in 2022, most new blocks at the peak quietly declined to include the sanctioned transactions, not because any rule changed but because the concentrated gate operators were identifiable and could be leaned on one at a time. Self-capture built the chokepoints; the state rode them. A medium engineered to have no censor grew one out of its own defaults.
The case adds a blade the attention economy never shows. This book runs on loss; its title names a pathology. Here is a medium with the opposite pathology: it never loses anything. The shadow is archived at perfect fidelity, forever, so exposure is retroactive, the analysis improving every year against a record that cannot decay. A transaction private enough today is not private enough in ten years, and nothing its sender does can change that. Set that against cash, whose shadow evaporated at the counter. Forgetting was the load-bearing lossiness of every pre-digital record, removed here without a decision. Run privacy through the book’s machinery and it comes out as engineered loss, the deliberate re-introduction of decay into the shadow. The book treats loss as the disease; this is the case that keeps the vocabulary honest, because loss is also the only thing that has ever stood between a person’s shadow and whoever would monetize it.
The camouflage is here too. “The chain is transparent, that’s just how blockchains work” is reification, a design decision filed under physics. It is not physics: proof systems in production for years now verify that every rule was followed without revealing who did what, so verifiability and visibility come apart. Public-by-default was one point in an option space, chosen when verification was the only job on the table, and the costume slips the moment a working alternative exists. So the equilibrium is not a fact about attention. It is a fact about media whose frozen defaults make observation the cheapest input. Wherever that holds, an industry forms to monetize the watching, and it is the medium’s native economics, not an abuse of it. Which is why the same move fails in both: patching leaks leaves the incentive untouched and the leak just moves, while changing the default takes away the subsidized input the whole industry condensed around. The fix lands at the default layer, on transparent money as on attention.
Why institutional carriers can’t keep up
Chapter 8 diagnosed the modern collapse as asymmetric and pointed here for the mechanism. Here it is. The platforms never had to suppress institutional carriers. They had an easier move: out-compete them in the attention market without ever engaging them. The platform’s marginal cost of attention is roughly zero, because users make the content, the algorithm selects, and ads pay. The depth journal, the lifetime academic, the investigative reporter, the careful textbook all have real costs per unit of attention. The platform can sell attention at zero; the carrier cannot.
A reader with a fixed budget and no line item for quality takes the free option, which is also the one engineered to be most engaging. That is not a moral failing; it is the market they were handed, priced at zero for engagement and at hours for depth. The carrier is not outargued or discredited. It is priced out. So the deep form keeps being produced and no attention is left to reach it, which is exactly Chapter 8’s asymmetry: preservation holds, training collapses. And it runs wider than training. As Chapter 8 already notes, elite preservation survives on endowment and prestige while broad preservation (local journalism, regional universities, adult education) collapses in the same squeeze. The mechanism is one; the reach is the whole of it.
The argument’s dents
I have pushed this argument against the strongest published objections to it, and it survives with two honest dents worth naming. First, most of what travels through any platform is neutral drift, not selection: Information Evolution in Social Networks (p.10) shows meme variants mostly following a Yule process of random copying plus rich-get-richer. So the political economy of selection shapes the meaningful slice of what spreads (the high-arousal, identity-flagged content the gates favor), not the neutral tail. Structural, yes; total, no. Second, the argument rests entirely on the medium being frozen selection. Postman (p.102) insists there is “no conspiracy here,” only a medium whose form dictates what counts as good, and he is right at the runtime layer he describes. The claim here lives one layer upstream: the form itself was set, once, at design time, by whoever built the medium. Without that upstream move Postman wins and the steering wheel is held by no one; with it the argument holds, and Chapter 11 will lean on it harder still. The other objections refine without denting: the medium amplifies pre-existing dispositions rather than causing them, and Mercier’s evidence that mass-persuasion campaigns routinely fail (*Not Born Yesterday*, p.259) is a within-campaign result, while the worry here is decades of demand-shaping, the timescale Chapter 7 carries. Campaign-scale persuasion failing is compatible with generation-scale want-shaping succeeding.
Where I land
Engagement maximization is not a bug, a moral failure, or an outside attack. It is the equilibrium of a particular arrangement: a medium owned by entities that monetize attention, a metric that proxies revenue rather than benefit, and a frozen option space set once and baked in. The business model is itself the captured selection-tuner, and that captured state is more stable than any adversary in this book, because it has no adversary, only an arrangement. The breakage circled throughout is not compression and not selection. It is who tunes the selection criteria, and at the dominant medium of this era the answer is the owners of an attention market priced to make engagement-optimized noise free and well-sourced depth unaffordable. Chapter 11 inherits this in a harder form, where one owner sets the gate, the option space, and the content-generator at once; Chapter 12 inherits the constraint, that any fix has to work in the attention market we have rather than wait for it to change.
Where I’m still uncertain
- Self-capture may prove too much. I claim self-capture is uniquely stable because the captured state is the institution operating normally. But that would make any institution whose normal operation is bad uniquely stable, which is unfalsifiable as stated. The distinction I want is that the captured state is profitable in a way mere incompetence is not, and I have not worked it out cleanly.
- The displacement bet is asserted, not argued. I treat displacement (selection arrangements outside the captured market) as the structurally available move and reform-from-within as the slow auxiliary. Both routes have mixed historical records, and the bet needs an argument not given here.
- The money case may be analogy, not instance. Attention monetizes the gate; transparent rails monetize the shadow. “Media whose defaults make observation cheapest” unifies them at one altitude, but the captured surfaces are different objects, and I have not checked whether the capture taxonomy files them in the same cell or merely adjacent ones. If adjacent, the equilibrium claim generalizes; if only rhyming, this section should be demoted from instance to illustration.
← Chapter 9: The Integration Problem · Chapter 11: AI as a New Kind of Node →