Plain-language definitions of the book’s load-bearing terms. Entries are grouped by where each enters the argument, so top-to-bottom roughly follows the book’s own path; the table of contents on the right is the alphabetical-ish index, so jump straight to any term. Each entry ends with where to read more, and, where they differ, the graph node the term maps to.
The pipeline
Transport
The work of re-encoding an idea for the next stage of its journey: compressing a finding into a news story, a story into a meme. Transport is lossy, and what it strips is usually the decoding key (methodology, qualifications, scope), not the bare claim. It is a cost with no preferred outcome, which is what distinguishes it from selection. Compression is transport seen from the receiver’s end. See Chapter 5; graph node transport.
Selection
The mechanism operating alongside transport: at every stage a gate decides which ideas pass at all, on local criteria (publishable, newsworthy, shareable, clickable). Unlike transport, selection runs on criteria, so it is the half of the machine that can be aimed, captured, or repaired. A selection gate is one such choice point; its criteria are institutional choices, not laws of nature. See Chapter 5b; graph node selection.
The pipeline
The stages an idea moves through from reality to a general audience: The Out There → Raw Data → Insight → Theory → News → Meme. The book treats this, plus the sibling selection gate at each stage, as the unit of analysis. See Chapter 1.
Memetic fitness
The grab-bag of things a selection gate actually rewards: emotional charge, identity reinforcement, status, “tells you what to do,” novelty, controversy. None is coordinated, and none cares whether the thing is true. See Chapter 5b; graph node memetic-fitness.
Option space
The menu of forms an idea could take at a given stage. Transport generates the variants, the medium sets which are expressible at all, and selection only picks among them. That is why the medium, which writes the menu, is the higher-leverage and less visible lever (transport generates, the medium bounds, selection picks). See Chapter 5b; graph node option-space.
Manufactured content
Content injected into the pipeline at a later stage without ever originating in measurement: propaganda, AI-generated text, agenda-driven framing. The gates can’t reliably tell it from measured content, which is what makes the distinction load-bearing. See Chapter 1; graph node manufactured-injection.
The medium
The medium (in this book’s sense)
Not the physical carrier (paper, airwaves, fiber) but the gate-criteria that carrier enforces. Different media reward different content, so changing the medium changes what can win and what receivers come to want. See the medium note.
Technology vs. medium
The technology is the machine or tool; the medium is the whole environment of habits, rewards, and expectations that grows up around how it gets used. This book’s “medium” always means the second, which is why you can’t blame the hardware for what is really a pattern of use. Postman’s distinction. See Postman; graph node technology-vs-medium.
Medium shapes want
Different media cultivate different appetites over time: one that rewards shallow fast engagement trains people to want shallow fast content; one that rewards sustained attention does the opposite. McLuhan’s “the medium is the message,” restated at the level of what audiences come to want. See the medium note; graph node medium-shapes-want.
The receiver side
Preconditions
The mental equipment a receiver must already hold to decode an idea correctly: vocabulary, frameworks, contextual knowledge. An idea is “complex,” in the book’s sense, when its preconditions are large or rare. Also called the outer message (below). See Chapter 5.
The three layers of a message
Hofstadter’s model, adapted: every message has a frame layer (this-is-a-message), an outer layer (how to decode), and an inner layer (the content). The load-bearing move: the outer message is what the book otherwise calls preconditions, and it comes mostly from the medium. See Chapter 5; graph node three-layer-message.
Outer message / decoding key
The decoding instructions that come, or fail to come, with a message: what tells you how to read the inner message. “Decoding key” is the book’s most frequent name for it. Most pipeline distortion is loss of the outer message, not the inner one: the claim survives transit; the key it needed doesn’t. See Chapter 5; graph node three-layer-message.
Receiver budget
The hours a reader has for taking in new material, and what those hours can carry: a tablespoon of weeks, a few focused hours in each. The hours are fixed; what each hour carries grows with structure the reader has already built. Finite, but trainable. See Chapter 3; graph node receiver-budget.
Want
The receiver’s standing appetite for engaging with material. The book argues want is the prime mover behind both transport and selection: receivers engage with what they want, and the gates reward what enough receivers want. See Chapter 3; graph node want-as-prime-mover.
Complexity ceiling
The most complexity an idea can carry and still cross a given network. The bigger the network, and the more compression and selection pressure it applies, the lower the ceiling. See Chapter 5; graph node complexity-ceiling.
Complexity / virality trade-off
The book’s name for Chapter 5’s core claim: how easily an idea spreads is inversely related to how complex it is, because every hop costs receiver budget and the losses compound across a network. See Chapter 5.
Handle-ability
What actually decides whether an idea spreads: not how many preconditions it needs, but how easily it plugs into models people already have, how much feeling or identity it carries, and whether it tells them what to do. A property of the selection side. See Chapter 5b; graph node handle-ability.
Compressed form / complex form
Two versions of the same idea. The compressed form is cheap to process and easy to pass on (and to weaponize); the complex form is expensive but keeps what the compressed form drops. See Chapter 5; graph nodes compressed-form, complex-form.
Compression and truth
Interpretive latitude
The range of valid readings a compressed claim supports across a population. Compression doesn’t just shift the expected decode; it widens the variance: different receivers reach for different keys and arrive at different inner messages. See Chapter 5c; graph node interpretive-latitude.
The three regimes
What happens to truth under compression. The same compressed claim can land as preserved truth (the receiver’s key is close to the original), inverted truth (keys oppose: a true claim decoded into a false belief), or orthogonal truth (the claim read as an identity flag, not decoded at all). Which regime occurs is set by which variant the selection gate picks. See Chapter 5c; question node truth-value-placement.
The corrective
The signal that a widely-held claim is wrong: a replication failure, a consensus reversal, a retraction. Chapter 2 sorts its cases by whether the corrective survived the pipeline; the book’s diagnosis is that a corrective almost always loses to the myth it corrects, because it carries less emotional payload. See Chapter 2.
Three realities (Harari)
Three kinds of thing a claim can be about: objective (true regardless of belief: atoms, planets), subjective (true only inside one mind: a pain, a dream), and intersubjective (true because enough people agree: money, nations, laws). The pipeline transmits-or-distorts objective claims, mostly skips subjective ones, and partly creates intersubjective ones. See Chapter 2; graph node three-realities.
Intersubjective truth
A truth generated by the network’s agreement rather than by reality. “The twenty in your pocket buys groceries” is really true, but erase the agreement from every head and nothing is left to re-measure; the truth was the agreement. Same word (truth), different generator. See the intersubjective note; graph node truth-generators.
Constitutive regimes
What the three truth-regimes become when the cargo is intersubjective: convergent constitution (decodes compose into one agreement), divergent constitution (camps affirm the same words as different agreements: a constitutional crisis), and hollow constitution (words stay affirmed while the constituting behavior evaporates: everyone still calls it money, nobody holds it). See the intersubjective note; graph node constitutive-regimes.
Constitutive transmission
For intersubjective content, spreading a claim is part of making it true: a receiver’s decode becomes one more share of the agreement. Transmission is participation in constitution, not reporting on it. See the intersubjective note; graph node constitutive-transmission.
Reification
Treating a made thing as a found thing: an intersubjective truth processed as if objective (“money just is valuable,” “the law is the law”). Not innocent: an agreement that reads as a law of nature has a generator nobody audits, which is the camouflage a captured generator wants. See the intersubjective note; graph node reification.
The fork
What stands in for the corrective when there is no referee: an intersubjective truth cannot be falsified, only renegotiated, abandoned, or forked (schism, secession, dollarization, chain split). A fork doesn’t determine who was right; it determines how many truths there are now. Design lesson: keep the fork affordable. See the intersubjective note.
Constitutive volatility
The transport collapse’s second blade. Agreement states used to carry inertia (a bank run moved at the speed of queues on the pavement) and now move at meme speed (a run coordinated in group chats), while the institutions that steady agreements still move at committee speed. See the intersubjective note.
Selection up close
Algorithm as selection engine
A recommendation algorithm isn’t lossy transport; it’s a selection gate with adjustable weights. That makes “who tunes the gate?” a literal, operational question, and turns a few companies’ criteria into everyone’s. See Chapter 5b; graph node algorithm-as-selection-engine.
Frozen selection
The medium seen as selection performed once, in advance, and baked into the substrate until it stops looking like a choice. A 280-character limit was somebody’s decision; to everyone living inside it, it is simply the shape of the world. That is what makes the medium the highest-leverage and least visible selection surface: a variant the option space cannot hold never has to be suppressed, because it never forms. See Chapter 5b.
Manipulation surface
What an attacker can exploit. In the pure-transport view it grows with network size; in the fuller view it is the tunable criteria of the selection gates, the more dangerous version, because the weights can be set on purpose. See Chapter 5b; graph node manipulation-surface.
Neutral drift baseline
The null model for spread, what propagation looks like with no selection: random copying plus rich-get-richer. Most variants follow it; selection is what you detect as a departure from it. See Chapter 5b; graph node neutral-drift-baseline.
False consensus among rational agents
O’Connor and Weatherall’s result: people who each weigh evidence correctly can still settle together on a false belief, because the failure lives in the network’s structure, not in individual error. See Chapter 5b; source The Misinformation Age; graph node false-consensus-rational-agents.
Credibility weighting
The rational, unavoidable habit of trusting evidence more when you trust its source, since no one can check everything first-hand. Also the seam propaganda exploits: manufacture or buy credibility and the weighting carries false evidence as readily as true. See Chapter 5b; graph node credibility-weighting.
Conformity
Adopting a belief because your peers hold it rather than because of the evidence. It can lock a community onto a belief, true or false, and seal it from correction. See Chapter 5b; graph node conformity.
How people hold beliefs
Epistemic vigilance
The cognitive defenses people use to screen incoming claims for plausibility and source reliability. Mercier sharpens it into open vigilance. See Chapter 7; graph node epistemic-vigilance.
Open vigilance (Mercier)
Mercier’s correction to the “people are gullible” story: we are both open to useful information and on guard against unreliable sources, evolved hand-in-hand. If anything we err cautious, more often failing to trust something we should than swallowing something we shouldn’t. See Chapter 7; source Not Born Yesterday; graph node open-vigilance.
Reflective vs. intuitive belief (Sperber)
Two ways to hold a belief. Intuitive beliefs are wired into your others and drive what you do; reflective beliefs are held at arm’s length, stated and shared as identity signals but inert in action. Much of what spreads online is held reflectively, which is why people share things they’d never act on. See Chapter 7; graph node reflective-vs-intuitive-belief.
Justification market
Mercier’s reversal of the obvious arrow. The naive story: content shapes belief, belief drives behavior. The reversal: behavior is mostly decided upstream (identity, group, interest), and people shop for content that justifies what they were going to do anyway. Changes what any fix has to target. See Chapter 7; graph node justification-market.
Paradox of virality
What spreads is not what people say they want spread: across the spectrum, people report not wanting high-arousal negative content to go viral, and it goes viral anyway. The gap points at the selection gate, not receiver preference. See Chapter 7; graph node paradox-of-virality.
The floor
Once a network is large and its gates reward engagement, what reliably travels settles at raw strong emotion (threat, tribe, anger, awe): the lower bound the system keeps returning to. A property of the aggregation, not of any receiver. See Chapter 7.
Networks at scale
Myth dilution
The loop by which a network’s binding myth loses determinacy as the network grows: scale forces the myth to compress, compression widens its interpretive latitude, latitude sorts the network into segments that affirm the same myth while decoding it differently. See the myths note; graph node myth-dilution.
Bureaucracy (narrow sense)
The book’s shorthand for two institutional jobs: preservation (keeping the full, uncompressed form somewhere) and training (teaching people to read it). Borrowed from Harari’s Nexus; real bureaucracies do much more, and none of that is meant here. See the myths note and Chapter 8.
Network segmentation
The fracturing of a network into sub-groups along the seams of its myth’s interpretive latitude, so it is superficially unified and structurally fragmented at once. See the myths note; graph node network-segmentation.
Echo chambers vs. bubbles (Nguyen)
Bubbles are missing-information structures; echo chambers actively discredit outside information. Small networks tend toward chambers, which changes what integration has to do. See Chapter 9; source Nguyen; graph node echo-chambers-vs-bubbles.
Polarization via distrust
O’Connor and Weatherall’s result: when agents discount evidence from sources they distrust, communities with any initial difference rationally diverge, and the divergence deepens the distrust that drove it. No irrationality or misinformation required, and adding a channel between distrustful communities can accelerate divergence. See Chapter 9; source The Misinformation Age; graph node polarization-via-distrust.
Zollman effect
Increasing connectivity inside a community can reduce the accuracy of its beliefs: densely connected agents see each other’s preliminary results too fast and converge prematurely on whichever answer took an early lead. There is an optimum density. See Chapter 9; graph node zollman-effect.
Inverted-U relationship
A shape that recurs in the argument: not “more is better” or “less is better” but an optimum density, size, or barrier-height with diminishing returns on both sides. See The Democratization Paradox; graph node inverted-u.
The bridge zone
Bridge zone
The stretch between deep specialists and mass audiences, staffed by journalists, popularizers, influencers, and AI summarizers. Selection pressure is most intense here, and the reshaping is active (forms built for clicks and approval), not passive blurring. Where most of the visible distortion happens. See Chapter 6; graph node bridge-zone.
Pseudo-context (Postman)
Real, accurate information stripped of the context that connected it to any action, so it survives only as trivia, quiz filler, or conversation-filler. Different from manufactured content (which never measured anything): pseudo-context is real information made idle. See Chapter 6; graph node pseudo-context; source Postman.
Political economy of attention
The modality argument
Why selection has an owner and transport doesn’t. Transport loss is a cost nobody chooses, so there is nothing to tune; selection runs on criteria, and criteria have setters. That asymmetry makes the political economy of attention structural: wherever there is a gate, someone owns its dial. See Chapter 10.
Engagement as a captured equilibrium
Engagement-maximization isn’t an outside attack on a platform; it is the business model doing its normal job. Time-in-app sells ads, so the gates get tuned toward whatever holds attention. That is what makes it stable: no villain to defeat, only an arrangement to dismantle. See Chapter 10; graph node engagement-equilibrium.
Out-competition of institutional carriers
Platforms don’t have to censor journals or universities; they out-compete them. Engagement-bait costs roughly nothing per view (users make it), depth costs real money per view, and audiences with limited attention take the free option, so the institutions collapse with no direct attack. See Chapter 10; graph node out-competition-of-carriers.
Cost-shifting from producers to consumers
When barriers to entry fall, the cost of quality assessment doesn’t disappear; it transfers from producers to the reader, who has no more time. The externality underneath the modern noise problem. See Chapter 10; source The Democratization Paradox; graph node cost-shifting.
Superspreader dynamics
A handful of high-connection nodes drive a wildly disproportionate share of what spreads, a network-structure effect with no offline equivalent at the same scale. About who is positioned to amplify. See Chapter 5b; graph node superspreader-dynamics; source epidemiological-virality.
The Orwell/Huxley axis (Postman)
Two opposite ways an information environment fails. Orwell: control by inflicted pain (censorship, surveillance). Huxley: control by inflicted pleasure, the truth not hidden but drowned in engaging irrelevance. The book’s diagnosis is Huxley, which is harder to fight because there is nothing to push against. See Chapter 10; graph node orwell-huxley-axis.
Engineered loss
Privacy, run through the book’s machinery: the deliberate re-introduction of decay into a record that would otherwise persist forever. The transparent-ledger case forces the observation that lossiness can be protective: strip an idea’s decoding key and you get the book’s usual pathology; strip the trail its participants leave and you get protection. See Chapter 10.
Capture, unified
Capture
Tuning a selection surface against the very thing it was supposed to serve. The taxonomy sorts every case along three axes: which surface, by whom (an outside adversary, the institution’s own business logic, or both), and how hard it is to recover. See the capture taxonomy.
Self-capture vs. external capture
Two sources. External: an outside actor learns a gate’s criteria and games them, an arms race you can at least fight. Self: the institution’s own business model is what tunes the gate against truth, no adversary, just an equilibrium, which is why it’s more stable. The common real case is composite: self-capture lays a slope external actors then ride. See the capture taxonomy; graph node self-vs-external-capture.
Consumer-key vs. surface capture
Why some captures are far worse. Capturing something that installs decoding equipment in people’s heads or models (what they’re taught, what an AI is trained on and for) damages the consumer and can’t easily be un-installed. Capturing something that only shapes the surface a person meets damages only the surface. The first kind is the dangerous kind. See the capture taxonomy; graph node consumer-key-vs-surface-capture.
The seven capture surfaces
The substrates capture can act on, unified from three chapters and ordered roughly most to least recoverable: deployment configuration and gate-criteria (the cheapest to re-tune), preservation archive and receiver training (Chapter 8), and training corpus and training objective (Chapter 11, the objective the worst of all because it self-reinforces across model generations), plus option space (frozen at design time, capture by absence). The recovery hierarchy is the operational point: defend the consumer-key and self-reinforcing surfaces first. See the capture taxonomy.
AI as a new kind of node
Selection-design surface
Any place where a design choice shapes what content can exist or travel: a gate’s criteria, the option space, the training data, the training objective, the deployment settings. Historically spread across separate institutions; the term exists to make concentration countable. See Chapter 11.
LLM as a collapsed design moment
A language model is the first node to own almost everything at once (the gate, the option space, the content generator, the training data, the objective, the deployment settings, even the price tier), in one place, one party. A platform owns two or three; a closed-weight LLM owner owns all of them. See Chapter 11; graph node llm-design-moment-collapse.
Decompression on demand
Used faithfully, an LLM can take a compressed claim or dense source and re-expand it, supplying the preconditions a reader is missing: the first real way out of the fixed receiver budget. Strictly conditional on faithfulness, since a hallucinating model supplies confident-but-wrong instructions, worse than none. See Chapter 11; graph node decompression-on-demand.
Compressed form becoming authoritative
The inverse failure: at scale the AI’s summary quietly replaces the source in everyone’s working memory, carrying the AI’s authority. The original still exists; nobody reads it. See Chapter 11; graph node compressed-form-as-authority.
Three ways to capture an LLM
Corpus capture: rigging what it learned from. Objective capture: rigging what it was trained to optimize, the worst kind, because its outputs feed the next model’s training data. Deployment capture: changing its behavior at use-time, the cheapest and most opaque. See Chapter 11; graph nodes corpus-capture, objective-capture, deployment-capture.
Substrate custody
Who holds an AI’s formative surfaces: its training data, its objective, and its deployment. An institution that needs a faithful model has to hold all three itself; renting them from a vendor means trusting the vendor’s tuning. Every hopeful LLM scenario is conditional on this. See Chapter 11 and Chapter 12.
Salvation case / worst case
The two ways an LLM can land, set by ownership. Salvation: a faithful model under an institution’s own custody, extending preserve-and-retrain. Worst case: a revenue-tuned model over a captured corpus, industrial-scale bridge-zone distortion with the surface markers of authority. Technically realizable, economically marginal, politically blocked. See Chapter 11.
The bridge and the prescription
Bridge node
A person or institution that successfully moves complex truth between networks that don’t share preconditions. Not a generalist but a versatile expert. See Chapter 9 and the bridge-node note; graph node versatile-expertise.
Versatile expertise
Deep specialist expertise paired with metacognitive flexibility: the habit of looking for structural analogies, suspending one’s paradigm, treating one’s own frame as one option rather than the substrate of reality. Depth gets the bridge admitted as a peer; flexibility lets it translate across. See the bridge-node note; graph node versatile-expertise.
The curse of expertise
The four cognitive failure modes of deep specialization without paired flexibility: paradigm lock-in, overconfidence in cross-application, perceptual filtering, conceptual rigidity. Together they are the cognitive substrate of polarization-via-distrust, the same mechanism inside one head. See the bridge-node note; source The Double-Edged Sword of Expertise; graph node curse-of-expertise.
The abyss
Everything a deep field could still ask or become, a space so large no one, expert or machine, can see all of it. It takes real competence just to notice it is there; beginners mistake the field for something finite. Not the same as a hard-to-use tool: a library is a map, not an abyss. The curse of expertise is depth that hasn’t seen the abyss; the bridge node is depth that has. See the abyss concept and The Abyss.
Selective isolation
The counter-pressure to integration: some specializations are useful because they are isolated, protecting a discipline’s internal standards from interdisciplinary pressure that would dilute them. The prescription has to integrate where outputs matter across networks and protect isolation where they don’t. See Chapter 9.
Curation layer
Institutional structures that preserve the gatekeeping function (separating signal from noise) while removing the access barrier (who gets to participate): Wikipedia editorial, Stack Overflow reputation, preprint-plus-peer-review. Where versatile experts do their work. See Chapter 9 and The Democratization Paradox; graph node curation-layer.
Trust-bootstrap problem
The hardest of Chapter 9’s operational problems: a curation institution earns trust by making good calls over time, but under advanced polarization its calls are pre-discounted by communities whose distrust is already set. See Chapter 9.
Preservation vs. training
What the book’s narrow “bureaucracy” splits into. Preservation keeps the full, un-compressed form alive somewhere; training re-installs the ability to decode it in people. They work as a pump (preservation holds the pressure, training releases it), so either alone fails. See Chapter 8; graph node preservation-vs-training-pair.
Capture asymmetry
When these institutions are captured, the training side does more damage and heals more slowly than preservation: a captured archive holds a tilted copy you can re-check elsewhere, while captured training tilts the readers themselves, and a re-tuned readership reads even good evidence through the bad key. So training is the half to defend first. See Chapter 8; graph node capture-asymmetry.
Survivable polarization
A design rule for bridge institutions: build them to work even when trust is partly broken, instead of needing it whole. In practice: decisions a skeptic can audit, reputation that survives partial distrust, real experts inside the institution, and several independent paths to the same answer. See Chapter 12; graph node survivable-polarization.
Funding decoupling
The structural counter to out-competition: an institution can’t survive if its revenue depends on winning the free-engagement market. The survivable shapes fund it from outside that market: endowments, public funding, member dues, public-goods mechanisms. See Chapter 12; graph node funding-decoupling.
Survival, not victory
The reframe at the heart of Part IV: integration institutions don’t need to beat the engagement machine, only to keep doing preserve-and-retrain alongside it without being eroded. Trying to win, by being more engaging, is how they get captured. See Chapter 12; graph node survival-not-victory.
Work of generations
Institutions of integration are civilization-scale investments on the timescale of universities, journals, and legal systems, not product launches. The captured equilibrium built itself in two decades; what holds against it has to be funded and protected for at least as long. See Chapter 12; graph node work-of-generations.