The Attention Machine

The long-standing Russian (and earlier Soviet) doctrine of influencing other countries by means short of war: disinformation, forgery, front organisations, and the amplification of existing social divisions. Online influence operations are this doctrine ported to a new medium — the goal is less to make you believe one specific lie than to corrode shared trust in institutions and in the idea of verifiable fact.

A simple framework given before the detail, into which the detail can then be slotted — Ausubel's idea that you learn complex material better when you first have a mental "map" to hang it on. The plain-language main text (Tier 1) is the advance organiser for this whole site: read it first and the depth blocks have somewhere to attach.

Not disagreeing more about policy, but disliking and distrusting the other side more — coming to see opponents as enemies rather than fellow citizens who are wrong. It is the kind of division that engagement-ranked feeds reward most, because contempt and us-versus-them content travels further than agreement. It corrodes the shared ground democracy needs even when people's actual policy views barely move.

The set of methods for establishing how old a user is — from self-declaration (a typed birthday, trivially defeated) through to document checks, facial age-estimation, or third-party verification. The central dilemma: the methods strong enough to actually work tend to require collecting identity data from everyone, including adults, creating a new surveillance risk in the act of protecting children. "Age assurance" is the broader term for approaches that aim to estimate age without full identification.

The construction of a predictive model of an individual from the traces they leave — what they click, how long they linger, what they share, who they follow. The profile is not a description you would recognise; it is a statistical object optimised to predict your next action. You never see it, and you cannot correct it.

Zuboff's term for the behavioural data left over after a service has been delivered to you — the click patterns, locations, and timings that aren't needed to run the product, but are captured anyway and fed into prediction. This surplus is the raw material of the business: it is refined into predictions about your future behaviour and sold in markets you never see.

The amount of information working memory is handling at once. Working memory is small and easily swamped; when a page demands more than it can hold, comprehension collapses — the facts "do not land." Cognitive load theory (Sweller) is the basis for keeping the main text spare and pushing detail into optional layers, so the reader chooses when to spend the capacity.

Networks of accounts that hide who is really behind them and act together to make a message look more popular or organic than it is — fake personas, paid amplification, bot-assisted sharing. The term comes from Meta's own enforcement vocabulary. What makes it "inauthentic" is the concealed coordination, not the content itself, which may be true, false, or merely divisive.

Interface designs that steer you toward the choice the company prefers rather than the one you would freely make: the giant "Accept all" button next to a greyed-out, three-clicks-deep "Reject". They exploit defaults, friction, and visual hierarchy. Under the GDPR a consent obtained this way is arguably not freely given — but the banners remain near-universal because the cost of ignoring the rule is low relative to the value of the data.

Representing a group as less than fully human — as vermin, disease, or invaders. It is not just insult: psychologically, it loosens the moral restraints that normally stop people from accepting or committing violence against others. Genocide scholars treat escalating dehumanising language as a recognised warning stage, which is why its mass amplification in Myanmar is documented so carefully rather than dismissed as mere offensive speech.

An influence strategy aimed not at making you support side A over side B, but at making you believe the whole system is rigged and no source can be trusted. It is cheaper than persuasion and harder to counter: you don't have to win an argument, only to exhaust people into disengagement. A population that trusts nothing is easier to rule and harder to mobilise in defence of institutions — which is why corroding trust, not winning votes, is often the actual objective.

A measure of how big a relationship is, separate from whether it is statistically real. Much of the dispute turns on this: a finding can be highly statistically significant (very unlikely to be chance) and still small in practical terms. Whether social media's measured effect on adolescent wellbeing is "small" or "meaningful at population scale" is a genuine disagreement about how to weigh the same numbers.

The practice of training a ranking system to maximise measurable behavioural signals — clicks, watch time, reshares, comments — used as a proxy for value. The proxy is convenient because it is easy to measure. It is also systematically misaligned with wellbeing: the content that holds attention is not the content that serves you.

Presenting two sides as evenly matched when the evidence is not — giving a fringe view equal airtime with an overwhelming consensus in the name of "balance". It is how a settled question (climate science, vaccine safety) can be made to look like a live controversy. The manufactured-doubt industry exploits it deliberately: you don't need to win the argument, only to make it look unsettled.

External examination of how a recommender system actually behaves — what it amplifies, to whom, with what effects — by parties who do not work for the platform. It requires genuine access to data and systems, not the platform's own summary reports. The recurring obstacle is the information asymmetry: the platform knows everything about its system and external researchers know comparatively little, which makes independent verification of any safety claim difficult. Closing that gap is itself a policy objective.

Data you never provided, generated by predicting your attributes from your behaviour — for example inferring a health condition from the articles you read. Under European data-protection law, inferences about you can be personal data, but you rarely know they exist, cannot see them, and cannot easily correct them. The profile that decides what you are shown is one you have no access to.

A term the WHO and UN used during COVID-19 for an overabundance of information — accurate and false mixed together — spreading faster than it can be evaluated, making it hard for people to find trustworthy guidance when they need it. The point of the word is the scale and speed: the problem is not that false health claims exist, which they always have, but that the feed distributes them at the velocity of a pandemic.

Building resistance to manipulation by exposing someone to a weakened form of the technique before they meet it for real — the psychological equivalent of a vaccine. Research led by Cambridge's Sander van der Linden shows that learning to spot tactics (scapegoating, false dilemmas, manufactured outrage) generalises across topics, where fact-by-fact debunking after the fact often does not. This is why the site favours teaching technique over arguing content.

The principle that an online service is not legally liable for content its users create, the way a phone company isn't liable for what people say on calls. It is what lets a platform host billions of posts — and moderate them — without being sued over each one. The hard question Section 230 left unanswered: does that immunity also cover the platform's own choice of which posts to amplify?

Under the GDPR a company may not process your personal data just because it wants to; it needs one of six "lawful bases", the most discussed of which is your freely-given, informed, specific consent. "Freely given" is the catch: consent buried in a pre-ticked box or extracted by a manipulative banner is not valid consent in law — which is why so much enforcement turns on how the choice was designed, not whether a box was clicked.

A study that follows the same people over a long period, measuring them repeatedly, rather than taking a single snapshot. It is the design best able to untangle cause from effect — does the feed worsen mood, or does low mood drive scrolling? — because it can see which came first. Its weakness is time: to know what childhood social media does across a life, you have to wait roughly that long.

Online slang for maximising one's physical attractiveness. "Softmaxxing" covers ordinary grooming and fitness; "hardmaxxing" covers extreme measures. Solea and Sugiura (2025) document how the term, along with the pseudoscientific "PSL" attractiveness scale, is used to repackage misogynist "incel" ideology as neutral self-improvement so it spreads more easily on mainstream platforms.

A strategy, not a state of knowledge: deliberately funding and amplifying contrarian research to make a settled question look unresolved, so that action can be delayed. Oreskes and Conway traced the same small playbook from tobacco to acid rain to climate. The tell is the goal — not to establish a competing truth, but to manufacture enough apparent controversy that "the science isn't settled" sounds reasonable.

Aiming a tailored message at a very narrow audience — sometimes individuals — defined by inferred traits rather than broad demographics. In politics, it means different voters see different, individually-optimised appeals that no one else sees, which makes the overall campaign hard to scrutinise or fact-check. Its real-world persuasive power is genuinely debated; its opacity is not.

A US procedure that pools many separate lawsuits making similar claims before a single judge for the shared pre-trial stage — gathering evidence once rather than thousands of times. It is not a class action: the individual cases keep their own identity and can be tried separately. MDL 3047 consolidates the social-media youth-harm suits; its scale (thousands of cases) is itself part of why the litigation is consequential.

The tendency for negative information to produce stronger and faster cognitive, emotional, and physiological responses than equivalent positive information.

The listings a search engine ranks by its own relevance signals, as opposed to paid results, which advertisers buy. On a typical commercial query the paid results sit on top and look almost identical to organic ones — the visual distinction between "ad" and "result" is deliberately faint, and many users do not notice where paid ends and organic begins.

Consuming a feed without interacting — scrolling, watching, reading, but not posting or messaging friends. It is the form most consistently linked to lower mood, plausibly because it is mostly upward comparison with none of the connection that makes real social contact restorative. The same hour spent messaging a friend is a different activity from an hour of silent scrolling.

The finding that averages hide opposite individual realities: a feature that lifts one person's mood lowers another's, so the population average can be near zero while real harms and benefits are large but cancel out. Valkenburg's work argues this is why the "small average effect" debate can mislead — the right question is often for whom, under what conditions, rather than how much on average.

A slang term (from "fattening the pig before slaughter") for a long-game investment fraud: the scammer invests weeks building a genuine-feeling relationship before introducing a fake, high-return investment platform. The trust is the product; the "investment" is the slaughter. Naming the structure matters because, once you know the relationship is the setup, the later pitch loses its power.

Debunking corrects a false claim after it has spread; prebunking teaches the manipulation technique in advance, so students recognise the move whatever specific claim it carries. The advantage is coverage and durability: there are endless false claims to debunk, but a small number of recurring techniques to inoculate against, and technique-level learning transfers across topics.

A branch of law that holds the maker of a product responsible when the product is built in a foreseeably harmful way — independent of anything a user did with it. Applied to platforms, the claim sidesteps Section 230 entirely: it is not about the content users posted (immunised) but about the company's own engineering choices (not immunised). The related "failure to warn" claim argues the company knew of harms and did not disclose them.

Targeting people by inferred psychological traits — personality, values, anxieties — rather than by demographics like age or location. A controlled study reaching over 3.5 million people found that ads matched to a person's personality produced up to 40% more clicks and up to 50% more purchases than unmatched ads. The same technique that sells running shoes can sell a political message.

The gradual drift from mainstream to extreme content, one small recommended step at a time — often called the "rabbit hole." Each suggestion is only slightly more extreme than the last, so no single step feels like a leap, but the cumulative direction is consistent. The engine isn't ideological; it has learned that more extreme content holds attention, and it optimises for attention.

The system that auctions each ad slot in the moment a page loads. To run the auction, a "bid request" describing the user — location, inferred interests, identifiers — is broadcast to many potential bidders at once. The privacy problem is structural: the data is shared with every participant in the auction, not only the winner, hundreds of times a day per person. The CJEU's IAB Europe ruling (Case C-604/22) found the framework underlying it raises serious GDPR problems.

The class of algorithms that decide what each user sees. They predict, from your past behaviour and the behaviour of similar users, which item you are most likely to engage with — and rank your feed accordingly. They are the engine of Facebook, YouTube, and TikTok.

Crucially, they optimise for predicted engagement, not for accuracy, importance, or your wellbeing. Those are simply not inputs to the model.

The finding, from the 2010s on, that a large share of published results — especially in psychology and parts of medicine — could not be reproduced when independent teams repeated the studies. It sounds like an indictment of science, and partly it is. But replication is itself a scientific method, and the crisis was found, named, and acted on by scientists — which is the self-correction the rest of this section is about, working in public.

Tim Wu's term for silencing speech not by banning it but by drowning it — overwhelming the information space with propaganda, distraction, troll campaigns, and automated content until genuine speech cannot be found or trusted. It is the modern censor's tool precisely because it never has to admit to censoring anything. It also explains why "let everything be said" no longer guarantees anything is heard: when attention is the scarce resource, whoever can manufacture the most noise wins, regardless of truth.

Leon Festinger's 1954 observation that people work out their own worth by measuring themselves against others, and that we reach hardest for comparisons when we are unsure of ourselves. The instinct is normal and useful in a small community. A feed that shows you thousands of carefully curated lives and faces a day turns it into a source of near-constant, mostly unfavourable comparison.

Naomi Eisenberger's fMRI studies showed that social rejection activates the dorsal anterior cingulate cortex — a brain region also active during physical pain. Approval activates the ventral striatum, the same region active when people receive food or money.

This is why social rejection genuinely hurts, and why a "like" can feel like a small chemical reward. The brain processes these in shared circuits, not separate ones.

The tendency to treat other people's behaviour as evidence of what is safe or smart — strongest under uncertainty, and strongest when those people are admired or seem similar to us. A scam delivered as an ad is easy to dismiss; the same scam delivered as a tip from someone you follow, or as a crowd of strangers all "winning," bypasses the scepticism an ad would trigger. The endorsement does the persuading.

Accounts run by an operator while posing as an unconnected, ordinary person — a fake "American voter" actually run from a troll farm abroad. At scale, a coordinated network of sockpuppets manufactures the appearance of a genuine grassroots movement (sometimes called "astroturfing"), which is what made the IRA's operation effective: the divisive content looked like it came from real fellow citizens.

The principle in US constitutional law that the Bill of Rights — including the First Amendment — limits what governments may do, not what private parties may do. A consequence many people find counter-intuitive: when a social platform removes your post, it has not violated your First Amendment rights, because it is not the state. The same act by a government would be censorship in the constitutional sense; by a company it is, legally, an editorial choice. This is exactly why the real debate is about private power, not constitutional rights.

Images, audio, or video generated or altered by AI to depict people saying or doing things they did not. The technical barrier has collapsed: a convincing voice clone now needs seconds of sample audio, and the result is cheap enough to produce at scale. The significance is less any single fake than the "liar's dividend" — once people know anything could be fake, genuine evidence can be dismissed as fabricated too.

The DSA's central innovation: instead of policing individual posts, it makes the largest platforms responsible for the society-level effects of their systems — harm to elections, civic discourse, public health, fundamental rights. They must assess these risks and mitigate them. It shifts the legal question from "is this post illegal?" to "does the design of this system cause foreseeable harm at scale?"

A four-step design loop — trigger, action, variable reward, investment — popularised by Nir Eyal in his 2014 book Hooked, written as a manual for building products people return to without thinking. It is a clear, honest description of the machinery, written from the inside. The same author later wrote a follow-up on how to resist it — which tells you something about what the loop does.

The long-standing argument — associated with John Stuart Mill and, in US law, with Justice Oliver Wendell Holmes — that the remedy for bad speech is more speech, not suppression, because truth tends to win an open contest. It is one of the strongest cases for a light touch. Its critics do not reject it so much as note its hidden assumptions: it pictures a town square of human speakers, not an engagement algorithm choosing what a billion people see. When reach is sold and automated, the "market" is no longer deciding between ideas on their merits.

A reward schedule where the payoff comes after an unpredictable number of tries, rather than every time or on a fixed timetable. In mid-twentieth-century experiments, B.F. Skinner found this schedule produces the most persistent, hardest-to-stop behaviour — which is exactly why slot machines use it. A feed that sometimes rewards a refresh, and a notification that might be something good, run on the same principle. The uncertainty is the hook, not a flaw.

The discipline of confirming a claim against independent evidence before publishing — and being willing to kill a story that won't stand up. It is the single feature that most cleanly separates journalism from propaganda: propaganda is not trying to find out whether something is true, it is trying to produce a reaction. Verification is slow and expensive, which is exactly why engagement-optimised systems route around it.

The DSA's tier for platforms with more than 45 million monthly active users in the EU — roughly a tenth of the population. VLOPs carry the heaviest obligations precisely because their scale makes their design choices a matter of public consequence rather than private business. The Commission, not national regulators, supervises them directly.

Europol's term for criminal networks that outsource violence: those who order and profit from it stay hidden, while the risk is pushed onto recruited operatives — increasingly minors with no criminal record, who are seen as less likely to be caught or to inform. Europol's EU-SOCTA 2025 threat assessment identified it as a fast-growing model across Europe, and the agency set up a dedicated taskforce in 2025 in response.

Berridge and Robinson distinguished two reward systems. "Wanting" (also called incentive salience) is driven by dopamine in the mesolimbic pathway and produces the urge to seek. "Liking" is mediated by opioid systems and produces the experience of pleasure.

The two can be dissociated experimentally. You can want something intensely and not enjoy it when you get it.

A search that ends on the results page — the user gets their answer from a snippet, knowledge panel, or AI summary and never clicks through to a source. Convenient for the user, but it strips away the context that lets you judge an answer: who wrote it, how they're funded, whether they correct mistakes. The search engine becomes the answer rather than a route to one.