The Broadcast and the Model.

---

Abstract

For six decades the study of media effects has returned a reassuring verdict: persuasion is hard, mass messages move people little, and confident manipulation is mostly myth. That verdict is real but bounded — an estimate taken under one condition, symmetric broadcast exposure, not a law of human persuadability. Citing it as reassurance about the personalized, machine-mediated regime now arriving is not a category error but a scope-condition violation: a boundary estimate extrapolated past its boundary, a warning conditional on the boundary actually being crossed. The argument proceeds in two registers, holding documented effect sizes apart from forward projections that each name a falsifier, and concedes the record in full — including the findings least convenient to any alarmist reading: the microtargeting nulls, the roughly five-percent ceiling on psychographic targeting, machine–human parity in single-message persuasiveness, and the diminishing returns to model scale. Because personalization is not the primary lever, the argument does not rest on it. What the measured record now shows is that adaptive, multi-turn dialogue already moves both epistemic beliefs and electoral preference — through fact-density, not personalization — at magnitudes the evidence will bear, and now with a first durability datum: on re-contact weeks later, roughly one-third of the preference shift persisted and two-thirds had decayed, a profile the broadcast-era lab/field ledger would have predicted. Whether adaptation adds anything beyond matched fact-density, and whether any of it survives unforced exposure, are the cells still open. The contribution is correspondingly narrow: a scope-condition correction to how the minimal-effects literature is invoked, updated for the conversational-AI era, and a measurement agenda. The paper situates itself against Bennett and Iyengar’s (2008) “new era of minimal effects”: they saw the scope conditions eroding and reasoned that fragmentation would sustain weak persuasion through selective exposure; the lever now arriving — conversational adaptation — layers onto their account rather than inverting it, while their selection dynamics are precisely what may keep it from scaling.

Keywords: media effects, persuasion, artificial intelligence, conversational AI, computational propaganda, information asymmetry, surveillance, public sphere

§ IThe Question and the Two Registers

Two kinds of thinking run through this paper, and its discipline is to keep them apart. One is retrospective: what does the documented record show about how mass communication has moved people? It answers in effect sizes and scope conditions. The other is prospective: what can a channel built on behavioral data and machine modeling be expected to do? It answers in trends, assumptions, and the observations that would prove it wrong. I label these Register A — what we know — and Register B — what we can expect, and admit a projection only if it names its falsifier. This is good practice rather than a novel instrument; I flag it because the characteristic failure of the genre is to let a Register-B intuition wear Register-A authority. The genre here is review-and-reframe: the empirical claims are conceded from others, the projections are flagged open, and the yield is a measurement agenda.

Both sides of the existing debate share an unexamined premise. Techno-panic and minimal-effects complacency look opposite — influence is nearly total; influence is nearly nil — but they measure the same object, the message, and quarrel only over its potency. Each assumes the broadcast structure: one message, many receivers, symmetric exposure, a shared space in which the message can be seen, sourced, and contested. The narrow thesis of this paper is that the minimal-effects record was compiled inside that structure and is scope-conditioned to it. Treating it as a timeless law — rather than as a precise description of message potency under symmetric exposure — extrapolates a boundary estimate beyond its boundary. The correct charge is a scope-condition violation, or out-of-distribution extrapolation, not a category error. And the charge is prospective: by the evidence assembled in §5, exposure is still substantially shared, so the violation is not yet badly committed — the warning is “do not extrapolate minimal effects to the regime now arriving,” not “citing them about the present is already wrong.”

This scope-condition insight is not new, and saying so is part of making it well. Bennett and Iyengar (2008) argued that the foundations of the minimal-effects paradigm — stable group anchors, the two-step flow, a limited media menu — were eroding, and a pointed exchange followed over whether selective exposure and reinforcement would follow from fragmentation (Holbert, Garrett, and Gleason 2010). What this paper adds is not the diagnosis that the conditions are changing but a claim about the lever that changes them. Writing before conversational machines, Bennett and Iyengar reasoned that fragmentation would tend to sustain weak persuasion by sorting audiences into reinforcing environments. The mechanism isolated here — adaptive, information-dense interaction at scale — is one their account could not anticipate, and it does not so much invert their prediction as layer onto it: they reason at the selection margin (who chooses to encounter what), while this lever operates conditional on engagement (what happens once a person is in the dialogue). The two stack — and the stacking cuts against easy alarm, because selective exposure is plausibly the very thing that gates the lever’s reach. Who volunteers for an eight-minute exchange built to change their mind? Recent election-context experiments moved opposed voters’ stated preferences under assigned engagement — a setting on which a reinforcement account is silent rather than one it forbids — so what remains genuinely open is reach under unforced exposure. And one consequence is stated now rather than discovered later: if adaptive persuasion fails under unforced exposure, that is the framework’s central projection failing (§6’s F3), not the selection account absorbing the result as confirmation. The paper does not get to keep the alarm and bank the reassurance on the same observation.

The boundary can be named in the vocabulary of information economics, where it has lived since Akerlof (1970), Spence (1973), and Grossman and Stiglitz (1980): it is information asymmetry between the parties to an exchange. Let Δ stand for that asymmetry between persuader and target. It is useful as a heuristic on one condition — that it not be collapsed into a single quantity. Δ has two components that need not move together: a targeting component (how well the persuader can model this receiver, drawing on behavioral data) and a capability/deployment component (how well the persuader can marshal and deploy argument the receiver cannot match or evaluate). The later sections show these components pointing in different directions — the targeting one weak in the evidence, the capability one strong — which is exactly why Δ must be split rather than summed. Δ is not yet independently measurable; its empirical content lives in the projections of §§7–8 and their falsifiers, stated plainly rather than letting an unobservable variable do covert work.

§ IITwo Caricatures to Clear

Two caricatures block the road, and each is a way of not asking the structural question.

The first is the hypodermic-needle revival: the recurring belief that a new medium makes the public directly programmable. Its founding examples are myths. The 1938 War of the Worlds “panic,” long taught as the origin of powerful-media theory, was largely manufactured afterward — by newspapers discrediting radio and researchers wanting a dramatic founding moment — with almost no one verifiably deceived (Pooley and Socolow 2013). The companion legend, that television “lost Vietnam,” does not survive the coverage record either: Hallin (1986) found that critical coverage rose roughly as elite consensus fractured, a correlation he reads as indexing rather than as proof that television drove opinion — an account dominant without being unanimous.

The second caricature is more respectable and more dangerous to the argument: minimal-effects complacency, the inference that because mass persuasion has been measured and found weak, it will stay weak. The response is not to dispute that record — §4 defends it at full strength — but to locate it. It is a body of measurements taken under symmetric broadcast exposure. To read it as a law of persuadability is the scope-condition violation named in §1. The complacent and panicked readings make the same mistake from opposite ends: both hold the structure fixed and argue about the number. This paper holds the number and interrogates the structure.

§ IIIWhat We Know I: The Broadcast-Era Mechanisms

The broadcast era produced a real but bounded science of influence, and Register A reports it at its measured size. Agenda-setting is the most robust mechanism: across sixty-seven studies the media agenda and the public agenda correlate at about 0.49, a moderate association whose causal direction is still contested (McCombs and Shaw 1972; Luo et al. 2019). Framing has the cleanest causal warrant because it was tested experimentally — Iyengar (1991) showed that episodic versus thematic presentation shifts where viewers assign responsibility. Cultivation is real but small: a five-decade meta-analysis puts it near 0.11, and even that shrinks under simultaneous demographic controls (Gerbner and Gross 1976; Hermann, Morgan, and Shanahan 2021). The spiral of silence runs near 0.10 overall, rising to about 0.34 only among close contacts (Matthes, Knoll, and von Sikorski 2018). Two further frameworks are routinely misread as effects research: the propaganda model describes gatekeeping, not measured audience conversion (Herman and Chomsky 1988), and Lippmann (1922) and Bernays (1928) supply the genealogy rather than the evidence. The honest summary: correlations clustering between 0.10 and 0.49, mostly cross-sectional, and — the point that carries forward — measured almost entirely under shared, symmetric exposure.

§ IVWhat We Know II: The Effect-Size Record

This is the chapter that earns the paper’s standing, and it does so by conceding the most deflationary evidence in the field rather than fighting it.

The deflation. Media research was born deflationary: Klapper’s (1960) reinforcement doctrine, Lazarsfeld’s small persuadable share (Lazarsfeld, Berelson, and Gaudet 1948), and Hovland’s (1959) lab/field gap — a third to half of an audience moving in the lab, almost no one in the field. The modern verdict has only hardened. Kalla and Broockman’s (2018) synthesis of forty-nine field experiments puts the best estimate of persuasive effect on candidate choice in U.S. general elections at approximately zero. Coppock, Hill, and Vavreck (2020) found small and stubbornly uniform advertising effects across fifty-nine experiments. A roughly two-million-person experiment found no detectable average turnout effect of presidential digital advertising (Aggarwal et al. 2023). Removing political ads from some sixty thousand users’ feeds for six weeks before the 2020 election produced a null across every pre-registered outcome (Allcott et al. 2026).

The Cambridge Analytica deflation, in full. The most-cited proof that the new tools broke the ceiling is the most deflated. Trait inference from Facebook Likes tops out near 0.4 — under a fifth of a trait’s variance (Kosinski, Stillwell, and Graepel 2013). The cleanest persuasion estimate is small, commercial, and contested on delivery-optimization grounds (Matz et al. 2017; Eckles, Gordon, and Johnson 2018). The best synthesis to date finds the chain weak at its first link: digital footprints predict about five percent of personality variance (Perla et al. 2026), before any loss from inference to message-matching to persuasion is counted. A regulator with the firm’s internal documents concluded it could not accurately predict personalities and did not deploy psychographics as the legend holds (Information Commissioner’s Office 2018). For 2016’s fake-news stories to have swung the election, each would have needed the force of roughly thirty-six television advertisements — a rate nothing in this literature approaches (Allcott and Gentzkow 2017). The structural argument built later does not depend on this case having worked, and is strengthened by the fact that it did not.

An outcome taxonomy — because* “persuasion” *is not one thing. A distinction the field routinely blurs is load-bearing here: the studies above measure different dependent variables that do not move alike. Three classes matter. Epistemic/factual beliefs (is this claim true?) are comparatively movable by good argument. Identity- and preference-anchored commitments (which party, which brand, which candidate) are sticky. Behavioral/mobilization outcomes (turnout) move on norms and logistics, not attitude. The classes are not commensurable, and the fault line is consequential: the paper’s worry — industrialized persuasion — is classically about the preference/identity class, while the deflationary record above is largely a record of that class proving hard to move. The sections that follow state which class each later claim concerns, and resist comparing magnitudes across classes.

What actually moves people. Sobriety is not the claim that nothing works; large effects exist, but they are not modeling effects. The largest single effect in the field is a social-pressure mailer that raised turnout 8.1 points by disclosing a recipient’s voting record to neighbors — norms, not a private model (Gerber, Green, and Larimer 2008). The most valuable targeting signal is the public record of who votes, not the private inference of who one is (Nickerson and Rogers 2014). Where campaigns persuade, the mechanism is informational: voters move most on the candidates they know least about, i.e., where a genuine information gap closes (Broockman and Kalla 2023).

§ VWhat We Know III: The Structural Shift Already Underway

The structure under which §§3–4 were measured is changing, and the change is documented present, reported with measured figures. Attention is the scarce resource (Simon 1971), and the attract-and-resell model has run for nearly two centuries (Wu 2016); what is new is the extraction of behavioral data as the raw material for prediction (Zuboff 2019). The substrate is vast and partly audited: one broker held seven hundred billion data elements, another added three billion data points a month (Federal Trade Commission 2014). U.S. digital advertising revenue reached $258.6 billion in 2024, with each impression auctioned on a behavioral profile in under a tenth of a second (Interactive Advertising Bureau 2025). Modeling capacity is concentrating: a court found Google holding about ninety percent of publisher ad servers (United States v. Google 2025), the Google–Meta duopoly takes roughly forty-seven percent of digital ad spending, and dominant media ownership ran from about fifty firms to five over two decades (Bagdikian 2004).

One premise must be stated with care, because it is more contested than alarm allows. The picture of sealed “private models” and filter bubbles is empirically tempered: individual choice outweighs the algorithm in limiting cross-cutting exposure (Bakshy, Messing, and Adamic 2015), and most Americans’ media diets are moderate, with concentrated echo chambers confined to a small if disproportionately visible minority (Guess 2021). So the honest claim is narrow: the conditions of shared, symmetric exposure under which the minimal-effects record was compiled are eroding at the margin — through cable, the algorithmic feed (whose isolating effect is itself contested), and now individuated addressability — without yet having produced the hermetic personalization the strong story imagines.

§ VIThe Scope Condition

This section is the operating system the rest of the paper runs on, offered as framework rather than finding. The minimal-effects estimates are boundary conditions of symmetric broadcast exposure, and extrapolating them to the personalized regime now arriving is a scope-condition violation. The diagnosis that those conditions are eroding belongs to Bennett and Iyengar (2008); the contribution here is to identify the lever that does the eroding — a lever that layers onto their account rather than inverting it, and that their own selection dynamics may gate. The more dramatic framing — that we are moving from a regime of messages to a regime of per-target models, “a different machine” — is a heuristic, a motivating image, not a claim the current evidence earns. It earns demotion for a specific reason: the empirical lever that actually raises machine persuasion is largely message- and capability-side, which is closer to the old regime than to a surveillance-driven new one.

Δ, the persuader–target information asymmetry, must be split or it misleads. Its targeting component (a per-target model built from behavioral data) and its capability/deployment component (the marshaling of argument the target cannot match or evaluate) are distinct, and the evidence has them pointing in opposite directions. A persuader can be data-rich and argument-poor, or the reverse. Summing them into one Δ buys a tidy thesis at the cost of truth; keeping them separate lets §7 report that the targeting component is weak and the capability component strong. This also disciplines an obvious objection: a well-staffed advertising agency of 1965 and a skilled debater both enjoyed capability asymmetry, and the broadcast era had plenty of both while minimal effects held. So capability asymmetry alone is not new.

A third discipline on Δ must be added, and it is a limitation to scope rather than repair: Δ is dyadic. It models one persuader and one target, and the political world contains rival persuaders. If the adaptive lever is cheap, both sides deploy it, and the aggregate effect can net toward zero even while each side’s gross effect is real — Coppock, Hill, and Vavreck’s uniform smallness is readable not only as a fact about persuadability but as an equilibrium fact about symmetric, competitive campaigning. The framework’s honest domain is therefore the asymmetric case — persuaders facing no matched counter-persuader: a state actor abroad, commercial claims below the threshold of organized rebuttal, a first mover in an unregulated channel — and the transition window before symmetry is restored. Inside the symmetric case, a widening Δ on both sides predicts arms-race expenditure, not net movement, and the measurement agenda of §10 inherits that distinction.

What, then, is new? Not personalization-from-surveillance, which the evidence does not support, and not denser argument as such, which is a difference of degree within the message frame. It is also not individuation at scale: mail-merge, programmatic advertising, and robocalls already sent a different artifact to each of millions. The genuinely new element is real-time adaptation — a system that conducts a responsive, multi-turn argument that updates to each interlocutor — with scale as the multiplier that would convert a laboratory capability into a population-level one. The debater can adapt but cannot scale; the broadcaster scales but cannot adapt; a deployed conversational model could do both at once. Whether that is a difference of degree or of kind is left open here — it is the question the paper most wants measured, not one to settle by assertion.

The paper commits to a single numbered set of falsifiers, carried into §7. (F1) Targeting null: model-personalized messaging keeps matching a strong untargeted message as data and models improve. (F2) Capability plateau: machine persuasion levels off without exceeding a strong generic message. (F3) Adaptation doesn’t scale: adaptive multi-turn systems fail to beat single-shot persuasion once deployed under unforced exposure. (F4) Synthetic media stays sub-message: its persuasive impact remains at or below documented message-effect magnitudes. F3 carries the most weight, because the adaptive tier is where the one candidate for a genuine break lives.

§ VIIWhat We Can Expect: The Capability Tiers

Register B proceeds by tier, not by a single forecast, reporting each tier’s documented trend, its load-bearing assumption, and its falsifier. The tiers do not rise together; the argument rests its weight on the third, not the first.

Tier 1 — Targeting / personalization (the component the argument does not lean on). The trend is mixed and, on balance, unfavorable: a minority of studies find personality-matched lift (Matz et al. 2024), but the better-identified tests find sociodemographic microtargeting no better than untargeted messaging (Hackenburg and Margetts 2024), and footprint-based personality prediction sits near five percent of variance (Perla et al. 2026) — the first link in the chain, with every later link leaking further. One positive entry must be tallied here rather than set aside: a structured-debate study in which a model given basic demographics beat human opponents 64.4% of the time, the lift coming from personalization and vanishing without it (Salvi et al. 2025) — minimal-demographic tailoring in a debate format, the tier’s strongest showing and its boundary. The assumption under which this component would matter is that richer person-level data lifts effect past a strong generic message as models improve. Falsifier F1 — continued nulls as data and model quality rise — is already partly observed, which is why targeting is treated as the component most likely to be empty, though the Salvi entry keeps it from being written off outright.

Tier 2 — Information-dense generation (a degree shift in the message frame). The strongest study to date finds machine persuasiveness rises chiefly through post-training and information-dense prompting — not personalization or scale — with higher persuasiveness tracking lower factual accuracy (Hackenburg et al. 2025). Held to the same standard as Cambridge Analytica’s odds ratios: the reported boosts of “up to 51%” and “27%” are relative increases on an already-modest base — typical configurations move attitudes only single-digit points on the studies’ scales — and the often-quoted ~25-point figure is the persuasion-optimized model on opposition voters specifically, an upper bound rather than the central estimate. This is a denser, better-argued message, not a per-target model; it intensifies the old phenomenon rather than replacing it. Falsifier F2 — machine persuasion levels off without exceeding a strong generic message — is gestured at by the diminishing-returns-to-scale and human-parity findings (Hölbling, Maier, and Feuerriegel 2025). A distinct corollary is that these gains may depend on the target’s inability to verify the claims; if so, the remedy sits on the target’s side (§10).

Tier 3 — Adaptive multi-turn interaction (where the novelty lives, and where the argument is most exposed). Multi-turn adaptive dialogue moves belief in both outcome classes as of late 2025 — and durability, long the missing column, is now a measured quantity, with the measurement mixed. On the epistemic side, an eight-minute exchange cut conspiracy belief by about twenty percent — some 12 to 17 points on a 100-point scale — and held at two months, with accurate argument (Costello, Pennycook, and Rand 2024) — a finding that must now be cited as contested: Science issued an Editorial Expression of Concern (June 11, 2026) over inconsistencies in the application of screening criteria and a code-merging error in the public dataset, the authors reporting that a corrected pipeline reproduces the results in direction, significance, and size, with the journal’s evaluation still open. The effect’s independent replication stands on its own data and is driven by the facts deployed rather than the AI messenger (Boissin et al. 2025). On the preference side, pre-registered, election-context dialogues moved opposition voters — about 1.5 to 3.9 points in the United States on a 100-point scale (a 2×–5.6× multiple of the 2016/2020 video-advertising benchmark, depending on the comparison; the often-quoted “roughly four times” reads off the upper half of that range), and about 10 points in Canada and Poland (Lin et al. 2025). (A debate study in which a model beat human opponents 64.4% of the time is tallied under Tier 1: its lift came from personalization and vanished without it, so it weighs in the targeting column, not adaptation [Salvi et al. 2025].) The lever in both classes is fact-density, not personalization or psychological manipulation.

Three gaps remain. The first is mechanism: no cited design compares adaptive dialogue against a static message of matched fact-density, so adaptation’s increment over dense argument is asserted nowhere in the record — and the tiers’ own numbers keep the question honest: Tier 2’s persuasion-optimized static messaging reaches its ~25-point upper bound on opposition voters, while the adaptive dialogues here move 1.5 to 3.9 points in the same population class. The designs and scales differ too much to subtract one from the other, but their order is the reverse of what an adaptation-first story predicts; degree-versus-kind stays open exactly as §6 concedes. The second is direction: in both flagship studies persuasiveness rises as factual accuracy falls — in the election experiments, the model advocating right-leaning candidates made more inaccurate claims across all three countries — so the malign mirror (persuading toward falsehood by the same fact-flooding) needs no new mechanism. The authors’ proposed account is fact-exhaustion: a model pressed for ever more claims runs out of accurate ones and begins to fabricate, which makes the worry less “lies persuade better” than “persuasion optimized past the supply of true facts,” and points the remedy at the target’s verification capacity (§10). The third gap, now load-bearing, is deployment and persistence: every one of these studies assigned its participants, and none has tested unforced exposure — but persistence is no longer unmeasured. The election study’s own follow-up re-contacted participants weeks later and found roughly one-third of the preference shift persisting, two-thirds decayed (Lin et al. 2025). That number must be read against the template it repeats. Hovland’s lab/field gap has a modern persuasion record behind it in which large assigned-exposure effects decay savagely — televised ad effects measurable in the week of airing and gone within a week or two (Gerber et al. 2011), pretreatment dynamics that shrink lab effects in saturated information fields (Druckman and Leeper 2012) — while deep canvassing showed a decade ago that responsive dialogue persuades durably (Broockman and Kalla 2016) and industrialized nothing. So the rival reading stands in this text as the live rival it is: the 2025 assigned-exposure experiments may be new entries in the lab column of §4’s ledger — Hovland’s inversion re-run with a machine interlocutor — rather than Register-A evidence of a new lever, and the surviving third is the datum both readings must now fight over. The nearest neighboring durability test points the deflationary way: a vaccine-intention chatbot trial found the chatbot’s effect present at fifteen days but gone by forty-five, while standard public-health materials still carried a modest effect at forty-five (Sehgal et al. 2026) — which points the same way as F3. Falsifier F3 is accordingly sharp: not “does preference move at all” (it does, in panel) but “does it move among people who did not choose a mind-changing conversation, and does the third that survives re-contact survive the field.” The gate is no longer evidence-free: disclosed-AI political outreach draws systematic penalties — judged less acceptable, more threatening to personal autonomy, more corrosive of trust, independently of message content (Jungherr and Rauchfleisch 2026) — direct evidence that consent to the conversation is where the lever binds. And the falsifier now specifies its design rather than gesturing at one: recruit for something other than persuasion, disclose the machine, leave engagement optional, and measure preference at thirty and ninety days against an assigned-dialogue arm. Its chief threat remains the lab/field-and-time gap conceded in §4, and the seam is selection — which the election study’s own authors name when they note that engagement is “a high bar to clear.”

Tier 4 — Synthetic media. Generation outpaces detection, but the documented effect runs more through eroded certainty and trust than through direct deception (Vaccari and Chadwick 2020). The assumption is scale plus degrading detection, with the epistemic-corrosion channel mattering as much as deception. Falsifier F4 — persuasive impact at or below documented message-effect magnitudes — would make this broadcast-era persuasion at higher volume, not a regime change.

An evidence map for the adaptive tier. Table 1 tabulates the multi-turn/adaptive studies by outcome class and setting, with an explicit laboratory-to-field discount, and deliberately computes no pooled estimate: the classes are not commensurable. The adaptive lever is now demonstrated in both classes — epistemic (benign, durable in panel, presently under an expression of concern) and preference (movement among the opposed, one-third of it surviving re-contact) — through the fact-density lever the framework predicts. What the map still shows empty is not a class but a setting: every study assigned its participants, and none tested unforced exposure, so against the field record’s savage attenuation of preference effects (Kalla and Broockman 2018) the magnitude under real, unforced exposure remains unmeasured even though existence in panel no longer is.

Table 1. Adaptive / multi-turn AI persuasion: an evidence map (native units; no pooled estimate by design).

§ VIIIConsolidation: A Research Proposal, Not a Result

A consolidation argument is worth posing but must be labeled honestly. The intuitive version — that mergers concentrate modeling capacity and thereby widen Δ — aims at the targeting component, which §7 finds weak. If a richer per-target model buys little persuasive lift, then consolidating per-target data inherits that near-zero value; pooling data-broker holdings does not, on present evidence, buy persuasive power.

The version that could bite runs through a different asset. The active levers in §7 — post-training and adaptive deployment — are produced by behavioral-feedback data and the capacity to deploy: the engagement data that post-train a persuasion-capable model, the compute to do so, and the distribution chokepoints through which a system reaches individuals at scale. Whether acquisitions concentrate those assets, and whether that concentration translates into persuasive advantage, is an open empirical question this paper poses rather than answers.

The regulatory doctrine is worth stating because it measures none of this. Merger review reaches price and, increasingly, conduct: a lone dissent in 2007 proposed a market for behavioral-targeting data and was not followed (Harbour 2007); the 2025 monopolization ruling reached Google’s ad-tech dominance through conduct, not a data-market theory (United States v. Google 2025); the judgment for Meta — entered on the merits after a six-week bench trial, on the finding that the agency had not proven present monopoly power (FTC v. Meta 2025) — showed data-advantage theories outliving the market-share monopoly the doctrine needs to act. But even a reformed merger doctrine would be aimed at data holdings, whereas the lever the evidence supports lies in training and deployment. Naming that mismatch — and the genuine thinness of measurement on either side — is this section’s contribution.

§ IXThe Ledger: What We Claim and What We Do Not

The empirical claims (Register A) are confined to §§3–5: the broadcast mechanisms at their measured magnitudes; the near-zero persuasive effects of campaign messaging and the deflation of Cambridge Analytica; and the documented present of attention markets, extraction, addressability, and concentration. The projections (Register B) are confined to §7’s tiers, each with a falsifier, none asserted to have happened. The heuristic (§6, and Δ) is labeled as such: the scope-condition argument is the claim; “message→model” is the motivating image.

What the paper declines to claim is as important. It does not claim that personalization-from-surveillance is the primary persuasion mechanism — the evidence makes it secondary at best, though tailoring may add at the margin. It does not claim that consolidation has been shown to widen the effective asymmetry; §8 is a proposal. On the strength of the 2025 election experiments it does claim that assigned multi-turn dialogue can move preference commitments in panel — single digits on a 100-point scale among opposition voters, with roughly one-third of the shift surviving re-contact weeks later; it does not claim that these effects survive unforced exposure, that the surviving third scales, or that adaptation rather than fact-density is the operative ingredient. It does not claim a demonstrated net harm: the advocacy direction was randomized for ethics, the lever is argument rather than psychological manipulation, and whether such persuasion is on balance harmful, benign, or mixed — the same machinery debunks conspiracies and may mislead by omission — is exactly what deployment studies must establish. It does not claim effects are or will be total; the asymmetry heuristic describes a rising marginal effect, not an irresistible one. And it attributes no coordinated intent to any named actor: the account is structural and incentive-based — there is no cabal in it, only capabilities and the markets that reward them.

§ XImplications: Change the Measurement, and Invest in Resistance

A structural argument earns its keep by correcting the instruments before anyone reaches for a remedy. The diagnostic shift is to stop citing 1960 as comfort and instead measure what the evidence says carries effect: the persuasive efficacy of adaptive, information-dense systems, assessed directly and longitudinally, and the degree to which capability rather than targeting is the operative lever. This is a measurement agenda, not a policy menu, and it asserts no instrument as following directly from it.

There is, however, one remedy the evidence already supports, and it falls out of Tier 2’s corollary. If information-dense persuasion depends on the target’s inability to verify claims, then raising the target’s side of the asymmetry is the lever with traction. That is the inoculation, or prebunking, literature, foundational since McGuire (1964) and now demonstrated at scale: short prebunking videos delivered to roughly 5.4 million users improved recognition of manipulation techniques at about five cents per view, with effects consistent across the political spectrum (Roozenbeek et al. 2022). Two true statements must be kept apart: inoculation is well-evidenced in general; whether it specifically neutralizes information-dense AI persuasion is untested, resting on the Tier-2 corollary that such persuasion depends on the target’s inability to verify. So the honest claim is directional — the productive place to look for leverage is the target’s evaluative capacity rather than the message — especially under the authors’ proposed fact-exhaustion mechanism, in which the accuracy deficit grows precisely as a model is optimized for persuasion.

§ XIA Research Program

The paper closes by handing its open questions forward as work — three studies, and one political-economy question beside them. The first isolates the two components of Δ: well-identified field experiments that separate the targeting contribution from the capability/deployment contribution, since the argument’s whole structure depends on their not co-moving. The second is the one the argument most needs and is least able to supply: whether adaptive, multi-turn persuasion survives deployment under unforced exposure rather than assigned dialogue, and whether the third of the panel shift that survives re-contact survives the field’s longer clock — §7 sketches the design: off-persuasion recruitment, a disclosed machine, optional engagement, preference measured at thirty and ninety days against an assigned arm. This is the question on which the lab/field-and-time gap bears directly, and which outside scholars independently stress in noting that getting voters into long chatbot conversations may prove a niche activity rather than a mass channel. The third tests the §10 remedy: whether prebunking, demonstrated against manipulation techniques in general, actually blunts information-dense AI persuasion in particular. Beside these sits §8’s question — whether acquisitions concentrate the behavioral-feedback and deployment assets the active levers use. The two-register method and a passage-level evidence corpus are offered as reusable hygiene, useful rather than novel.

---

References

Aggarwal, M., Allen, J., Coppock, A., Frankowski, D., Messing, S., Zhang, K., Barnes, J., Beasley, A., Hantman, H., & Zheng, S. (2023). A 2 million-person, campaign-wide field experiment shows how digital advertising affects voter turnout. Nature Human Behaviour, 7, 332–341.

Akerlof, G. A. (1970). The market for “lemons”: Quality uncertainty and the market mechanism. Quarterly Journal of Economics, 84(3), 488–500.

Allcott, H., & Gentzkow, M. (2017). Social media and fake news in the 2016 election. Journal of Economic Perspectives, 31(2), 211–236.

Allcott, H., Gentzkow, M., Levy, R., et al. (2026). The effects of political advertising on Facebook and Instagram before the 2020 US election. Nature Human Behaviour. https://doi.org/10.1038/s41562-025-02328-w

Bagdikian, B. H. (2004). The new media monopoly. Beacon Press.

Bakshy, E., Messing, S., & Adamic, L. A. (2015). Exposure to ideologically diverse news and opinion on Facebook. Science, 348(6239), 1130–1132.

Bennett, W. L., & Iyengar, S. (2008). A new era of minimal effects? The changing foundations of political communication. Journal of Communication, 58(4), 707–731.

Bernays, E. (1928). Propaganda. Horace Liveright.

Boissin, E., Costello, T. H., Spinoza-Martín, A., Rand, D. G., & Pennycook, G. (2025). Dialogues with large language models reduce conspiracy beliefs even when the AI is perceived as human. PNAS Nexus, 4(11), pgaf325.

Broockman, D., & Kalla, J. (2016). Durably reducing transphobia: A field experiment on door-to-door canvassing. Science, 352(6282), 220–224.

Broockman, D. E., & Kalla, J. L. (2023). When and why are campaigns’ persuasive effects small? Evidence from the 2020 U.S. presidential election. American Journal of Political Science, 67(4), 833–849.

Coppock, A., Hill, S. J., & Vavreck, L. (2020). The small effects of political advertising are small regardless of context, message, sender, or receiver. Science Advances, 6(36), eabc4046.

Costello, T. H., Pennycook, G., & Rand, D. G. (2024). Durably reducing conspiracy beliefs through dialogues with AI. Science, 385(6714), eadq1814. [Editorial Expression of Concern issued June 11, 2026; the authors’ corrected pipeline is reported to reproduce direction, significance, and size, with the journal’s evaluation open.]

Druckman, J. N., & Leeper, T. J. (2012). Learning more from political communication experiments: Pretreatment and its effects. American Journal of Political Science, 56(4), 875–896.

Eckles, D., Gordon, B. R., & Johnson, G. A. (2018). Field studies of psychologically targeted ads face threats to internal validity. Proceedings of the National Academy of Sciences, 115(23), E5254–E5255.

Federal Trade Commission. (2014). Data brokers: A call for transparency and accountability. FTC.

FTC v. Meta Platforms, Inc., No. 1:20-cv-03590-JEB (D.D.C. 2025).

Gerber, A. S., Gimpel, J. G., Green, D. P., & Shaw, D. R. (2011). How large and long-lasting are the persuasive effects of televised campaign ads? Results from a randomized field experiment. American Political Science Review, 105(1), 135–150.

Gerber, A. S., Green, D. P., & Larimer, C. W. (2008). Social pressure and voter turnout: Evidence from a large-scale field experiment. American Political Science Review, 102(1), 33–48.

Gerbner, G., & Gross, L. (1976). Living with television: The violence profile. Journal of Communication, 26(2), 76–97.

Grossman, S. J., & Stiglitz, J. E. (1980). On the impossibility of informationally efficient markets. American Economic Review, 70(3), 393–408.

Guess, A. M. (2021). (Almost) everything in moderation: New evidence on Americans’ online media diets. American Journal of Political Science, 65(4), 1007–1022.

Hackenburg, K., & Margetts, H. (2024). Evaluating the persuasive influence of political microtargeting with large language models. Proceedings of the National Academy of Sciences, 121(24), e2403116121.

Hackenburg, K., Tappin, B. M., Hewitt, L., Saunders, E., Black, S., Lin, H., Fist, C., Margetts, H., Rand, D. G., & Summerfield, C. (2025). The levers of political persuasion with conversational AI. Science. https://doi.org/10.1126/science.aea3884

Hallin, D. C. (1986). The uncensored war: The media and Vietnam. Oxford University Press.

Harbour, P. J. (2007). Dissenting statement, In the matter of Google/DoubleClick (FTC File No. 071-0170).

Herman, E. S., & Chomsky, N. (1988). Manufacturing consent: The political economy of the mass media. Pantheon.

Hermann, E., Morgan, M., & Shanahan, J. (2021). Television, continuity, and change: A meta-analysis of five decades of cultivation research. Journal of Communication, 71(4), 515–544.

Holbert, R. L., Garrett, R. K., & Gleason, L. S. (2010). A new era of minimal effects? A response to Bennett and Iyengar. Journal of Communication, 60(1), 15–34.

Hölbling, L., Maier, M., & Feuerriegel, S. (2025). A meta-analysis of the persuasive power of large language models. Scientific Reports. https://doi.org/10.1038/s41598-025-30783-y

Hovland, C. I. (1959). Reconciling conflicting results derived from experimental and survey studies of attitude change. American Psychologist, 14(1), 8–17.

Information Commissioner’s Office. (2018). Investigation into the use of data analytics in political campaigns. ICO.

Interactive Advertising Bureau / PwC. (2025). Internet advertising revenue report: Full-year 2024 results. IAB.

Iyengar, S. (1991). Is anyone responsible? How television frames political issues. University of Chicago Press.

Jungherr, A., & Rauchfleisch, A. (2026). The hidden costs of AI-mediated political outreach: Persuasion and AI penalties in the US and UK. arXiv:2603.27413. [Preprint.]

Kalla, J. L., & Broockman, D. E. (2018). The minimal persuasive effects of campaign contact in general elections: Evidence from 49 field experiments. American Political Science Review, 112(1), 148–166.

Klapper, J. T. (1960). The effects of mass communication. Free Press.

Kosinski, M., Stillwell, D., & Graepel, T. (2013). Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences, 110(15), 5802–5805.

Lazarsfeld, P. F., Berelson, B., & Gaudet, H. (1948). The people’s choice: How the voter makes up his mind in a presidential campaign (2nd ed.). Columbia University Press.

Lin, H., Czarnek, G., Lewis, B., White, J. P., Berinsky, A. J., Costello, T., Pennycook, G., & Rand, D. G. (2025). Persuading voters using human–artificial intelligence dialogues. Nature, 648, 394–401.

Luo, Y., Burley, H., Moe, A., & Sui, M. (2019). A meta-analysis of news media’s public agenda-setting effects, 1972–2015. Journalism & Mass Communication Quarterly, 96(1), 150–172.

Lippmann, W. (1922). Public opinion. Harcourt, Brace.

Matthes, J., Knoll, J., & von Sikorski, C. (2018). The “spiral of silence” revisited: A meta-analysis. Communication Research, 45(1), 3–33.

Matz, S. C., Kosinski, M., Nave, G., & Stillwell, D. J. (2017). Psychological targeting as an effective approach to digital mass persuasion. Proceedings of the National Academy of Sciences, 114(48), 12714–12719.

Matz, S. C., Teeny, J. D., Vaid, S. S., Peters, H., Harari, G. M., & Nave, G. (2024). The potential of generative AI for personalized persuasion at scale. Scientific Reports, 14, 4229.

McCombs, M. E., & Shaw, D. L. (1972). The agenda-setting function of mass media. Public Opinion Quarterly, 36(2), 176–187.

McGuire, W. J. (1964). Inducing resistance to persuasion: Some contemporary approaches. Advances in Experimental Social Psychology, 1, 191–229.

Nickerson, D. W., & Rogers, T. (2014). Political campaigns and big data. Journal of Economic Perspectives, 28(2), 51–74.

Perla, R., et al. (2026). The (in)effectiveness of psychological targeting: A meta-analytic review. Psychology & Marketing. https://doi.org/10.1002/mar.70073

Pooley, J., & Socolow, M. J. (2013). Checking up on The Invasion from Mars: Hadley Cantril, Paul F. Lazarsfeld, and the making of a misremembered classic. International Journal of Communication, 7, 1920–1948.

Roozenbeek, J., van der Linden, S., Goldberg, B., Rathje, S., & Lewandowsky, S. (2022). Psychological inoculation improves resilience against misinformation on social media. Science Advances, 8(34), eabo6254.

Salvi, F., Horta Ribeiro, M., Gallotti, R., & West, R. (2025). On the conversational persuasiveness of GPT-4. Nature Human Behaviour, 9, 1645–1653.

Sehgal, N. K. R., Rai, S., Tonneau, M., Agarwal, A. K., Cappella, J., Kornides, M., Ungar, L., Buttenheim, A., & Guntuku, S. C. (2026). Large language model chatbot conversations vs public health materials and parental HPV vaccination intentions: A randomized clinical trial. JAMA Network Open. [Published version of the trial circulated in 2025 as arXiv:2504.20519.]

Simon, H. A. (1971). Designing organizations for an information-rich world. In M. Greenberger (Ed.), Computers, communications, and the public interest (pp. 37–72). Johns Hopkins Press.

United States v. Google LLC, No. 1:23-cv-00108 (E.D. Va. 2025).

Vaccari, C., & Chadwick, A. (2020). Deepfakes and disinformation: Exploring the impact of synthetic political video on deception, uncertainty, and trust in news. Social Media + Society, 6(1).

Wu, T. (2016). The attention merchants: The epic scramble to get inside our heads. Knopf.

Zuboff, S. (2019). The age of surveillance capitalism: The fight for a human future at the new frontier of power. PublicAffairs.

Revised July 2026 in response to the three-referee external panel; the decision memo and response are filed under reviews/005. The 13-slide companion deck — The Broadcast and the Model — sits beside this paper; the corpus index, claim outline, framework, and prior manuscript versions are filed under cowork. Comments welcome at contactme​@​marshallcahill.com.