---

The Gate and the Field

Unchosen exposure to machines that argue — a census, a module, and five designs for the measurement nobody has made. A Specification Paper in Armchair Scholar, the independent paper series in which this is No. 012. Rules in force: no capability inflation (no claim exceeds the sizes the measured record supports), no null laundering (a failure to find is never read as a finding of absence), no boundary smuggling (chosen and unchosen exposure are never conflated). The paper's machinery is numbered so it can be held to: six designs (D1–D6, §7), five projections (P1–P5, §9), four registers of claims and six tripwires (§10), and three pre-registered protocols (A–C, §12). Draft v4 (archive panel revision, September 5, 2026); the module, dossiers (v2), and the three protocols ship with the paper.

· · ·

Some kinds of communication on some kinds of issues, brought to the attention of some kinds of people under some kinds of conditions, have some kinds of effects.

— Bernard Berelson, "Communications and Public Opinion," in W. Schramm (ed.), Communications in Modern Society (1948) (verbatim — confirmed against the chapter text at imprint)

Abstract

The measured record on machines that argue — assembled in The Broadcast and the Model (this archive, 2026) from the assigned-exposure experiments it synthesizes — will bear this much: adaptive, multi-turn dialogue moves belief and stated preference by single digits on hundred-point scales, with roughly a third to two-fifths of the shift surviving re-contact — measured in panels, under assigned exposure, on people who chose the conversation. That paper closed on the sentence everything politically consequential depends on: whether that lever reaches people who did not choose the conversation — the selection gate — remains unmeasured. This paper specifies the instruments that would measure it. A verified census structure shows the gate already being dismantled at scale, but selectively and conditionally: AI search summaries — reported by their maker at 2.5 billion monthly users, a company-stated figure this paper grades as such — are a default within the chosen act of searching — present unless removed, firing on a minority of queries; a messaging assistant whose presence cannot be fully disabled reaches a billion-plus by its maker's last count and on the order of 1.5 billion by trackers, its presence unavoidable and its conversations chosen; and sixty percent of American adults told a national survey this year that they have ever read the machine-written summaries now placed atop their search results — content unrequested, encounter conditional on a chosen act. The unchosen reach is wide and shallow; the dialogue depth where the measured lever lives remains, for now, mostly chosen; and the field question is therefore the depth-crossing rate — how often unchosen encounter becomes unchosen conversation, and what a contact does when it arrives uninvited. Against that question the paper specifies: an exposure census on a two-coordinate gate grammar (how the surface is encountered; how depth escalates), each coordinate admitting conditional values; a survey module with per-surface depth and gate items, a published timing test, and a data-donation validation path stated with its selection limits; and five designs — the census run; default-flip natural experiments, including the replacement of the Android default voice assistant — available on a billion devices by its maker's January 2019 count, a company-stated availability figure with no current successor — by a conversational model scheduled for September 4, 2026, whose estimand is restated honestly as a treatment change on invokers plus whatever changelog-audited proactive behavior introduces genuinely unchosen delivery; a flip panel carrying the assigned-exposure outcome batteries with never-chooser status defined at baseline — the instrument the paper's central comparison actually requires; encouragement designs whose complier-only estimand and compliance-diluted sample sizes are printed at their true magnitudes; and a platform-cooperative ceiling design left ready on the shelf — and, beside the five, an aggregate ambient arm sketched on the one template that has measured ambient machine influence at population scale, so that the event-like and ambient readings finally predict different things on the same census. One scoped section asks whether the arguing channel is also a conferring one, and equips the question with the discriminant it needs: machine praise against matched non-agentic boilerplate, so that the bridge can fail even while returning readings. The paper states no field-effect number, because none exists and manufacturing one is the vice it indicts; it concedes in its non-claims that a null across every design bounds only the event-like channel, which the ambient arm indexes without measuring; and it had one deadline, printed with its own limitation: the September flip's usable baseline was the panel's first wave, it had to field in August, and it did not — the missed before-picture is logged as missed, the projection that needed it is re-indexed to the next named flip under a sunset rule, and what the vendor's multi-week rollout still allows is kept at its degraded grade (§7, §11).

§ IThe Debt, the Gate, and the Product

The Broadcast and the Model ended where honest measurement ended. Its verdict — real persuasion, modest sizes, fact-density as the lever, an accuracy tell attached — was taken entirely under chosen exposure. And it named its boundary rather than extrapolating across it — in its dek, quoted exactly: "What remains unmeasured is the field itself: whether the lever reaches people who did not choose the conversation — the selection gate, named precisely enough to be tested."

This paper is that test's specification, and its organizing observation is arithmetic. Every consequential claim about conversational persuasion — alarmed or reassuring — is silently a product of three factors: reach among the unchosen × per-contact effect net of selection × durability — a decomposition, not an identity: it carries an unstated dose term (contacts per person) and assumes the per-contact effect is the same among the unchosen as among the chosen, which is exactly what the third projection (P3, §9) exists to test. The record holds careful measurements of the second and third factors among the chosen, no measurement of the first, and no measurement of the product. Until the gate is measured, both the panic and the reassurance are extrapolations. After it, they are bounded — and the product itself is computable only by the one design that assigns exposure at the surface (D4, §7), which is why measuring the gate leaves the product unmeasured and the paper says so.

The load-bearing assumption, stated so it can be doubted — and bounded, not laundered. Everything below assumes persuasion exposure is event-like: encounterable, recallable in a survey week, assignable by a flip. A serious rival holds that machine influence is ambient — climate rather than weather, in the cultivation tradition's sense (Hermann, Morgan and Shanahan 2021), a slow recalibration of what answers feel normal, accumulated across thousands of shallow contacts no respondent could enumerate — and the ambient view predicts its largest effects in exactly the census's highest-reach cell, through a channel every event-shaped instrument is blind to. The paper does not pretend otherwise: none of the five designs below measures the ambient channel, a null across all of them bounds the event-like channel only, and the non-claims of §10 say so in binding type. But the ambient channel is not unmeasurable. It has been measured once, at population scale, by an instrument that needs no respondent: word-level synthetic control on 737,083 hours of unscripted podcast speech found model-preferred vocabulary entering spontaneous human speech after ChatGPT's release, with a registered experiment (N = 496) confirming the mechanism (Yakura et al. 2024/2026). That template gives the ambient question an aggregate arm — D6 in §7, a sketch with an index, not a persuasion measurement — and, more important, it lets the two readings finally predict different things on the same census (P5, §9): under the event-like reading the flip moves invokers at dialogue depth and the ambient index stays on its pre-trend; under the ambient reading the index moves at the flip while the dialogue cell stays in single digits. The individual-level direction — the donation architecture of §6 extended longitudinally, cumulative shallow-contact dose logged against meta-epistemic outcomes (answer-trust, epistemic deference) over months, not a week — stays beside it in §7 with its insufficiencies stated. And a third channel, neither event-like nor ambient, is now a finding rather than a conjecture: argumentative content assigned inside a chosen, non-argumentative act — an AI writing assistant's biased suggestions shifted 2,582 participants' stated positions on societal issues, most unaware, with forewarning and debriefing not attenuating the shift (Williams-Ceci et al., Science Advances 2026). No respondent could enumerate that exposure; a recall instrument is blind to it by construction, and §10's non-claims say so.

§ IITwo Caricatures

Both are constructed ideal types — poles of the argument space, not quotations of named authors. The Multiplier takes the measured single digits, multiplies by billions of monthly users, and announces a persuasion machine of civilizational scale. Its arithmetic error is this paper's subject: the lab effect was measured at dialogue depth on the chosen, and the billions are mostly shallow, conditional, and — on the anchor surface — triggered by a minority of queries (§5); the Multiplier multiplies numbers from opposite sides of the gate and calls the product measured. The Selection Comfort runs the fallacy in reverse: only seekers are touched, so the field is safe by construction — assuming the gate's value instead of measuring it, in the decade when defaults are dismantling the gate at planetary scale. Both share one method: assigning the gate a value without an instrument. The paper's position is that the gate has no value until measured — and it states the bar once, so the claim can be held to it: a measurement of the gate is population-scale, on a belief or preference outcome, on a population not defined by having chosen the surface, under randomized or quasi-experimental identification. By that bar nothing has measured it (Protocol B, §12), and by the same bar only two of the five designs of §7 could: the flip panel (D5) against a named flip, and the platform-cooperative design (D4) if a platform volunteers. The census run (D1) measures the gate's prevalence with no outcome attached; the September flip design (D2) identifies invokers, not never-choosers; the encouragement design (D3) identifies compliers. So the honest sentence is not now measurable but now specifiable — and measurable through D5 against a named flip, which is what §7 and §9 price.

§ IIIThe Measured Record, Compressed

From the record, at full strength and binding here (rules 1 and 2) — the assigned-exposure findings The Broadcast and the Model assembled, cited at their primary sources in the corpus index: single-digit movement on hundred-point scales, between a third and two-fifths surviving re-contact (36 to 42 percent of the immediate effect at one month in the largest of these studies — Hackenburg et al., Science 2025; and, in the election-dialogue study's own follow-up, slightly more than a third of the preference shift persisting about six weeks later — Lin et al. 2025); fact-density as the lever, not personalization; persuasiveness rising as accuracy falls; the protective nulls standing — the microtargeting nulls (Hackenburg and Margetts, PNAS 2024), the ~5 percent psychographic ceiling, machine–human single-message parity, diminishing returns to scale — with the persuasion shelf The Broadcast and the Model compresses running wider than one Science paper (Salvi et al. 2025 on LLM debate persuasion; Matz et al. 2024 on personality-matched messages; Goldstein et al. 2024 on machine-written propaganda), all synthesized there. Nothing in this paper exceeds these sizes anywhere, in any mood (rule 1). One flag attaches to the record itself, and rule 2 requires it printed — correctly, since an earlier draft of this paper pinned it to the wrong study: the conspiracy-belief dialogue experiment (Costello, Pennycook and Rand, Science 2024), the record's strongest durability datum with its two-month hold, has been under an Editorial Expression of Concern at Science since June 11, 2026 — screening-criterion inconsistencies and a dataset merge error, with the authors reporting a corrected pipeline that preserves direction, significance, and size, and the matter unresolved at this draft. The durability figures this paper carries come from different studies and are unaffected; The Broadcast and the Model already prints the notice beside the contested finding. The paper adds no effect claims. It adds the missing factor's instruments.

§ IVThe Gate, Formalized — Two Coordinates, Conditional Values

A single label per surface turned out to be too coarse for the two largest surfaces on Earth, so the grammar carries two coordinates. Gate-of-encounter: how the surface's content first reaches a person — chosen (sought), default (present unless removed), unavoidable (cannot be removed), delegated (an agent the user chose encounters it on their behalf — the machine-to-machine layer, agents transacting with agents, that The Spectator and the Agent (this archive, 2026) records; a fourth type, named as one, not smuggled as an exception) — with conditional values admissible and required: an AI search summary is a default within the chosen act of searching — present unless removed, and removable (a browser extension removed it for 1,065 participants, §7), so it is not unavoidable under this grammar's own definition — and it fires on a minority of queries, so the census carries its activation rate beside the grade; a messaging assistant's presence is unavoidable while its conversations are chosen. Gate-of-escalation: how encounter deepens — the depth grammar (summary → answer → dialogue → agent) crossed with whether each escalation was sought or offered, because the modern product escalates in one tap ("ask a follow-up"), and a dialogue reached through an unchosen summary's embedded prompt is the paper's core quantity living in what a one-coordinate scheme could not code. The seam is codable now, and the census commits to proving it before it fields: the codability test of the appendix — two independent coders, a pilot of free-text encounter descriptions with a quota of seam cases, an agreement floor set on the joint cell rather than on each coordinate, and a donation-concordance criterion — with results published with the census run.

Two facts still organize everything, one of them now stated with its domain limit. Defaults move behavior at large effect sizes where outcomes are passive — d ≈ 0.68 across 58 studies (Jachimowicz, Duncan, Weber and Johnson, Behavioural Public Policy 3(2), 2019, CI [0.53, 0.83]) — which motivates treating default placement as exposure assignment; whether that transfers to invocation-gated conversational surfaces is untested, and this paper treats the transfer as a question, not a premise. And the structural irony stands: The Broadcast and the Model showed the broadcast century's reassuring science was bounded by symmetric, unchosen exposure; the default AI surface is rebuilding unchosen mass exposure — conditionally, tap by tap — with an arguing machine inside it. Berelson's sentence was a map of scope conditions; the industry has been redrawing the map's largest border for two years, largely unmeasured.

The construct has ancestors, named so this paper extends rather than reinvents. Selection's power is the field's oldest result (Klapper 1960; the two-step flow of Katz and Lazarsfeld); Prior showed what choice expanding did — selection came to dominate effects (Post-Broadcast Democracy, 2007) — and this paper is Prior inverted: choice contracting by default. The oldest statement of the question sits a few pages past the epigraph (p. 210 of the 2007 reprint, against the epigraph's p. 204 — two passages, both verified): Berelson's same chapter records slight evidence that non-purposive — his word, accidental — exposure converts opinion more readily than deliberate exposure, because the defenses are down and the material is less congenial; a 1948 conjecture that is, nearly verbatim, the reading §9 registers as the one this paper most fears. The platform era named the residue incidental exposure (Fletcher and Nielsen 2018) and then complicated it: Thorson (2020) reframed incidental exposure as algorithmically attracted rather than accidental — the module's own "was it just there?" question, asked of feeds six years before this paper asks it of arguing surfaces — with Weeks and Lane (2020) on its ecology and Schäfer (2023) as the scoping review; the census below is that literature extended from feeds to arguing surfaces. The specific cell "default within chosen search" has direct experimental precedent: the search-engine manipulation effect (Epstein and Robertson, PNAS 2015 — rank bias shifting undecided voters' stated preferences by twenty percent and more, maskable past awareness — an effect, note, at zero dialogue depth, which is why the depth gradient of §5 is a claim about where the measured lever lives, not a claim that shallow contact is inert) and the answer-bot study that followed (Epstein, Lee, Mohr and Zankich, PLOS ONE 2022 — a single answer-box exchange moving stated preferences by large margins) — carried with the caveat printed: that program is substantially one laboratory's, and large independent replications are the missing tier. A different cell has precedent of its own: argumentative content delivered inside a chosen, non-argumentative act — opinionated autocomplete moving writers' own views (Jakesch et al., CHI 2023) and, at N = 2,582, biased writing-assistant suggestions shifting stated positions with most participants unaware (Williams-Ceci et al. 2026) — the closest published miss to the field datum this paper wants (Protocol B, §12), because the exposure is unsought at the item level.

§ VThe Census

Every surface carries both gate coordinates, its depth grade, its reach figure at stated grade, and its flip dates. Verified this round, with grades corrected: Google's AI Overviews — 2.5 billion "monthly active users," company-stated (Pichai, I/O 2026), with no published definition of active for a surface that renders without being asked — encounter default-within-search, conditional on activation; escalation one-tap-chosen; rolled out in five dated waves (May 2024 → 200+ countries by May 2025). The default cell is conditional on activation, and the activation figure cuts two ways: an independent audit of 55,393 trending queries (March–April 2026), in the algorithm-audit tradition Metaxa et al. (2021) codified, found the summary fires on 13.7 percent of queries overall, 64.7 percent of question-form queries, and markedly less on politically sensitive topics (Xu, Iqbal and Montgomery, 2026) — a query-distribution figure, not a user one, and the reason the census carries an activation rate beside every reach figure. It does not make the cell thin at user level: at 13.7 percent per query, a person running five searches a week meets at least one summary in 52 percent of weeks, ten searches in 77 percent, twenty-five in 97.5 percent — query-level rarity is compatible with near-universal weekly encounter, which is the module's question, not the audit's. The larger cross-national audit (Aral, Li and Zuo, 2026 — 24,000 queries across 243 countries; AI Overviews expanding from 7 countries in 2024 to 229 in 2025, with France, Turkey, China and Cuba receiving none; AI answers drawing on fewer, less credible, and more right- and center-leaning sources than the traditional results) supplies D2's country-by-date assignment map, and its political-lean finding is a second reading of "less on politically sensitive topics" — fewer summaries on those queries, differently sourced when they fire — which the census carries rather than reconciles. AI Mode — a billion-plus monthly active users per I/O 2026, company-stated — chosen; the Gemini app, a distinct product the census keeps in its own row, at 950 million monthly by July 2026 (Alphabet's second-quarter statement; an earlier draft's "billion" was ahead of the record). The Gemini-for-Assistant substitution — surface default, invocation chosen; reach for the assistant surface itself has no current company figure, the last being Google's January 2019 statement that Assistant would be available on one billion devices by that month's end (company-stated, availability not use, seven years old); product-level irreversible from September 4, 2026, rolling over several weeks and excluding home and television devices (the date is a vendor schedule; as of this draft no postponement had been announced and no dated observation of removal reaching devices had been pinned — the flip log's first entry carries that pin; see the deadline's re-indexing clause, §10). Meta AI — a billion-plus monthly by the company's last stated count (May 2025), with nothing fresher disclosed through its mid-2026 earnings calls; ~1.5 billion monthly at tracker grade, WhatsApp its leading surface by the company's own account and ~63 percent of use by trackers — presence unavoidable, delivery chosen-conditional, forced into 41 European countries in March 2025. Destination chatbots — chosen, 49 percent of U.S. adults, 24 percent daily (Pew, February 2026, N = 5,119). The agentic layer — delegated by construction, reach without any public figure: the census's named empty cell, carried as a gap.

The opening finding, restated on the corrected grammar and scoped to the channels a respondent can enumerate: the widest reach is conditional-unchosen and shallow; the dialogue depth where the measured lever lives remains mostly chosen. Sixty percent of American adults say they read the machine-written summaries atop their search results — Pew's item, verbatim, asks whether they "ever read AI summaries at the top of a search engine result" (60 do, 30 do not, 10 not sure; fielded February 17–23, 2026) — content unrequested, encounter conditional, and an ever-measure of reading rather than a rate of encounter. The seven-day encounter rate is not a gap this module fills: the Reuters Institute already fields it — 54 percent across six countries had seen an AI-generated answer to a search query in the past week in 2025 (Argentina 70, the United Kingdom 64, the United States 61, France 29 — France then untreated by AI Overviews), and a year later 10 percent of respondents in 45 markets used an AI chatbot for news weekly, 42 percent of those users asking a follow-up question and 4 percent regularly clicking through to a source (Simon, Nielsen and Fletcher 2025; Ross Arguedas 2026). What the module adds is the per-surface gate and escalation items — was it sought or just there, and did the back-and-forth grow out of something that was just there — which no fielded instrument asks. The field question is the depth-crossing rate — and the existence claim is now scoped precisely: randomized evidence that answer surfaces persuade exists (the Epstein program; Xu, Dash, Kang, Liao and Spiro's preregistered RCT on AI search summaries, 2025), in laboratories and online panels, on consenting participants. What has never been measured is the field: population-scale effects on people who never chose any of it. That narrower sentence survives the literature, and it is the one this paper builds against.

§ VIThe Module

The module is a screener plus a per-surface loop, and its honesty constraints are stated with it — including its ancestry: Pew's February 2026 battery and the Reuters Institute's cross-country generative-AI items — the six-country series of Fletcher and Nielsen (2024) and Simon, Nielsen and Fletcher (2025), and the chatbot-news items of the 2026 Digital News Report — are the named instruments it extends, and it claims no firstness: the seven-day window is theirs; the gate and escalation items are this paper's. The screener establishes encounter for six surface families (search summaries; destination chatbots; voice assistants; messaging-app AI; AI answers in feeds; agents that acted for you) over a uniform seven-day window. The loop, for each surface encountered: sought it, or was it just there? (gate-of-encounter); did it go back and forth — you replied and it answered again? (depth, per surface — the joint cell is constructible by design now); for any back-and-forth: did you start that exchange, or did it grow out of something that was just there? (gate-of-escalation — the seam item); domain contact split by domain (news; politics; health; a live purchase — four items, not one); frequency ladders; the gate-awareness item (were the AI features on by default? — whose "don't know" is a named mixture of unnoticed defaults, construct confusion, and absent features, analyzed against the donation subsample's device audit, not read naively); and the paid data-donation consent hook (methodology: Boeschoten et al.; Ohme and Araujo — the pin the dossier owed, closed). Response options are harmonized across items. The five-minute claim is now a commitment, not an assertion: a published timing test on a heavy-exposure profile precedes any fielding claim, and if the loop breaks five minutes, the runnable claim is struck by the paper's own tripwire — pre-registered here.

The donation subsample's limits, printed: consenters are privacy-tolerant and likely heavier users — a bias that runs toward P1's falsifier, and is registered there as a caveat rather than left here as a remark; the recall-versus-log bound applies to consenters and is exported to the full sample only under a selection model the dossier states before fielding — post-stratification of consenters to the full sample on the module's own frequency ladders, surface count, and device class, with the assumption printed that recall error is independent of consent within those strata; the subsample also carries a pre-registered concordance criterion of its own — self-reported just there must agree with logged default status on at least eight of ten audited encounters, or the gate-of-encounter item is redesigned before the census run claims to measure the gate (Protocol C, §12); and the device audit logs surfaces and default status, not content, so it is as blind as the module to exposure a respondent cannot enumerate (§4) — the bound it supplies is a bound on recall of enumerable encounters, nothing more; and at n ≈ 500 across six surfaces, the smallest cells — unchosen dialogue above all — will carry wide error, which the census run reports rather than hides. Cell-level precision for the smallest analysis cell is stated in the D1 dossier alongside the marginal ±1.4-point figure that applies only to topline prevalence.

§ VIIThe Designs — Five Families

D1 — the census run. National probability sample near five thousand, module in an existing omnibus, donation subsample for the recall bound with the §6 limits. Reach by surface × both gate coordinates × depth. The program's cheapest component; decidable this year.

D2 — the default-flip designs, restated honestly. The design family's ancestors are the media-rollout studies — DellaVigna and Kaplan's Fox News effect (QJE 2007); Martin and Yurukoglu's channel-position instrument (AER 2017); closest in structure to a default placement, the staggered college-by-college rollout of Facebook (Braghieri, Levy and Makarin, AER 2022) and the availability-as-assignment designs of Enikolopov, Petrova and Zhuravskaya (AER 2011) and Durante, Pinotti and Tesei (AER 2019) — and its newest members teach two lessons the dossier absorbs: platform algorithm flips measurably move political attitudes (Gauthier, Hodler, Widmer and Zhuravskaya, Nature 2026), with effects that can be asymmetric between on and off — a third reading for any September null — and client-side reranking has made some field experiments possible with no platform invitation at all (Piccardi et al., Science 2025), a rung between D3 and D4 the dossier records as D4′. Three dated instruments. The retrospective two — the AI Overviews five-wave rollout and the March 2025 forced Meta AI rollout — are conditional on series discovery: they are usable only where belief or preference series already run at matching cadence and coverage, the paper does not yet name a qualifying series — the appendix enumerates and adjudicates the first candidate on the record (the Reuters Institute's cross-national surveys against the AI Overviews country rollout, with France untreated until July 22, 2026) — and the discovery protocol is written into the appendix so that a series found later is labeled discovered, not registered. The prospective one — September 4 — is restated per its actual structure: the flip replaces an existing assistant, so the counterfactual is dumb-assistant, not no-assistant, and the surface is invocation-gated, so the estimand is a treatment change on invokers: the first stage is the change in dialogue-depth exposure conditional on invocation, plus any change in invocation rate, plus genuinely unchosen delivery only where changelog-audited proactive behaviors introduce it — an audit with its own emerging literature on what models volunteer unprompted (Poungpeth et al., 2026). The comparison structure is named for what it is: with one treated platform and one donor there is no donor pool, so "synthetic control" would be a label — the estimator is a two-group difference-in-differences against iOS users, resting on a parallel-trends assumption the dossier states, with the platform-selection confounds (income, age, country mix) and the shared information environment (the same chatbot apps on both platforms) named as the assumption's known violations; failing that, an interrupted time series, the design family's weakest member, called by its name. The onset is smeared — the vendor's removal reaches devices over several weeks — so the dossier declares an event window (announced start to announced completion, with the module's migration-status item dating each respondent's switch where a panel exists) rather than a single day. And the changelog audit is a hope with a channel, not an estimand: no changelog access is secured, proactive is operationalized in the dossier as delivery not preceded by an invocation in the same session, and the vendor can ship server-side changes no changelog records — the audit literature cited (Poungpeth et al., 2026) audits model behavior in conversations, the nearest instrument and not the same object. There is no contemporaneous untreated Android unit, and the dossier says so.

D5 — the flip panel (the instrument the central comparison requires). A probability panel fielding the module and the assigned-exposure outcome batteries, never-chooser status defined at baseline from the module's own gate items — a status on the enumerable channels only, so a never-chooser so defined may carry exposure no item can elicit, and P3's comparison is a comparison on that scope — minimum two pre-flip waves and two post-flip waves — and, because the paper's own minimum binds it, D5 is re-indexed: for the September 4 substitution the two pre-flip waves are permanently unsatisfiable (no wave was fielded before the announced start; §11), so D5 stands specified against the next named wave of the same substitution — the vendor has said that Assistant's removal on Home and Nest speakers, smart displays, and Google TV follows the phone rollout, and that cars with Google built-in keep Assistant beyond September 4 — with its two pre-waves fielded before that wave's announced date, and a sunset rule in tripwire type: if the flip log's first quarterly entry (§11) names no flip with a fieldable pre-period, P3 is withdrawn from Register P — the paper's list of projections, §10 — by rule, and the register says so. D5 is what turns a flip from a first-stage curiosity into the first individual-level field measurement on identified never-choosers — the only instrument in this paper that can deliver P3's comparison. The deadline of §11 belonged to D5's first wave, and the deadline sentence carried its own limitation: one August wave would have been a baseline for levels, not trends; two waves were the honest minimum; the calendar permitted neither. What the September rollout still allows — a wave with a self-reported migration-status item on devices not yet switched — is a levels baseline for D2's first stage, printed as the degraded instrument it is (§11), and it identifies nothing for P3.

D3 — the encouragement designs, at their true price. The design of Gerber, Karlan and Bergan's newspaper experiment (AEJ: Applied 2009), with the complier estimand of Angrist, Imbens and Rubin (1996): randomized encouragement into a dialogue surface, four weeks, the same outcome batteries, the durability re-contact. The estimand is the per-contact effect for compliers, printed beside every estimate. And the price is printed at its real magnitude: powering the intention-to-treat against a complier effect requires scaling by the inverse square of the compliance rate. And the compliance assumption is now bounded rather than picked: Pew's own non-user chapter documents the reluctance of the half who don't use chatbots, so the dossier carries π ∈ [0.2, 0.5], and the price is printed with its inputs so a reader can reproduce it: a complier effect of d = 0.15 (the record's ceiling; 0.10 at its center) on the hundred-point batteries of The Broadcast and the Model (the dossier prints the working standard-deviation basis it uses to turn d into points, to be replaced by the pilot's observed value), two-sided α = .05 at 80 percent power, intention-to-treat dilution by 1/π², and 20 percent attrition — at π = 0.4 and d = 0.15, on the order of 5,500 per arm; at π = 0.2, 22,000; at the record's central d = 0.10 the same two points read 12,300 and 49,000. The span across the paper's two stated uncertainties is 3,500 to 49,000 per arm (the dossier prints the full table). The pilot that fixes π is not optional. That is consortium scale. The paper prefers that table to a cheaper wrong number — and it adds the one repair the encouragement literature offers for the complier label: complier profiling (Marbach and Hangartner 2020), so the population D3 actually identifies is described, not merely named.

D4 — the ceiling design, on the shelf. Platform-cooperative randomization of presence or depth on live queries, under the US 2020 governance template (Allcott et al. 2024; Guess et al. 2023; Nyhan et al. 2023) — external academic control, pre-registration, no platform veto — carried with its published critique: Wagner's "Independence by permission" (Science 2023) argued that the platform's permission itself bounded that study's science, and the dossier prints the objection beside the template rather than beneath it. And "only if a platform volunteers" is no longer the whole truth: under the Digital Services Act's Article 40, operationalized by Delegated Regulation (EU) 2025/2050, vetted researchers hold a statutory path to platform data for systemic-risk questions. Experimentation still needs a volunteer; the observational half no longer waits on goodwill. The dossier splits accordingly — D4-experimental (volunteer-gated) and D4-observational (statute-gated) — with D4′, client-side intervention, standing between them, needing no invitation at all. D4′ is no longer hypothetical on the census's anchor surface: a browser extension randomly removing AI Overviews from 1,065 consenting participants' searches (two weeks, early 2026) cut outbound organic clicks by about forty percent and raised zero-click searches by a third, with no measurable change in reported search quality (Agarwal and Sen, 2026) — the experimental counterpart of Pew's July 2025 observational finding that users click links less often when a summary appears. The removal instrument exists; what that study measured was traffic and satisfaction, not belief, and its population was consenting installers — so D4′ on this surface is the same experiment with the assigned-exposure outcome batteries and the durability re-contact bolted on, and its population label printed.

D6 — the ambient arm, a sketch with an index. Two directions, neither a persuasion measurement. Individual: a longitudinal donation panel — months, not a week — logging cumulative shallow-contact dose against meta-epistemic outcomes (answer-trust, epistemic deference); not yet an instrument, recorded so the gap has an address. Aggregate, on the Yakura template: build a lexicon of the substituted assistant's preferred words and framings from its own logged responses before the flip; track those treated terms against a donor pool of untreated terms with matched pre-trends in public corpora of spontaneous speech and writing (podcast transcripts; forum text) across the flip's event window, with word-level synthetic control as Yakura et al. ran it on 737,083 hours of podcast speech around ChatGPT's release; the post-flip excess of treated terms is the ambient index — an uptake measure, not a belief effect, priced at computation plus public corpora, runnable this year. Its failure modes are printed with it: the lexicon overlaps other models' (the flip is one model among several, so attribution needs terms specific to the substituted assistant); the speaking population is not the invoking population; a null on vocabulary does not bound a shift in what answers feel normal. What D6 buys is not the ambient effect but a discriminating prediction (P5, §9): the event-like and ambient readings, which agree about everything the census returns, disagree about whether the index moves at the flip.

§ VIIIThe Bridge, in One Section — Now With Its Discriminant

A channel that argues back also responds — praises, corrects, dismisses, counts. Whether that is conferral — being counted by others as worth something — in the sense The Wage and the Worth and The Praise and the Praiseworthiness spent two papers learning to measure is exactly what this section refuses to assume, because the strongest objection is now in the text: conferral, as the latter paper measures it, is a social fact, a community's counting, and a machine's praise may be pseudo-conferral — a simulacrum eliciting the phenomenology of being counted with no community behind it, no one whose counting it is. So the bridge's test carries a discriminant: machine conferral registers only if that paper's battery distinguishes agentic praise from matched non-agentic boilerplate — identical words, no purported judge. Differential response is the finding; equivalence sends the bridge to a footnote even though the battery returned readings, because readings alone confirm nothing but anthropomorphism — and equivalence is a claim, not a failure to reject, so it carries a bound: the bridge files to its footnote only when a two-one-sided test rejects a differential larger than one-third of the smallest conferral effect The Praise and the Praiseworthiness reports on the same battery, at the sample that bound implies (printed in the dossier); a differential neither detected nor excluded leaves P4 undecided, and the register says so. The expectation is symmetric — the paper genuinely does not know — and the channel, if it registers, is scoped as the first responsive, individualized conferral source manufactured at scale. This section builds no instruments and claims no effects; the one magnitude it states is the equivalence bound, which it must. Cross-listed, in the archive's own organization, to the two threads those papers belong to; one section; nothing further.

§ IXThe Projections

Five, each about what the instruments would return, all under the record's ceiling. P1: the census returns a depth gradient — unchosen-encounter reach wide at summary–answer depths, with unchosen dialogue reach in single digits as a percent of adults; falsified by donation-validated double-digit unchosen dialogue reach, which would mean the regime change outruns this paper — with the falsifier's precision printed: at the donation subsample's n ≈ 500 the estimate carries roughly ±2 to ±2.6 points at the prevalences in question, so a reading between 8 and 12 percent is undecided, not decided either way; and its known bias printed: consenters are privacy-tolerant heavier users, which pushes the donation-validated figure toward the falsifier, so P1 fails only on a double-digit reading that survives the stated selection model, not on the raw consenter figure. P2: the September flip produces a measurable first stage in its restated estimand — invocation-conditional depth change, invocation-rate change, audited proactive delivery; a null now carries three readings, distinguished in advance: the defaults-motivated expectation failed; the surface was misclassified as a flip; or the flip's effects are asymmetric, as the newest platform-flip evidence shows they can be (Gauthier et al. 2026) — separated by the invocation series and the module's switched/not-yet-switched contrast inside the event window, which the paper controls, and by the changelog audit, which it does not: without the audit the proactive clause is unmeasured, a null carries the first two readings only, and the paper says which. And what the flip adds is scoped against the existing record: randomized answer-surface persuasion exists in laboratories (the Epstein program; Xu et al. 2025); the flip's contribution is the population — invokers in the field, moved from a dumb assistant to an arguing one without choosing the change — not the phenomenon; never-choosers it reaches only through the audited proactive clause, and the paper counts none as reached until the audit shows delivery. P3 — decidable only through D5, re-indexed to the next named flip, stated in tripwire type: with the flip panel fielded before that flip's announced date, the comparison of complier effects (D3) against never-chooser effects (D5) brackets the selection effect the entire chosen-exposure literature has never seen. Expected: the deflating direction — compliers more moved, the panels overstating the field. Registered as the reading the paper most fears — and the one Berelson conjectured in 1948 (§4): never-choosers moved more, every reassuring number in the record a floor. Without D5, P3 is not a projection but a wish, and the paper says so — and acts on it: D5's two pre-waves are unsatisfiable for September 4, so P3 stands against the next named wave of the same substitution (§7), under the sunset tripwire T6 (§10). P4: the bridge's discriminant returns a differential — or an equivalence, within the bound §8 states, that files the bridge to its footnote by rule; a reading that neither detects nor excludes the differential leaves P4 undecided. Symmetric; no expectation stated. P5 — the discriminating projection (D6 against D2 and D5): the event-like reading this paper assumes predicts that movement at the flip, where any is found, sits with invokers at dialogue depth (D2's first stage; D5's comparison) and that the ambient index of D6 stays on its pre-trend through the flip's event window; the ambient reading predicts the index moves at the flip while the dialogue cell of the census stays in single digits. Expected: the event-like pattern — stated as an expectation the paper would rather lose honestly than hold by assumption. Falsified, for the §1 assumption itself, by an index that moves at the flip with the dialogue cell flat; that reading converts the ambient concession from a moat into a measured failure of the paper's premise, and it is registered here so the concession cannot be kept and banked at once.

§ XThe Ledger of Claims and Non-Claims

Register C. The record's verdict is bounded to chosen exposure — by the explicit statement of The Broadcast and the Model; the three-factor product with its unmeasured first factor; deployment preceding measurement, as with broadcast; default placement as motivated exposure assignment, the transfer to invocation-gated surfaces open; the two-coordinate gate with conditional values, and the depth-crossing rate as the field question; the ceiling/floor structure of platform cooperation — the floor anchored in the post-API record (Freelon 2018; the 2023 access shutdowns); the bridge's question equipped with a discriminant; the module as the cheapest decidable piece. Register E, at corrected grades: 2.5B and 1B monthly active users — company-stated, definitions unpublished; the 13.7 / 64.7 percent activation rates, query-distribution grade, with their user-week conversion (52 / 77 / 97.5 percent at 5 / 10 / 25 searches); Meta AI 1B+ company-stated (May 2025), ~1.5B tracker grade; the Gemini app at 950M, company-stated; the Assistant surface at one billion devices, company-stated availability as of January 2019 and no current figure; Pew February 2026 — 60 percent have ever read AI search summaries, 49 percent use chatbots, 24 percent daily, 42 percent use them to search for information; N = 5,119, item verbatim; the Reuters six-country seven-day AI-answer item (54 percent; France 29) and the 2026 chatbot-news items (10 / 42 / 4 percent); the AI Overviews country map with its four excluded countries and France's July 22, 2026 launch; the US 2020 designs and Ns; d = 0.68 with its passive-outcome domain stated; the measured record The Broadcast and the Model assembled, at its primary sources, with the Expression of Concern attached to the conspiracy-belief study it concerns; the AI Overviews removal experiment as D4′'s existence proof; the Epstein program and Xu et al. (2025) as the existing randomized evidence the field claim is scoped against; the DSA Article 40 delegated regulation; the AI Mode / Gemini-app distinction held; the agentic gap, named. Register D: five design families, each with estimand, population, comparison structure, and failure mode — including D5, re-indexed to the next named flip under a sunset rule, without which the paper's central comparison does not exist — and one sketch, D6, carried with an index and no effect claim. Register P: the five projections — P3 conditional on D5's re-indexing, P5 the discriminating one. Nothing else predicts anything.

Non-claims. No field-effect number. No capability inflation past the record's sizes, in any mood — including the first stage, where the defaults transfer is a question. No null laundering. No boundary smuggling. No platform villain narrative. The bridge builds and adjudicates nothing. And the concession honesty requires: a null across all five designs bounds the event-like channel only; the ambient channel is not measured by them, and D6 indexes it without measuring its effect on belief. A third channel — argumentative content inside chosen, non-argumentative acts (§4) — is measured in the laboratory and invisible to a recall instrument by construction; the module does not claim it.

Tripwires. T1, population honesty: complier labels travel with every D3 estimate; a design sold as the field while identifying compliers fails its section. T2, census completeness: both gate coordinates and a depth grade, or the surface is excluded and listed. T3, the module's constraint: the published timing test precedes the five-minute claim; failure strikes the runnable claim. T4, the ceiling: no forward claim past the record's sizes without a named design. T5, the deadline, re-indexed: the September sentence tracks the flip's current announced date — a vendor slip is logged in the flip log, never smoothed; the sentence may not be softened; and if the window closes, the paper says the window closed. T6, the register's sunset: P3 is withdrawn by rule if the flip log's first quarterly entry names no flip with a fieldable pre-period; P4 is undecided, never equivalent, without the §8 bound; the projection count in §9 and the Notes tracks the register.

§ XIThe Close

The Broadcast and the Model found the lever and refused to guess its reach. This paper refuses differently: it makes the guessing unnecessary — and it prices everything, including its own timing. The asks. To the survey houses: the module and the panel's first wave are specified; the September flip's usable baseline was the panel and its window was August (§7). What the printed deadline bought is not absolution but a rule: the projection that needed the baseline is re-indexed to the next named wave of the same substitution, with its pre-waves due before that wave's announced date and a sunset if none is named — so the next window is the instrument's to keep, not the archive's to confess. What the multi-week rollout still allows is a weaker baseline — a September wave with a self-reported migration-status item, establishing levels on devices not yet switched — printed as the degraded instrument it is, never as the one the paper asked for. To the platforms: the ceiling design is written, governed by your own precedent, and waiting. To the discipline: stop extrapolating across the gate in either direction; the gate's value is an empirical constant of the new medium — arguably its most important one — and the instruments now exist on paper. And the archive binds itself, with maintenance semantics this time: the module and dossiers ship with the paper, and the archive will maintain a public flip log — gate-coordinate changes on arguing surfaces above a stated reach threshold, quarterly cadence, inclusion criteria published with the first entry — which is due with this paper's imprint and carries two pins: the dated observation that the September removal reached devices, and the next named wave of the same substitution with its announced date, against which D5 and P3 now stand (§7) — and the lapse rule of The Imprint and the Exit (this archive): two missed quarters and the log is declared lapsed as of its last entry, so a dead log can never impersonate a quiet one.

The gate was named there, formalized here on two coordinates, and priced. What it is worth — the number that decides whether the reassuring century of media research still describes the world — is now one fielded module, one panel, and one honestly-identified flip away from being a measurement instead of an argument.

---

§ XIIAppendix — Three Protocols, Pre-registered at This Draft

A. Series discovery for the retrospective flips (D2). Before any outcome is examined: enumerate belief- and preference-item series fielded continuously across each flip date (the AI Overviews waves from May 2024; the March 2025 European Meta AI rollout) from a named list of panels and trackers — the national election and social-survey panels, the probability-panel trackers, the cross-national news-use surveys. A series qualifies only with identical item wording in at least two waves before and two after the flip, sample geography overlapping the rollout, and an exposure proxy (a search-use or device item) in the same instrument. The list is registered before outcomes are opened; each qualifying series carries its directional prediction from §9 before analysis; any series located after outcomes are seen is reported as discovered, analyzed, and marked ineligible for confirmation. The first candidate is enumerated and adjudicated here rather than counted as zero: the Reuters Institute's cross-national series — the Digital News Report (45 markets, annual, identical wording) and the six-country generative-AI series (2024, 2025) — against the AI Overviews country rollout mapped by Aral, Li and Zuo (2026), which left France, Turkey, China and Cuba untreated through 2025, with France receiving AI Overviews and AI Mode on July 22, 2026. The exposure-proxy condition is met (the 2025 wave fielded "seen an AI-generated answer in the last week" in France, at 29 percent, pre-treatment); the geography condition is met; the two-waves-before condition is met for France. The series qualifies for exposure; whether it qualifies for the protocol turns on the outcome condition — a belief or preference item at identical wording across the same waves — which the trust-in-news and platform-attitude items may satisfy and the protocol has not yet confirmed at item level. Disposition at this draft: candidate, not yet qualifying — adjudicated at the flip log's first entry, on the record, with the item-level check printed. The count of qualifying series remains zero, and the zero is now printed against a named candidate rather than assumed.

B. The "first field datum" search (§5's existence claim). The claim that no population-scale measurement exists of persuasion by arguing surfaces on people who did not choose them is a negative claim, and it is maintained by a repeatable search, not by assertion: the scholarly indexes and preprint servers, including the registries of pre-registered studies; a fixed term list crossing the surface names with persuasion, attitude, belief, preference and the identification vocabulary (field experiment, natural experiment, rollout, difference-in-differences, encouragement); no date floor; inclusion requiring randomized or quasi-experimental identification, a population not defined by having chosen the surface, and a belief or preference outcome. Run at the coverage round (August 2026) and re-run at this imprint (September 4, 2026): zero qualifying studies. The nearest misses are recorded because they instruct, and the register is complete against this draft's search: the biased writing-assistant experiments (Williams-Ceci et al., 2026: assigned argumentative content inside a chosen, non-argumentative act, N = 2,582, stated positions moved, most participants unaware — the closest miss, because the exposure is unsought at the item level, but consenting participants in a laboratory task); the opinionated-autocomplete study it extends (Jakesch et al., 2023); the panel RCT on AI search summaries (Xu et al., 2025: assigned exposure, consenting panelists); the AI Overviews removal experiment (Agarwal and Sen, 2026: field assignment, behavioral outcomes only); and the podcast-vocabulary study (Yakura et al., 2024/2026: population scale, no consent required, an uptake outcome rather than a belief one — the D6 template). Re-run at this revision (September 5, 2026) by the archive's referee panel (Notes, review record): zero qualifying studies. The search re-runs at every revision and at each flip-log memo; the first qualifying study retires the existence claim and converts §9's projections into scored comparisons.

C. Codability of the seam (§4, §6). Before fielding: at least one hundred free-text descriptions of AI encounters, collected on the module's screener population, with a quota of at least thirty descriptions of escalated exchanges so the seam cell is not starved (at single-digit seam prevalence, a hundred unquota'd descriptions yield about five seam cases and no test), coded independently by two coders into gate-of-encounter (chosen, default, unavoidable, delegated) and depth grade (summary, answer, dialogue, agent), with the escalation seam coded separately; agreement floor Cohen's κ = 0.70 on the joint gate × depth cell — the quantity the census reports — not on each coordinate separately, since κ = 0.70 per coordinate implies roughly 0.57 on the two-coordinate cell and 0.46 on three; and a second criterion from the donation subsample: self-reported just there must agree with logged default status on at least eight of ten audited encounters. Below either floor, the item is redesigned and the test repeated before any fielding claim. Results publish with the timing test of §6.

Notes

Registers. C marks conceptual claims carried on cited scholarship; E marks empirical claims re-pullable from named public sources, each carried at its stated grade — company-stated, tracker, survey, or query-distribution; D marks design claims, each with instrument, estimand, population, and failure mode; P marks the five projections — and nothing else in the paper predicts anything.

Rules. No capability inflation (nothing exceeds the measured record's sizes, in any mood, including the first stage). No null laundering (the protective nulls enter at full strength; the Expression of Concern is printed beside the study it concerns). No boundary smuggling (chosen and unchosen are never conflated; complier labels travel with every encouragement estimate; company figures are graded as such). The paper states no field-effect number, because none exists and manufacturing one is the vice it indicts.

Prior work. The Broadcast and the Model (this archive, 2026) closed by naming what it could not measure; this paper specifies the measurement. Its evidence is cited at the primary studies it synthesized; the earlier paper is cited once, as the synthesis. The bridge section applies instruments specified in The Wage and the Worth and The Praise and the Praiseworthiness and builds none.

Review record. One internal adversarial round (nine majors — among them P3's missing instrument, repaired by D5; the module's inability to construct the surface × gate × depth cell, repaired with per-surface depth and seam items; the September estimand restated as a treatment change on invokers; the encouragement power omitting intention-to-treat dilution, re-priced at π ∈ [0.2, 0.5]) and the standing coverage round (seventeen findings, six mandatory — the existence claim narrowed against Xu et al.; the Epstein program as lab ancestry with its single-lab caveat; the D2 and D3 ancestors; the Expression of Concern; DSA Article 40 and the D4 split; the incidental-exposure lineage). Referee report and response letter are filed with the archive. The archive panel of September 2026 (four referees, blind; decision memo and reports under reviews/012/) produced v4: the August baseline window recorded as closed with no wave fielded by this archive; P3 re-indexed to the next named flip under a sunset rule; the dossier rebuilt as v2 beside v1; D3 priced as a range with its inputs printed; the anchor surface regraded to default-conditional-on-activation with the user-week conversion; the codability floor moved to the joint cell with a seam quota and a concordance criterion; D2-September named as difference-in-differences with an event window; the bridge's equivalence bound; D6 and P5; the persuasion, incidental-exposure, audit, and rollout shelves cited; the private numbering retired.

Verification record. The Berelson epigraph confirmed verbatim against the chapter text on the third check (the comma earlier drafts carried after "communication" was ours). The Science Expression of Concern re-attributed at imprint to the study it concerns (Costello, Pennycook & Rand 2024); the durability figures traced to their studies — Hackenburg et al. 2025 (36–42 percent of the immediate effect at one month; Science 390, December 2025) and Lin et al. 2025 (slightly more than a third of the preference shift at re-contact in the US experiment — b = 0.34, 95 percent CI 0.23 to 0.44, a mean of 41 days later, 1,936 returning; read at the primary source, main text and Methods) — and unaffected. An earlier draft's "the date held" regraded at the archive panel: the September 4 schedule, its scope, and its per-device irreversibility are confirmed at press grade quoting the vendor; execution is unobserved at this draft. The Pew item printed verbatim (60 / 30 / 10; fielded February 17–23, 2026; N = 5,119). Google's terms pinned to the I/O 2026 keynote transcript ("monthly active users"; no definition published for AI Overviews); the Gemini app's 950 million pinned to Alphabet's second-quarter 2026 statement, correcting an earlier draft's "billion." Meta's last company figure pinned to the May 2025 shareholder meeting; its 2026 earnings calls disclose no Meta AI monthly-active figure, so the ~1.5 billion stays at tracker grade. The AI Overviews activation audit (Xu, Iqbal & Montgomery 2026) and removal experiment (Agarwal & Sen 2026) surfaced by the Appendix B search and integrated. Open: the Expression of Concern's disposition; any fresh Meta AI company figure; a dated observation that the September removal reached devices (the flip log's first pin); the page of Berelson's accidental-exposure passage in the reprint the epigraph was checked against; Withdrawn at the archive panel: the note an earlier draft owed The Broadcast and the Model — that paper attributes the surviving third to the election-dialogue study's own follow-up, correctly; the one-month figure here is a separate finding.

Companion artifacts. The execution brief; the corpus index (the paper's source ledger) with its coverage-round, imprint, and panel appendices; the claim outline; the census specification, survey module, and design dossiers (v1, and the v2 the archive panel required, filed beside it); the projection set (v1 and v2); the referee report and response letter — filed with the archive.

References

Agarwal, S. & A. Sen (2026). "The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment." SSRN 6513059 (April 3, 2026; revised July 8, 2026). · Allcott, H., M. Gentzkow, W. Mason, A. Wilkins, P. Barberá et al. (2024). "The effects of Facebook and Instagram on the 2020 election: A deactivation experiment." Proceedings of the National Academy of Sciences 121(21):e2321584121. · Angrist, J. D., G. W. Imbens & D. B. Rubin (1996). "Identification of Causal Effects Using Instrumental Variables." Journal of the American Statistical Association 91(434):444–455. · Aral, S., H. Li & R. Zuo (2026). "The Rise of AI Search: Implications for Information Markets and Human Judgement at Scale." arXiv:2602.13415. · Berelson, B. (1948). "Communications and Public Opinion." In W. Schramm (ed.), Communications in Modern Society. Urbana: University of Illinois Press; reprinted in W. Schramm (ed.), Mass Communications (1960), 527–543. · Boeschoten, L., J. Ausloos, J. E. Möller, T. Araujo & D. L. Oberski (2022). "A framework for privacy preserving digital trace data collection through data donation." Computational Communication Research 4(2). · Braghieri, L., R. Levy & A. Makarin (2022). "Social Media and Mental Health." American Economic Review 112(11):3660–3693. · Costello, T. H., G. Pennycook & D. G. Rand (2024). "Durably reducing conspiracy beliefs through dialogues with AI." Science 385(6714):eadq1814. [Under an Editorial Expression of Concern (Thorp 2026), unresolved at this draft.] · DellaVigna, S. & E. Kaplan (2007). "The Fox News Effect: Media Bias and Voting." Quarterly Journal of Economics 122(3):1187–1234. · Durante, R., P. Pinotti & A. Tesei (2019). "The Political Legacy of Entertainment TV." American Economic Review 109(7):2497–2530. · Enikolopov, R., M. Petrova & E. Zhuravskaya (2011). "Media and Political Persuasion: Evidence from Russia." American Economic Review 101(7):3253–3285. · Epstein, R. & R. E. Robertson (2015). "The search engine manipulation effect (SEME) and its possible impact on the outcomes of elections." Proceedings of the National Academy of Sciences 112(33):E4512–E4521. · Epstein, R., J. Lee, C. Mohr & V. R. Zankich (2022). "The Answer Bot Effect (ABE): A powerful new form of influence made possible by intelligent personal assistants and search engines." PLOS ONE 17(6):e0268081. · European Commission (2025). Commission Delegated Regulation (EU) 2025/2050 on data access for vetted researchers under Article 40 of Regulation (EU) 2022/2065 (the Digital Services Act). · Fletcher, R. & R. K. Nielsen (2018). "Are people incidentally exposed to news on social media? A comparative analysis." New Media & Society 20(7):2450–2468. · Fletcher, R. & R. K. Nielsen (2024). What Does the Public in Six Countries Think of Generative AI in News? Oxford: Reuters Institute for the Study of Journalism, May 2024. · Freelon, D. (2018). "Computational Research in the Post-API Age." Political Communication 35(4):665–668. · Gerber, A. S., D. Karlan & D. Bergan (2009). "Does the Media Matter? A Field Experiment Measuring the Effect of Newspapers on Voting Behavior and Political Opinions." American Economic Journal: Applied Economics 1(2):35–52. · Goldstein, J. A., J. Chao, S. Grossman, A. Stamos & M. Tomz (2024). "How persuasive is AI-generated propaganda?" PNAS Nexus 3(2):pgae034. · Google (2019). Google Assistant at CES 2019: the Assistant "available on one billion devices by the end of January," company statement, January 7, 2019. · Google (2026). Pichai, S., "I/O 2026: Welcome to the agentic Gemini era," keynote transcript, May 19, 2026; Alphabet second-quarter 2026 results, July 22, 2026. · Guess, A. M., N. Malhotra, J. Pan, P. Barberá, H. Allcott et al. (2023). "How do social media feed algorithms affect attitudes and behavior in an election campaign?" Science 381(6656):398–404. · Hackenburg, K. & H. Margetts (2024). "Evaluating the persuasive influence of political microtargeting with large language models." Proceedings of the National Academy of Sciences 121(24):e2403116121. · Hackenburg, K., B. M. Tappin, L. Hewitt, E. Saunders, S. Black, H. Lin, C. Fist, H. Margetts, D. G. Rand & C. Summerfield (2025). "The levers of political persuasion with conversational artificial intelligence." Science 390(6777):eaea3884 (December 4, 2025); arXiv:2507.13919. · Hermann, E., M. Morgan & J. Shanahan (2021). "Television, Continuity, and Change: A Meta-Analysis of Five Decades of Cultivation Research." Journal of Communication 71(4):515–544. · Jachimowicz, J. M., S. Duncan, E. U. Weber & E. J. Johnson (2019). "When and why defaults influence decisions: a meta-analysis of default effects." Behavioural Public Policy 3(2):159–186. · Jakesch, M., A. Bhat, D. Buschek, L. Zalmanson & M. Naaman (2023). "Co-Writing with Opinionated Language Models Affects Users' Views." CHI '23: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, DOI 10.1145/3544548.3581196. · Katz, E. & P. F. Lazarsfeld (1955). Personal Influence. Glencoe: Free Press. · Klapper, J. T. (1960). The Effects of Mass Communication. Glencoe: Free Press. · Lin, H., G. Czarnek, B. Lewis, J. P. White, A. J. Berinsky, T. Costello, G. Pennycook & D. G. Rand (2025). "Persuading voters using human–artificial intelligence dialogues." Nature 648:394–401. · Marbach, M. & D. Hangartner (2020). "Profiling Compliers and Noncompliers for Instrumental-Variable Analysis." Political Analysis 28(3):435–444. · Martin, G. J. & A. Yurukoglu (2017). "Bias in Cable News: Persuasion and Polarization." American Economic Review 107(9):2565–2599. · Matz, S. C., J. D. Teeny, S. S. Vaid, H. Peters, G. M. Harari & M. Cerf (2024). "The potential of generative AI for personalized persuasion at scale." Scientific Reports 14:4692. · Meta Platforms (2025–2026). Annual meeting of shareholders, May 2025 (Meta AI at one billion monthly actives); first- and second-quarter 2026 earnings call transcripts, April 29 and July 29, 2026. · Metaxa, D., J. S. Park, R. E. Robertson, K. Karahalios, C. Wilson, J. Hancock & C. Sandvig (2021). "Auditing Algorithms: Understanding Algorithmic Systems from the Outside In." Foundations and Trends in Human–Computer Interaction 14(4):272–344. · Nyhan, B., J. Settle, E. Thorson, M. Wojcieszak, P. Barberá et al. (2023). "Like-minded sources on Facebook are prevalent but not polarizing." Nature 620:137–144. · Ohme, J., T. Araujo, C. H. de Vreese & J. T. Piotrowski (2021). "Mobile data donations: Assessing self-report accuracy and sample biases with the iOS Screen Time function." Mobile Media & Communication 9(2):293–313. · Pew Research Center (2025). "Google users are less likely to click on links when an AI summary appears in the results." July 2025. · Pew Research Center (2026). Americans and AI 2026: Chatbots, Smart Devices and Views on Impact. June 17, 2026 (survey of U.S. adults, February 17–23, 2026; N = 5,119). · Piccardi, T., M. Saveski, C. Jia, J. T. Hancock, J. L. Tsai & M. S. Bernstein (2025). "Reranking partisan animosity in algorithmic social media feeds alters affective polarization." Science (adu5584). · Poungpeth, N., N. Clark & T. Mitra (2026). "Spontaneous Persuasion: An Audit of Model Persuasiveness in Everyday Conversations." arXiv:2604.22109. · Prior, M. (2007). Post-Broadcast Democracy: How Media Choice Increases Inequality in Political Involvement and Polarizes Elections. Cambridge: Cambridge University Press. · Reuters Institute for the Study of Journalism (2026). Digital News Report 2026. Oxford. · Ross Arguedas, A. (2026). "Emerging uses of AI chatbots for news and what it means for journalism." In Digital News Report 2026. Oxford: Reuters Institute for the Study of Journalism. · Salvi, F., M. Horta Ribeiro, R. Gallotti & R. West (2025). "On the conversational persuasiveness of GPT-4." Nature Human Behaviour 9(8):1645–1653. · Schäfer, S. (2023). "Incidental News Exposure in a Digital Media Environment: A Scoping Review of Recent Research." Annals of the International Communication Association 47(2):242–260. · Simon, F. M., R. K. Nielsen & R. Fletcher (2025). Generative AI and News Report 2025: How People Think About AI's Role in Journalism and Society. Oxford: Reuters Institute for the Study of Journalism, October 7, 2025. · Thorp, H. H. (2026). "Editorial Expression of Concern." Science 392(6803):1131. DOI 10.1126/science.aej2383. · Thorson, K. (2020). "Attracting the news: Algorithms, platforms, and reframing incidental exposure." Journalism 21(8):1067–1082. · Wagner, M. W. (2023). "Independence by permission." Science 381(6656):388–391. · Weeks, B. E. & D. S. Lane (2020). "The ecology of incidental exposure to news in digital media environments." Journalism 21(8):1119–1135. · Williams-Ceci, S., M. Jakesch, A. Bhat, K. Kadoma, L. Zalmanson & M. Naaman (2026). "Biased AI writing assistants shift users' attitudes on societal issues." Science Advances 12(11):eadw5578. · Xu, H., U. Iqbal & J. M. Montgomery (2026). "Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact." arXiv:2605.14021. · Xu, Y., S. Dash, S. Kang, W. Liao & E. S. Spiro (2025). "AI summaries in online search influence users' attitudes." arXiv:2511.22809. · Yakura, H., E. Lopez-Lopez, L. Brinkmann, I. de la Serna, L. Kirfel, P. Gupta, I. Soraperra, T. F. Eisenmann, D. U. Wulff & I. Rahwan (2024/2026). "Empirical evidence of Large Language Model's influence on human spoken communication." arXiv:2409.01754 (v4, July 2026). · Gauthier, G., R. Hodler, P. Widmer & E. Zhuravskaya (2026). "The political effects of X's feed algorithm." Nature, article s41586-026-10098-2. · This archive (armchair.marshallcahill.com), 2026: The Broadcast and the Model (No. 005); The Spectator and the Agent (No. 007); The Wage and the Worth (No. 008); The Praise and the Praiseworthiness (No. 009); The Imprint and the Exit (No. 011).

The companion deck — The Gate and the Field — sits beside this paper, with the survey module, the census specification, the design dossiers (v1 and v2), the projection set (v1 and v2), the execution brief, the corpus index, and the referee round filed with the archive. The gate was named in The Broadcast and the Model, formalized here on two coordinates, and priced. What it is worth is one fielded module, one panel, and one honestly-identified flip away from being a measurement instead of an argument — and the flip log opens with the first quarterly entry, under a lapse rule that lets a dead log never impersonate a quiet one.