Blog · Review Essay · Last reviewed June 16, 2026

Propaganda and the Administration of Belief

Jacques Ellul's Propaganda: The Formation of Men's Attitudes is not just a book about lies, slogans, posters, or state messaging. Its harsher claim is that modern propaganda is an environment: organized communication, technical administration, social pressure, measurement, education, entertainment, news, and habit converging until people experience adaptation as common sense.

Belief administration, in this review, means the organized shaping of what people treat as normal, credible, urgent, actionable, or professionally responsible. It does not require a single falsehood. It works when repeated formats, metrics, defaults, rankings, narratives, and institutional routines train the next acceptable move.

The audit question is therefore concrete: which institution selected the source, which ranking or retrieval rule moved it, which interface made it feel complete, which sponsor or model produced it, which action path followed it, and which record would let an outsider reconstruct that route later?

The Book

Propaganda: The Formation of Men's Attitudes was first published in French as Propagandes in 1962 and appeared in English from Knopf in 1965, translated by Konrad Kellen and Jean Lerner. Internet Archive's record for the 1965 Knopf edition lists xxii, 320, vii pages, bibliographical references, and subjects under propaganda. Google Books likewise lists the Knopf Doubleday edition at 320 pages. Open Library identifies the 1973 Vintage edition as part of a work first published in 1962 and lists the 1973 ISBNs 0394718747 and 9780394718743. Amazon's current listing for the Vintage paperback gives January 12, 1973, 352 pages, and the same ISBN-13.

Ellul was a French political and social scientist, Protestant theologian, and philosopher of technology. Britannica describes him as best known for his analysis of "technique," the larger social regime in which methods are organized around maximum efficiency. The International Jacques Ellul Society's biography places propaganda alongside technology and Marx as subjects he regularly taught, and notes his background in law, resistance work during the Vichy period, local public service, and a long academic career around Bordeaux.

That biography matters because Propaganda is not detachable from The Technological Society. Ellul's propaganda is not merely false speech. It is communication under technical conditions: planned, continuous, researched, organizational, and tied to the administrative need to make people act inside large systems. The propagandist is not only a demagogue. The propagandist can be a bureaucracy, party, media system, school, corporation, platform, civic campaign, or public culture that learns how to produce adjustment.

Propaganda as Environment

The book's most useful move is to shift attention from message to milieu. A narrow theory asks whether a statement is true or false, who sent it, and what motive the sender had. Ellul asks a wider question: what kind of society needs continuous formation of attitudes, and what institutions make that formation ordinary?

This is why the book still reads as contemporary. A feed does not persuade only by carrying a lie. It persuades by deciding what returns, what trends, what looks popular, what feels urgent, what becomes searchable, what receives a thumbnail, what appears beside what, what disappears into friction, and what is repeated until it becomes background. An answer engine does not shape belief only when it hallucinates. It shapes belief by compressing sources into a single voice, choosing which disputes become visible, and teaching users what kind of answer feels complete.

Ellul is especially sharp on the loneliness of mass communication. Modern subjects can be isolated as individuals while processed as members of a crowd. That structure now describes much of digital life. The user sits alone with a phone, search box, companion bot, dashboard, or productivity agent. Behind the interface, the user is also a member of audience segments, lookalike populations, risk groups, engagement cohorts, workplace metrics, advertising categories, voter models, and training data. The interface feels personal. The administration is statistical.

The important distinction is that propaganda is not identical with persuasion, publicity, education, moderation, or interface design. It appears when the environment repeatedly narrows a person's field of plausible interpretation while presenting that narrowing as neutrality, efficiency, safety, professionalism, or ordinary convenience. That narrower definition keeps the word from becoming a universal insult and makes it useful for audit.

That audit frame matters because a propaganda system can be accurate in fragments while misleading in arrangement. A feed may surface true posts in a distorted order. A search page may cite real sources while suppressing source-class diversity. A dashboard may display valid measurements while making unmeasured harms administratively invisible. An AI answer may cite documents while erasing disagreement, uncertainty, or the path from source to synthesis. The harm is not always a bad sentence. Sometimes it is a managed field of attention.

That makes Propaganda a missing ancestor for Manufacturing Consent, Network Propaganda, Invisible Rulers, and The Chaos Machine. Later books map ownership, media ecosystems, influencers, recommender systems, and networked publics. Ellul supplies the older diagnosis: propaganda becomes strongest when it appears not as propaganda but as the normal informational atmosphere of a functioning society.

Integration, Not Just Agitation

One reason the book feels more useful in the AI era than many older propaganda studies is its distinction between explosive persuasion and integrative adjustment. People often imagine propaganda as agitation: inflaming crowds, naming enemies, driving action, mobilizing crisis. Ellul also cares about integration propaganda: the quieter process by which people are adapted to an existing social order.

Integration is less cinematic and more durable. It tells people how to be modern, productive, realistic, informed, healthy, patriotic, employable, safe, rational, data-driven, innovative, resilient, or compliant. It does not always shout. It can arrive as training, public service messaging, corporate culture, school platforms, lifestyle media, wellness programs, productivity dashboards, risk scoring, best-practice documentation, or expert synthesis.

This is where AI systems become politically interesting before they become spectacularly deceptive. A workplace copilot can teach what a good memo sounds like. A school platform can teach what a good explanation looks like. A benefits chatbot can teach which claims are legible. A health portal can teach which symptoms deserve attention. A civic assistant can teach which policy options are available for thought. A recommender can teach what people like you supposedly like. Each system forms attitudes by narrowing the next move.

That does not make every interface propaganda. It means the propaganda question should be asked wherever a system repeatedly trains people to fit an institutional pattern while presenting that pattern as neutral assistance. The core issue is not whether the screen says something ideological. It is whether the screen makes a specific form of life easier to accept, harder to contest, and easier to measure.

Rational Propaganda

Ellul's account of rational propaganda is one of the book's best AI-era tools. Propaganda need not be anti-factual. It can work through facts, statistics, maps, charts, expert language, polls, forecasts, and administrative categories. The point is not that numbers are fake. The point is that numbers can become persuasive because they look like reality after institutional processing.

This matters for dashboards, benchmarks, model evaluations, risk scores, predictive systems, performance metrics, and AI-generated reports. A score can persuade a manager that a worker is underperforming. A model card can persuade a buyer that a system is safe enough. A benchmark can persuade a public that one model is smarter. A risk flag can persuade an agency that a household needs intervention. A summary can persuade a reader that a dispute has been settled. The machinery of persuasion is often the aura of disciplined objectivity.

Read beside Trust in Numbers, The Tyranny of Metrics, and the benchmark problem, Ellul's argument becomes concrete. Technical truth can become social command when institutions skip the interpretive step. The number may be useful. The danger begins when the number is treated as if it has already resolved what should be done.

The AI Reading

The obvious AI reading is synthetic propaganda: generated articles, fake local news, deepfake candidates, cloned voices, bot swarms, automated microtargeting, and cheap persuasive copy. That risk is real, but Ellul points to a deeper one. AI does not merely create more messages. It changes the environments in which people learn what messages are worth believing.

Answer engines shift trust from a visible field of sources to a synthesized response. Companion systems shift persuasion from public address to private emotional continuity. Agentic assistants shift action from deliberate choice to delegated execution. Recommendation systems shift public attention through invisible ranking. Enterprise copilots shift institutional memory into model-mediated summaries. Each layer can be useful. Each layer can also become a belief-administration surface.

The recursive pattern is the key. A system frames a question. The user acts inside that frame. The action becomes data. The data improves the frame. Other users encounter the updated frame as evidence of what people do, ask, choose, fear, or prefer. Over time, the machine does not merely reflect reality. It helps produce the reality it later measures.

This is why Propaganda belongs beside the answer-engine problem, The Hype Machine, The Filter Bubble, and AI persuasion. The future of propaganda is not only a better fake. It is a better-adjusted interface: one that knows the user's context, speaks in the right register, remembers the prior confession, supplies a convenient explanation, and routes the next action before the user has fully named the choice.

The harder current case is not a fake post traveling alone but a synthetic public made legible to machines: generated comments, automated replies, cloned voices, influencer briefs, optimized headlines, and answer-engine citations that create evidence of attention for the next system to summarize. Once public reaction is partly manufactured, later synthesis can mistake the manufactured trail for social reality unless provenance, timing, coordination signals, and source classes stay visible.

The Belief Route

Ellul becomes most useful for AI governance when propaganda is treated as a route, not a mood. A belief route is the trace from source material to public action: source selection, retrieval, ranking, sponsorship, synthesis, display, personalization, social proof, correction, logging, and the button or workflow that follows. If that route cannot be reconstructed, no one can tell whether a person saw evidence, a sponsorship, a ranked sample, a generated summary, a synthetic crowd, or an institutionally preferred next step.

The route has to be audited at several points. Input: what source classes were included or excluded, and how fresh were they? Movement: what ranking, recommender, targeting, or retrieval rule made the item visible? Frame: what title, summary, thumbnail, confidence cue, label, citation, or answer format made it feel credible? Conversion: what action path followed the belief - share, buy, vote, report, apply, comply, disclose, escalate, or delegate? Memory: what record remains for correction, appeal, research, or incident review?

This connects Ellul's environment argument to recursive reality. A system routes a claim, people act on the routed claim, their action becomes a signal, and the signal changes future routing. The result can be a machine-mediated common sense: not a single lie, but a loop in which measured reaction becomes evidence for the next intervention.

Governance and Safety

Ellul's useful governance move is to separate content control from environmental control. Removing a false post is not the same as governing a system that ranks, personalizes, synthesizes, recommends, advertises, labels, remembers, and routes action. A belief-administration surface can be lawful, helpful, and factually competent while still training users to accept a source hierarchy, a metric, an institutional category, or a preferred action without enough friction to inspect it.

A useful governance map has four layers. Content governance asks whether a particular artifact is false, illegal, deceptive, sponsored, or synthetic. Distribution governance asks how ranking, recommendation, advertising, search, and moderation move that artifact through a public. Synthesis governance asks how answer engines and copilots compress many artifacts into one voice. Agent governance asks what the system can do after persuasion: file, buy, route, recommend, escalate, or execute. Belief administration becomes dangerous when those layers are treated as separate problems even though users experience them as one environment.

The governable object is the whole persuasion route. For an answer engine, feed, chatbot, or agent, that route includes the source corpus, retrieval and ranking logic, personalization variables, sponsorship and ad status, synthetic-media markers, model and prompt version, generated summary, action buttons, logs, correction path, and incident record. A safety review that inspects only the final output misses the environment that made the output authoritative.

By June 16, 2026, the current policy context is no longer only theoretical. The European Commission describes the Digital Services Act as requiring very large online platforms and search engines to assess systemic risks linked to their services, including risks to fundamental rights, public security, and electoral processes, and to address transparency around advertising, recommender systems, and content moderation. It also points to independent audits, vetted researcher access, non-profiling recommender options, and public ad repositories for VLOPs and VLOSEs. That approach does not solve propaganda. It does recognize that the architecture of attention can be a public-risk system.

The AI layer adds provenance and disclosure duties. The Commission's June 10, 2026 Code of Practice on Transparency of AI-Generated Content supports Article 50 of the EU AI Act, whose transparency obligations are scheduled to apply from August 2, 2026. The Commission frames those obligations around marking and detection of AI-generated content, labeling of deepfakes and certain AI-generated publications, and the risk of deception and manipulation. NIST's AI Risk Management Framework gives a broader management vocabulary - govern, map, measure, and manage - for documenting context, impacts, testing, monitoring, feedback, and accountability over an AI system's lifecycle. C2PA supplies technical specifications for certifying the source and history of media content.

The safety implication is practical. A platform, answer engine, companion system, civic chatbot, school tool, or workplace copilot should be reviewed not only for bad outputs but for the attitudes it trains. Does it preserve source trails? Does it reveal ranking and retrieval boundaries? Does it label sponsorship and synthetic media where relevant? Does it keep correction and appeal paths usable? Does it log enough of the route from evidence to output for audit? Does it test persuasive effects, especially for vulnerable users, political contexts, medical contexts, employment contexts, and public emergencies? Does it prevent fabricated social proof from becoming evidence for the next summary?

The minimum artifact for such systems is a belief-route log: source classes, retrieval and ranking settings, personalization variables, sponsor and ad status, synthetic-media markers, model and prompt versions where relevant, generated summaries, visible labels, action buttons, correction paths, and incident records. The log does not need to expose private user data to the public. It does need to make independent review possible under appropriate privacy, security, and researcher-access controls.

This is also where provenance must be kept in proportion. A C2PA credential, watermark, label, or model disclosure can help people inspect origin and alteration. It cannot prove that a claim is true, fair, representative, or democratically legitimate. The same distinction runs through provenance and content credentials, synthetic consensus firebreaks, public registers, and claim hygiene: accountability needs records of how belief was routed, not just a badge attached to the final artifact.

A credible safety case would therefore preserve more than labels. It would keep public-interest registers for high-risk persuasion surfaces, retain ad and sponsorship records, document source-class exclusions, publish meaningful recommender options, support vetted researcher access where law and privacy allow, and connect serious failures to AI incident reporting. The point is not to ban influence. It is to make consequential influence inspectable before it hardens into common sense.

Source Discipline

The book is strongest when used as a diagnostic of mechanisms, not as a license to call every disliked message propaganda. A disciplined reading separates five layers: the claim being made, the evidence offered for it, the institution that selected or repeated it, the incentive that made that selection likely, and the effect on public interpretation. Collapsing those layers turns critique into another machine for belief.

For current claims, this page relies on primary or official sources where possible: the European Commission and AI Act Service Desk for EU law and implementation context, NIST for AI risk-management vocabulary, C2PA for provenance specifications, and catalog or publisher records for book metadata. Secondary reviews are used for reception and criticism, not for legal or technical claims.

Evidence classes should stay labeled. An ad-library entry proves a paid placement existed; it does not prove persuasion succeeded. A provenance credential can show origin or editing history; it does not prove truth. A platform transparency report can show enforcement categories; it does not reveal every ranking effect. A peer-reviewed study can estimate mechanisms; it may not describe the current product. Screenshots and testimony can surface harm; they need chain-of-custody and context. Regulator filings and audit reports can be stronger, but only within their scope.

For AI-mediated communication, that discipline becomes operational. A generated answer should not be evaluated only as a paragraph. Ask what documents were retrieved, what source classes were excluded, what freshness policy was used, whether ranking was personalized, whether commercial or institutional relationships shaped the result, and how a correction would propagate. A citation is not the same as a retrieval audit. A label is not the same as accountability. A model card is not the same as evidence that the system behaves well in a particular public setting.

Keep separate evidence burdens for falsity, coordination, synthetic generation, reach, harm, sponsor identity, state attribution, and persuasive effect. A campaign can be synthetic without being effective. A claim can be false without being coordinated. A state-linked operation can fail to travel. A recommender can amplify authentic speech in a way that still creates public risk. Source discipline means naming the proven mechanism and leaving the unproven mechanism unclaimed.

The same caution applies to Ellul. His totalizing style makes invisible systems easier to see, but it can tempt readers into unfalsifiable suspicion. The useful question is not "is this propaganda?" as a moral verdict. The useful questions are narrower: what is repeated, what is hidden, what is made easy, what is made expensive, who benefits from the trained response, and what evidence would show the system is failing to preserve independent judgment?

Where the Book Needs Friction

Propaganda is powerful because it is totalizing, and risky for the same reason. Ellul often writes as if modern propaganda is nearly inescapable once technical society reaches a certain scale. That gives the book diagnostic force, but it can understate failure, resistance, boredom, counter-publics, institutional conflict, local knowledge, satire, refusal, and the fact that people often misunderstand, ignore, remix, or reject the systems trying to shape them.

Daniel Lerner's 1964 review of the French Propagandes in American Sociological Review was sharply critical, treating the book as more revealing of Ellul's method than of propaganda as a measurable social process. Randy Kluver's 1995 paper on Ellul's contribution to rhetorical theory notes a related criticism from social-scientific readers: Ellul's claims could appear tautological if the power of propaganda is assumed and then used to explain the evidence.

Those criticisms are worth keeping. A useful AI-era reading should not turn "propaganda" into a universal accusation. If every interface, institution, metric, school lesson, public-health campaign, search result, news story, training document, and chatbot reply is called propaganda in the same sense, the term stops helping. The task is to identify specific mechanisms: repetition, enclosure, source invisibility, emotional targeting, social proof, metric authority, absence of appeal, institutional dependency, and the conversion of behavior into confirmation.

The book also has the limits of its period. Its media world is organized around press, radio, film, television, parties, states, mass society, and Cold War ideological blocs. It does not know social media, recommender systems, search engines, model training, platform moderation, synthetic video, or personalized agents. The concepts travel, but they need updating. Today's propaganda environment is less centralized, more participatory, more metric-driven, and more likely to arrive as convenience.

What This Changes

The practical lesson is to audit the environment before arguing only about content.

For any AI-mediated system, ask what attitudes it forms. Does it train users to trust synthesis without source inspection? Does it make institutional categories feel natural? Does it convert uncertainty into a score? Does it route dissent into a dead end? Does it turn a temporary measurement into a durable identity? Does it make the preferred action easier than the contested action? Does it provide correction paths, human accountability, uncertainty labels, source trails, and ways to leave?

Then ask who needs the attitude. A state may need compliance. A platform may need engagement. A workplace may need measurable productivity. A vendor may need procurement confidence. A campaign may need emotional certainty. A school may need administrable learning. A model provider may need users to accept synthetic answers as normal public knowledge. Propaganda is often easiest to see after asking which institution benefits when a person stops asking the next question.

Propaganda remains valuable because it refuses the comforting idea that belief is shaped only by bad actors saying false things. Belief is shaped by environments, procedures, formats, incentives, measurements, and the daily training of attention. AI makes that lesson more urgent. The most consequential propaganda may not look like a poster. It may look like a helpful answer, a clean score, a friendly companion, a summarized record, a ranked feed, or a workflow that quietly teaches people which reality is operational.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books