Blog · Review Essay · Last reviewed June 23, 2026

If Then and the People Machine of Political Prediction

Jill Lepore's If Then: How the Simulmatics Corporation Invented the Future is a history of an early data company that tried to make voters, consumers, publics, and war zones computable. Its value now is not that Simulmatics was secretly competent. Its value is that the company's ambition still feels familiar: collect data, model the population, target the message, and treat prediction as authority over people who never saw the model.

For this review, political prediction means a model-backed claim about how groups will respond, joined to an institutional choice to act on that claim through message design, targeting, ranking, resource allocation, or suppression. Prediction becomes political power when the affected public cannot see the categories, inspect the evidence, or answer back before the forecast changes its world.

The practical lesson is to audit the model-to-message chain: data source, category, simulation, decision right, delivery channel, measured response, and correction path. Without that chain, prediction becomes a private substitute for public contestation.

The Book

If Then was published by Liveright in 2020. The National Book Foundation record lists the full title, Jill Lepore as author, Liveright / W. W. Norton & Company as publisher, ISBN 9781631496103, and the book as a 2020 National Book Awards nonfiction longlist title. Lepore follows the Simulmatics Corporation, opened in 1959, through electoral politics, advertising, journalism, social science, civil-rights crisis management, and Vietnam War research.

The book should be read as history of a governing temptation, not as proof that one company invented the modern platform. Simulmatics often overpromised. That is part of the point. The durable object is the institutional appetite: clients wanted the public to become calculable, consultants wanted social science to become a sales instrument, and computers lent old ambitions a new authority.

The story's central device is the "People Machine," Simulmatics' name for a system that used polling, demographic categories, behavioral assumptions, and computer runs to simulate public response. In Lepore's telling, the machine's glamour mattered as much as its performance. It sold an image of political control: the idea that publics could be decomposed into types, measured, simulated, and nudged by tailored messages.

Current Context

As of June 23, 2026, the Simulmatics pattern lives across three layers: paid political advertising, platform ranking, and generative persuasion. A campaign or civic actor can buy targeted ads, test messages, use recommender-friendly content, generate variants, analyze responses, and feed the results back into future outreach. The point is not that today's systems are the same as Simulmatics. It is that prediction, delivery, and measurement have fused into one operational loop.

European law now names part of that loop. The Digital Services Act requires very large online platforms and search engines to assess and mitigate systemic risks, including risks to civic discourse and electoral processes, and the Commission's 2024 election guidelines translate those duties into expectations around recommender systems, advertising transparency, crisis response, audits, and researcher access. Regulation (EU) 2024/900 on political advertising entered full application on October 10, 2025; the Commission says it requires clear labels and information about sponsor, cost, election or referendum context, and targeted audience when targeting or ad-delivery techniques are used. Commission Implementing Regulation (EU) 2026/818 sets technical arrangements for the European online political-ads repository.

Those rules still leave a harder question. Paid political ads are only one channel of influence. Organic ranking, search summaries, recommender amplification, direct messages, chatbots, influencer networks, synthetic polling, and generative reply systems can shape civic perception without looking like a classic ad buy. A review of predictive politics therefore has to ask not only who paid for a message, but what system selected the audience, what evidence justified the category, what delivery logs survive, and whether the public can contest being modeled.

The governance problem is not that campaigns learn from evidence. Public politics has always involved polling, canvassing, message discipline, and coalition arithmetic. The problem is closed-loop behavioral administration: the same system predicts a group, chooses the message, delivers it through a private surface, measures response, and treats that response as validation for the next round while the modeled public sees only its assigned fragment.

The People Machine

Simulmatics sits at the junction of several histories that usually get told separately. It belongs to the history of Cold War social science, because it borrowed confidence from psychology, communication research, systems analysis, and counterinsurgency. It belongs to the history of campaigns, because it offered politicians a way to treat voters as segmented response surfaces. It belongs to media theory, because it imagined communication less as public argument than as stimulus management.

The useful definition is this: the People Machine was not only a computer program. It was an institutional assembly of polling archives, demographic categories, social-science theories, consultants, clients, sales language, and mainframe time. Its output was a prediction, but its function was authority. It told decision-makers that a disputed public could be queried like a system.

The sharper definition is predictive administration: a governing actor treats model output as permission to act before the governed public has spoken in public. The model need not be accurate to matter. It only has to persuade the client, enter the workflow, and attach itself to money, media access, institutional authority, or coercive capacity.

In 2026 language, the Simulmatics pattern is not prediction alone. It is prediction plus action rights: the client does not merely receive a forecast, but a warrant to segment, target, test, fund, suppress, or redirect communication. That distinction matters because a forecast that stays in a notebook is evidence; a forecast wired into message delivery, platform ranking, policing, credit, benefits, or campaign strategy becomes governance.

The central ethical failure is category capture. A group is named for administrative convenience, scored for likely reaction, and then approached through that score. The category may contain real evidence, but it also starts shaping how the institution sees the people inside it. If the category is private, wrong, outdated, proxy-laden, or politically convenient, the public has no way to refuse the translation.

The case that anchors the book is the 1960 presidential election. In her New Yorker account, Lepore writes that Simulmatics opened for business on February 18, 1959, and that its 1960 election project sorted voters into 480 types and issues into 52 clusters. The article describes a voting-behavior simulation that let campaign actors ask what different moves might do to narrow slices of the electorate. The machine's real product was not a secret insight. It was the authority of the simulation, the feeling that the country's reactions had already been computed.

That is why the book feels so relevant to AI without being a book about contemporary AI. The company tried to operationalize a fantasy that still structures many technical systems: if the model is good enough, the public can be known in advance. Prediction becomes a substitute for listening. Segmentation becomes a substitute for politics. A population becomes legible as a set of probabilities before it is treated as a community of people who can answer back. The companion warning is in AI Snake Oil: weak prediction can still become powerful when institutions use it to launder judgment.

Simulation as Governance

The strongest reading of If Then is not "they invented Facebook." That claim is too neat. The stronger claim is institutional: Simulmatics shows how organizations use computation to change the meaning of uncertainty. A messy public becomes a dashboard. A disputed political question becomes a targeting problem. A conflict becomes an optimization task. A model then offers administrators, consultants, and clients the feeling that the social world has become more governable.

This is where Lepore's narrative connects with the site's recurring concern about recursive reality. A model does not merely represent a world from outside. Once institutions act on it, the model feeds back into the world it describes. Messages are targeted because a simulation predicted susceptibility. Public behavior then changes under targeted pressure. The next model treats that changed behavior as fresh evidence. Over time, prediction and production become difficult to separate.

The important word is not simulation but counterfactual. Simulmatics sold what-if worlds to clients who wanted to know what might happen if a speech changed, a message shifted, a group was isolated, or a policy was reframed. Modern versions can run those what-ifs with ad auctions, recommender tests, language-model-generated copy, voter files, and platform analytics. The governance issue is who sees the experiment and who is unknowingly enrolled in it.

That is why source discipline matters for simulations. A defensible simulation should disclose data provenance, category construction, missing variables, validation results, uncertainty, and decision rights. Without those, the what-if becomes a private constitution: it names which futures are worth preparing for while the people inside the modeled future remain outside the room.

Counterfactual tools are most dangerous when they turn from planning aids into behavioral experiments on unknowing publics. An institution may be entitled to ask "what if?" internally. It should not be entitled to hide which hypotheses it acted on, which groups were exposed, which alternatives were never tested, and which measured responses were later folded back into the system as if they were natural preference.

That feedback loop is also the political danger. When the public is approached as an object of behavioral management, democratic communication thins out. The citizen is no longer primarily someone to persuade in common view, but someone to classify, score, and address through a channel optimized for effect. The simulation becomes a quiet form of governance because it decides which possibilities institutions prepare for, which groups they notice, and which forms of refusal they treat as noise.

Governance and Safety

The current governance context turns If Then from history into a practical diagnostic. The EU Digital Services Act election guidelines, formally adopted in April 2024, tell very large online platforms and search engines to mitigate systemic risks to electoral processes under Article 35. The Commission lists relevant DSA obligations around terms and statements of reasons, recommender systems, crisis response, transparency, independent audits, online advertising transparency, and data access. That is the modern legal answer to the People Machine: not merely "was a message false," but "how did the delivery system shape democratic risk?"

Regulation (EU) 2024/900 on political advertising goes further into the targeting layer. The Commission says it entered full application on October 10, 2025, requires political ads to be clearly labelled, and requires information such as who paid, cost, and targeted audience when targeting or ad-delivery techniques are used. Article 18 permits personal-data-based targeting or ad delivery for online political advertising only under conditions including data collected from the data subject, separate explicit consent for political advertising, and no profiling using special-category personal data. Article 13 points to a European online political-ad repository; Commission Implementing Regulation (EU) 2026/818 set technical arrangements for common data structure, metadata, authentication, and a common API on April 9, 2026.

These rules also show the limits of ad transparency. Paid political advertising is only one persuasion channel. Organic ranking, recommender amplification, influencer networks, direct messages, public chatbots, synthetic polling, and generative reply systems can shape politics without fitting the old ad-buyer model. The DSA partly reaches recommender and platform risk, but no single disclosure page makes a simulated public democratically accountable.

NIST's AI Risk Management Framework supplies a broader standards vocabulary: risk is managed across design, development, use, and evaluation, with organizational functions such as govern, map, measure, and manage. Applied to political prediction, that means a serious safety program cannot stop at model accuracy. It has to document data provenance, category construction, intended use, affected groups, targeting rules, evaluation limits, human review, retention, audit access, and incident response.

A civic prediction audit should preserve data provenance, category definitions, targetable audience criteria, excluded and proxy variables, model or simulation assumptions, message variants, delivery logs, outcome metrics, human approval records, retention rules, and independent access for lawful oversight. It should also distinguish prediction from mandate: a forecast may justify inquiry, but it does not by itself justify suppressing, manipulating, or ignoring the people forecasted.

The minimum governance record is a prediction-use record: what was predicted, for whom, from which data, with what validation, under which uncertainty, by whose authority, connected to which message or intervention, delivered through which system, measured by which outcome, retained for how long, and open to which challenge path. That record is the difference between responsible strategy and invisible administration.

Three tests follow. First, publicity: can affected outsiders learn that prediction shaped the encounter? Second, contestability: can they challenge the category, data, or use? Third, reversibility: can the institution stop, correct, or roll back an intervention when the evidence fails? If those tests fail, better model accuracy may only make unaccountable power more efficient.

The practical controls are plain. Political and civic prediction systems need purpose limits, data minimization, sponsor disclosure, targeting disclosure, retention rules, independent researcher access where lawful, non-profiled recommender options for civic contexts, and public records that survive after the campaign. Those controls connect this review to AI and election integrity, AI persuasion, platform governance, recommender systems, political ad libraries, and platform risk assessment.

The AI-Age Reading

Modern AI systems make the Simulmatics ambition broader, faster, and more intimate. Campaigns, platforms, search engines, recommender systems, ad exchanges, workplace analytics, data brokers, and AI assistants now gather signals continuously rather than waiting for surveys and mainframe runs. Language models add another turn: they do not only select messages. They can generate them, personalize them, test them, translate them, and wrap them in conversational warmth.

The new risk is synthetic scale. Generative systems can produce many tailored variants cheaply, but the governance issue is still old: who can see the test, who is targeted, what is optimized, and whether persuasion remains public enough to be contested.

The more intimate risk is answer-shaped prediction. A user may not experience a campaign ad at all; they may experience search, chat, recommendation, donation flow, voter-information lookup, or community guidance. If the system has inferred the user's likely vulnerability or preferred frame and adapts the encounter accordingly, the old line between information and influence becomes too thin to govern by interface labels alone.

The relevant continuity is not technical identity. Simulmatics was crude by today's standards. The continuity is the governance posture: people are most manageable when they are rendered as data profiles, placed inside simulations, and addressed through interfaces that hide the institutional hand. In that sense, If Then belongs beside The Filter Bubble, The Attention Merchants, The Black Box Society, and The Age of Surveillance Capitalism. It gives those later systems a prehistory.

The book also clarifies why AI governance cannot be treated as a purely model-internal property. A perfectly obedient targeting system can still serve a manipulative institution. A prediction engine can be accurate and anti-democratic at the same time. A simulation can reduce uncertainty for administrators while reducing agency for the people being administered. The risk does not require an AI system to be conscious, divine, or AGI. It requires data, incentives, interfaces, and an institution willing to act as if the forecast were consent.

That is why the privacy frame matters. Contextual integrity asks whether information flows fit the social setting in which data was given. Data brokers show how those settings can be broken apart and recombined into targeting fuel. The People Machine is primitive beside the contemporary data economy, but it names the same ethical fault line: a person gives off traces in one context, and an institution reuses them to predict, rank, persuade, or manage the person in another.

Limits of the Book

If Then is strongest as archival narrative and weakest when the subtitle is read too literally. Some reviewers have pressed on the gap between Simulmatics' claims and what its systems actually accomplished. That criticism matters. The company should not be granted the mythic power it tried to sell.

But the gap between capability and sales pitch is part of the lesson. Institutions do not need perfect prediction to reorganize themselves around predictive authority. A weak model can still change budgets, staffing, policy, campaign strategy, and public imagination if enough decision-makers believe that it sees the future. The dangerous object is not only the working machine. It is the institution that starts acting as if the machine has made human judgment obsolete.

The book also needs care around causality. Simulmatics did not need to be technically decisive to matter historically. Failed or exaggerated systems can still train buyers, journalists, academics, and politicians to imagine the public as something computable. The archive is therefore evidence of both technical practice and institutional desire.

Read this way, Lepore's book is a caution against technical amnesia. Today's AI politics did not appear from nowhere. It inherits older dreams of mass persuasion, administrative research, behavioral control, and machine-mediated certainty. The names changed. The appetite remained.

What This Changes

If Then clarifies a rule for AI-era institutions: prediction is never just epistemology. It is a proposed distribution of power. Someone defines the categories, someone buys the forecast, someone acts on the output, and someone else becomes the object of the action.

The practical questions follow. What population was modeled? Which variables were excluded? Which sensitive traits or proxies entered the system? What outcome was treated as success? Who could inspect the categories? Who benefited from targeting? How long were records kept? Could the affected public answer back, or only be measured?

That lesson sits beside Republic.com 2.0, The Hype Machine, The Tyranny of Metrics, How Data Happened, and Prediction Machines. A society can be governed through the dashboard before it notices that the dashboard has become political. Lepore's history gives the dashboard a genealogy, and it gives AI governance a warning: do not let simulated publics replace public contestation.

Source Discipline

This review separates four evidence types. Lepore's book is an archival synthesis and interpretation. Her New Yorker article and the Computer History Museum discussion are public-facing accounts with specific dates, actors, and project descriptions. EU law, Commission guidance, and NIST standards sources are current governance materials, not evidence about Simulmatics itself. Critical reviews are used to test the scale of Lepore's claims, especially where the company's marketing outpaced its demonstrated capacity.

Current legal sources are cited for obligations and dates, not as evidence that any jurisdiction has solved predictive politics. Reviews are used for reception and criticism, not for law.

Claims about prediction should keep four layers separate: model capability, client belief, deployed intervention, and public impact. Simulmatics could overstate capability and still shape client behavior. A modern campaign can have strong delivery logs and still lack evidence of persuasion. A law can require disclosure and still miss organic or conversational influence. Collapsing those layers turns critique into mythmaking.

That separation keeps the claim narrow. Simulmatics is not treated here as the direct ancestor of every platform, or as proof that every prediction system fails. The stronger and safer claim is that predictive systems become politically dangerous when their evidentiary limits are hidden from the people they target, while institutions treat the forecast as permission to act.

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