Politics Cannot be Simulated

There’s more to democracy than decision-making

Idealized Greek Landscape with resting Shepherds, Karl Friedrich Schinkel (1781-1841)

Civic life is the practice of taking responsibility with others for the shared conditions of common existence. Today, many of us hold a view of politics that is a long way from this ideal. We equate civic life with voting for representatives who make decisions on our behalf or engaging in limited forms of consultation when the opportunity arises. Modern politics is a system for making decisions rather than citizens.

Our predicament is partly a function of scale. The polities of the ancient world, though deeply limited in many respects, were small enough that those who held the right to engage in collective decision-making knew their peers and could formulate laws that would directly affect them. In gargantuan modern states, this kind of engagement is a practical impossibility.

Representative democracy exists, in part, to free us from the demands that direct self-rule would otherwise make of us. Classical proponents like Edmund Burke and James Madison argued that representation also improves the quality of governance by passing public preferences through a deliberative body capable of refining them and attending to the long term good.

On Madison’s account in Federalist 10, the legislature exists “to refine and enlarge the public views” by passing them through “a chosen body of citizens, whose wisdom may best discern the true interest of their country.” In Federalist 51, he argues that “ambition must be made to counteract ambition,” that political structures should compensate for the imperfection of those who govern and those who are governed. In other words, representative democracy can produce stable public decisions without requiring citizens to be highly formed civic actors.

Of two modes of civic life, formation and aggregation, the latter is taken to be the primary role of an advanced democracy. As political systems strain under the complexity of the modern world, AI threatens to remake institutions like education, entertainment, economic systems, the media, and various professions. Thinking machines do cheaply the things institutions were built for (such as drafting legal contracts or writing and researching) and make possible things they were not designed for (like one-on-one tutoring for every student or autonomous agents that act on a person’s behalf).

These twin pressures demand new kinds of institutions, especially political institutions to help societies navigate the diffusion of AI into public and private life. Various kinds of “AI for democracy” projects fit this account by appearing to remake democratic decision-making through, for example, summarization techniques for lawmakers or fostering agreement on contentious issues.

This assessment gets things backwards. AI represents, at least in the context of democracy, a small-c conservative turn that may entrench an aggregative view of civic life at exactly the point at which new possibilities allow us to better ask what civic life is and why it matters. The technology could prove to be less a centrifugal force pulling political institutions apart and more a centripetal force gluing a particular kind of institutional arrangement into place.

A view that takes civic life to be about primarily producing better and more legitimate decisions through preference aggregation has been dominant for two centuries. The AI and democracy project takes that settlement and makes it more efficient, capable, pervasive, and considerably harder to dislodge. It is the terminus of the Madisonian project, a sophisticated apparatus for generating politically resonant outputs without bearing the cost of citizen formation.

City on the hill

The first recognizable instance of politics-as-formation was the Athenian polity, a self-governing political community that emerged on the peninsula of Attica in southern Greece around the late sixth century BCE. Across two and a half thousand square kilometers, roughly the size of modern Luxembourg, the Athenian polity was made up of the city proper, surrounding villages and farmland, the silver mines country at Laurion in the south, and the port of Piraeus a few miles from the city.

From urban core to rural hinterland, all citizens who met the criteria were part of the same political whole. A farmer thirty miles from the city and the craftsman in the agora belonged to the same community and could expect, if they chose to make the journey, to vote in the same assembly. The polity was small enough that everyone within it was capable of representing everyone else.

Roughly 250,000 to 300,000 residents fell under the authority of Athens in the fourth century BCE. Of these, only adult male Athenians born of Athenian parents, say, about 30,000 men, were full citizens with political rights. Factoring out wives, daughters, slaves, foreigners, the political community was around ten to twelve per cent of the population.

The Assembly or Ekklesia was a gathering of all citizens together as a single body. It met about forty times a year on a hill called the Pnyx, a quarter of a mile south-west of the Athenian agora. The body dealt with the things that today we would call “high politics,” such as whether to go to war, whether to accept peace terms, what taxes to levy, what laws to pass, or how much grain to import. It also included grislier episodes like voting to slaughter the entire population of Mytilene and putting its own generals to death after the battle of Arginusae.

Another institution was the Council of Five Hundred, the Boule, which functioned as a kind of executive committee for the Assembly. It set the agenda and supervised the administration of the state. Its members were chosen by raffle, with fifty citizens each year drawn at random from each of the polity’s ten tribes. Most male Athenians spent a year of their adult life as part of the executive government of their state.

Together, these bodies meant that the participating citizen might attend the Assembly several times a year, serve a full year on the executive Council, sit as juror in major trials, and hold one or more rotational offices. For those who could take part, institutions allowed citizens to live a particular kind of political life. A handful of professional administrators could have made decisions far more quickly, yet the Athenians chose to bind themselves to more cumbersome arrangements because the essence of the system was to produce citizens.

Aristotle described this idea as hexis, a durable state of character cultivated through repeated action and habituation. We become just by doing just acts, temperate by doing temperate acts, and brave by doing brave acts. The citizen is formed by undergoing the discipline of ruling and being ruled in turn. They grow by attending the Assembly, sitting on the Council, hearing cases in the courts, and being held to account for their decisions.

Aggregation vs. formation

Civic life is taken to serve primarily one of two ends:

• Aggregating preferences. The process by which the preferences of citizens are gathered and weighed against one another to be converted into collective decisions. The citizen is a source of opinions that the political system needs in order to govern well.

• Forming citizens. The process by which citizens are made and remade. The output of democratic participation is the participant herself insofar as she leaves with a capacity to be a better civic actor.

Where the ancient model was mostly concerned with forming citizens, modern democracies are for the most part sprawling systems for converting preferences into governance. More participants produce more inputs to be aggregated, so in principle there is no upper bound on the size of the political community to which the procedure can apply. Modern institutions can hoover up preferences from millions through regular elections, but only at an extremely low level of resolution.

The basic claim of the aggregative family is that democracy is valuable because it generates good collective decisions, that the citizen is essentially an information source that produces preferences or opinions to be subordinated to the greater whole. Democratic procedures (like voting, deliberation, polling, or citizens’ assemblies) are all tools for processing those inputs into legitimate outputs.

Even within the aggregative tradition, some like John Stuart Mill argue that representative democracy can make citizens as well as decisions. For Mill, even occasional participation in public functions has a salutary effect on the citizen who takes part. “He is called upon,” Mill writes of the private citizen, “to weigh interests not his own; to be guided, in case of conflicting claims, by another rule than his private partialities; to apply, at every turn, principles and maxims which have for their reason of existence the general good.”

Analyses of this type acknowledge that formation is a laudable goal and hope that representative institutions could supply some version of it through juries, local office, and the modest acts of public engagement available to a private citizen. In practice, however, the settlement does not bear out these hopes. Over the century and a half since Mill’s reflections, the aggregative purposes of the system have steadily eclipsed its formative functions. Few of us would today say that the systems we live inside sharpen our faculty of self-governance or fellow-feeling.

Parallel to the aggregation tradition is the formation-first approach, best characterized by Alexis de Tocqueville in his travels across America in the 1830s. Tocqueville believed that the habits of self-rule in the small things prepared people to rule themselves in the great ones. The more citizens tested their judgments in the company of others, the more resilient their characters became. Democracy stood or fell on these everyday practices, for it was only through the rehearsal of judgment in everyday life that citizens could learn the art of governing themselves.

Tocqueville knew what happened when those habits were no longer present. In the France he grew up in, centuries of monarchical centralization had drawn the business of ordinary life away from local communities and into the hands of the state, so that by the time the revolution spilled into the streets, the citizens who were asked to govern themselves had long since lost the practice of doing so.

The provincial assemblies that had once given ordinary people a say in their own affairs had been largely abolished; the guilds and local associations that had trained people in the small arts of collective decision-making had been suppressed; and the aristocratic class that had once served, however imperfectly, as a counterweight to central power had been reduced to a decorative caste with titles but no responsibilities.

Tocqueville imagined how this kind of decay might unfold in a country that had skipped its aristocratic adolescence. One danger was that citizens might willingly hand over the work of self-government to a mild and far-seeing administrative state. He described a “tutelary power” that “does not tyrannize, but it compresses, enervates, extinguishes, and stupefies a people, till each nation is reduced to nothing better than a flock of timid and industrious animals, of which the government is the shepherd.”

Democratic deficit

Representative democracy faces two distinct challenges within the aggregative tradition. The first is a bandwidth problem, which reminds us that modern institutions can only collect citizens’ views in a heavily compressed form - like a vote or poll - that fails to capture most of what each citizen actually thinks. The second is the deliberative problem. Here the kind of refining and filtering Burke and Madison thought representative bodies were for has been overtaken by the volume of political business and the professionalization of representatives.

The AI for democracy field is directed at the first problem. It concerns capturing citizens’ views at higher fidelity and aggregating those inputs at scale. The deliberative deficit barely features beyond the comparatively narrow work of synthesizing information so that officials can deliberate with better inputs. The formative concern, which is about what civic life does to the citizens who take part in them, is not a significant part of the project.

A survey by Seth Lazar and Lorenzo Manuali connects the field’s work to four modes of usage:

• Summarization: Synthesizing information in a way that officials or citizens can act more effectively. One developed example is the Recursive Public project, which gathered public input on AI policy and used LLMs to summarize the resulting discussions and group similar arguments into clusters in semantic space.

• Aggregation: Generating new statements that the model predicts will attract broader support than any of the originals. The Generative Social Choice project, for example, took participants’ views on contested issues and prompted a model to propose a slate of statements that a majority of participants reported feeling “excellently” or “exceptionally” represented by.

• Representation: Predicting user preferences by standing in for her in a simulated polity. See, for example, the Augmented Democracy project, which seeks “to create personalized AI representatives to augment their ability to participate directly in many democratic decisions.”

• Facilitation: Engaging participants in a structured conversation whereby the model plays the role of a discussion leader to keep conversations on topic, ask follow-up questions, draw reluctant participants into the exchange, and produce summaries at intervals. See Rappler’s aiDialogue project.

All of the major labs have worked on projects that fall within these categories, from Anthropic’s Collective Constitutional AI process designed to crowdsource Claude’s constitution to Google DeepMind’s Habermas Machine that helps groups of individuals find consensus on political topics.

These tools and others like them make it possible to aggregate preferences more efficiently. They also provide the resulting decisions with a kind of legitimacy that any modern polity needs in order to function. The decisions that constitute political life must be made somehow, and a procedure that draws as fully as possible on its citizens is by most reasonable measures an improvement on the status quo.

Like the vast majority of AI and democracy projects, each of the examples described above are concerned with aggregation rather than formation. An accurate clustering of a thousand Americans’ opinions about a frontier model’s content policy does not give anyone the experience of having governed themselves. A more efficient deliberative platform that drafts consensus statements between strangers does not produce the habits of self-rule that the Athenian Assembly or the New England town meeting cultivated through consequential practice over the course of an ordinary life.

The outputs look superficially similar in that citizens have played a role, but they did so without the inward labor through which civic life makes citizens. A citizen who has practiced forming and revising her own views in conversation with her peers is a better judge of political questions than someone whose preferences have been collected and aggregated by a model. Rather than asking us to trade decision quality for civic virtue, the formative tradition suggests both are the joint product of the same practice.

An obvious response is to ask whether AI might be designed to support a formative account of civic life. The temptation should be resisted. Formation is constituted by the experience of being responsible for consequences alongside fellow citizens whose lives are bound up with yours. There is no AI system, however carefully designed, that can deliver that.

While not straightforwardly citizen-making themselves, there are uses of AI that could lower the cost of entry to formative practices. A tool that helps a citizen convene like-minded peers does not help us grow, but it does make the work of civic life easier to start. A system that translates a council planning document into something the resident affected by it can understand does not produce a citizen capable of self-rule, but it may lower the transaction cost of participation through which such a citizen might be made.

Formative years

Tocqueville thought that juries were slower than judges and more likely to err on questions of law. If the only purpose of the courtroom was to produce the right verdict, the jury would have been abolished long ago. “The jury teaches every man not to recoil before the responsibility of his own actions,” he wrote, “and impresses him with that manly confidence without which political virtue cannot exist.”

Today, we think about governance as the process by which decisions are taken on our behalf and implemented by specialists. This is in part a function of the highly complex nature of the modern world, which depends on large-scale coordination and deep technical specialization.

Even so, the decisions that constitute political life are taken at scales that far exceed the ceiling of the Tocquevillian model. The areas that remain to citizens at the local level are administrative bodies operating downstream of decisions taken elsewhere rather than deliberative bodies that make decisions and live with the results.

AI systems may well make political institutions more legitimate, more efficient, and more responsive to our collective wants. On its own terms, this is a good outcome if it helps democracies make better decisions according to the wishes of their people. But collecting and aggregating preferences is not how we become more capable agents who can better take part in civic life.