Simulating democracy

Two of the most persuasive books about voters argue that the voter this chapter needs barely exists. In Democracy for Realists, Christopher Achen and Larry Bartels take the textbook story — informed citizens evaluate policies, choose the candidate whose platform fits their preferences, and elections translate the public's will into government — and spend a book dismantling it as the "folk theory" of democracy .^[1] In their telling, people vote their social identities and group loyalties, roughly the way they inherit a religion. They punish incumbents for droughts. They reward economic growth that began before the incumbent took office. The rational, policy-weighing citizen of the civics textbook bears little resemblance to how democracy runs.

Bryan Caplan pushed past ignorance to something worse. The Myth of the Rational Voter set the public's economic views beside professional economists' and found the gaps were not random .^[2] Voters lean anti-market, anti-foreign, toward make-work, and toward pessimism — four systematic distortions. Random errors wash out in a large electorate; directional ones compound, and they push democratic outcomes away from what most economists would call welfare-improving policy.

I take the skeptics seriously enough to start from their side of the ledger. If identity and loyalty carry most of the vote, anything an information tool can touch is at best a minority share of the decision. So the claim in this chapter is deliberately modest. It is not that information fixes democracy; it cannot. It is that to whatever degree some slice of the vote responds to perceived policy impacts, the quality of those perceptions matters, and the quality is improvable. Even if only a fifth of the vote is policy-driven and the rest is identity all the way down, cleaning up that fifth still moves outcomes. The whole argument lives inside the fifth.

The perception problem

The last chapter estimated what people think: a distribution of opinion, estimated and benchmarked. A distribution is only an input. What a democracy does with it is aggregation — millions of noisy, partial pictures of the world compressed into one binding decision — and this chapter is about what the noise does on the way in.

Consider a voter I am inventing on purpose. Sarah teaches school, has two kids, and faces a choice between two candidates: one would expand the Child Tax Credit, the other would repeal her state's income tax. Which leaves her family better off? Without doing the arithmetic she has to guess. The state income tax stings visibly on every pay stub. She remembers the expanded pandemic-era credit helping. The ads on both sides are built to persuade her. Her eventual vote blends what she values, what she believes each policy would do, what her party and her neighbors support, and noise — a candidate's charisma, the morning's headlines. Even if her values are perfectly clear, her vote may fail to track them, because she cannot see the true impacts. She votes on a noisy signal.

Multiply Sarah by a national electorate and the aggregate inherits her noise: the outcome only loosely tracks what voters would have chosen with clear sight. Achen and Bartels would call even this too generous, since most votes never involve policy evaluation at all. Caplan would add that the voters who do evaluate lean predictably wrong. Grant both. The narrower question stands: where a signal exists, does its quality matter?

A toy, on purpose

To reason about that, I built a small simulation called Democrasim. The emphasis belongs on small. It is not a validated production system like PolicyEngine, and not a benchmarked estimator like the opinion work of the last chapter. It makes simplifying assumptions political scientists would rightly attack, and nothing it produces is a finding about the world. It is a thought experiment in code, built to make one relationship concrete enough to reason about, and I present it as exactly that.

Each simulated voter carries three things. First, a set of weighted preferences across policy dimensions — how much they care about the economy versus the environment versus social issues: their true values. Second, an accuracy level, which sets how close their perception of a policy's effects sits to the truth. Third, a bias — a systematic lean beyond random noise, such as consistently overstating a tax burden or understating an environmental cost. The whole model rests on one line: perceived impact equals true impact plus noise plus bias, with the noise shrinking as accuracy rises. Voters compare candidates on perceived impacts weighted by their preferences, and the majority wins. Then you vary the accuracy and watch what happens to the welfare quality of the winner.

To see the mechanism, give the toy an election worth arguing about, with round numbers invented for the purpose. Candidate A expands the child credit and pays for it with a small rise in the top marginal rate. Candidate B repeals the state income tax and pays for it by cutting food assistance. For a middle-income family with children and no food assistance, say the credit expansion is worth $1,500 a year and the tax repeal $700, while the offsetting rate increase never reaches them: they do better under A, even though "repeal the income tax" sounds like the bigger deal. An accurate perceiver picks A. A noisy one may pick B, against her own family's interest, on a vague sense of which candidate is the tax-cutter. Scale the comparison across an income distribution and it develops a shape — lower-income families do far better under A, upper-income families modestly better under B — and whether the election reflects that shape depends on how much of the vote is signal rather than framing.

What the runs show, in the only sense a toy can show anything: with accuracy high, outcomes track welfare, and the candidate whose policies would actually improve lives tends to win. With accuracy low, the link frays; good and bad platforms win indistinguishably, because the signal is too corrupted to sort them. The relationship behaves like a threshold — below a certain accuracy, elections become essentially random with respect to welfare; above it, the coupling strengthens fast. Bias bends the outcome differently than noise does. A population that systematically overvalues tax cuts relative to equivalent benefit increases elects tax-cutters regardless of the welfare arithmetic: Caplan's argument reproduced in miniature, with the reminder that it operates through perception, upstream of any voting rule.

I want to be careful not to overclaim. Democrasim's "true" impacts are numbers the model invented, so it can demonstrate a dynamic but never measure the world. Its voters maximize welfare through policy evaluation — the very folk theory Achen and Bartels took apart — and it ignores identity, loyalty, and nearly everything else that actually decides elections. What survives the caveats is one relationship: to the degree that any of the vote responds to perceived impacts, the quality of the perceptions governs whether elections can reward good policy at all.

Read backwards, the toy explains the book. If outcomes track welfare only where perceptions of policy impacts are accurate, then machinery that makes accurate perception cheap — rules encoded exactly, populations calibrated honestly, forecasts graded in public — stops being a convenience for analysts and becomes an input to whether self-government works.

The stack exists to raise the accuracy term in this model.

What the evidence shows

The real evidence on information and voting is mixed, and worth taking straight. Information moves votes when it reaches people. In Brazil, Claudio Ferraz and Frederico Finan studied federal corruption audits whose results were released before municipal elections: where auditors found corruption, voters punished the incumbents — but the effect concentrated in municipalities with local radio stations to carry the results .^[3] Information that existed but did not travel changed nothing. American nonpartisan voter guides sit at the weaker end of the same spectrum: they tend to raise political knowledge without much moving vote choice citation pending, because in a polarized electorate most voters have committed long before the guide arrives. Arthur Lupia and Mathew McCubbins showed that voters run on trusted-source shortcuts rather than full information ,^[4] which cuts both ways — a good shortcut approximates an informed choice; a bad one entrenches identity over evaluation. And accurate information never competes on equal terms, because a well-crafted misleading claim travels farther and sticks longer than a dry correction. Democrasim treats noise as random. In the world, some of the noise is engineered by people with an interest in the distortion, which means the toy understates the problem.

Kansas ran the expensive version of the experiment. In 2012 the state enacted large income-tax cuts that the governor billed as a "real live experiment" in supply-side growth .^[5] Independent analysts projected large revenue shortfalls instead. The shortfalls came — years of budget crises and school-funding cuts — until June 2017, when a bipartisan legislative supermajority repealed the centerpiece cuts over the governor's veto. The experiment ran five years, and the tuition was paid by the citizens of Kansas in billions of dollars of lost revenue and degraded services. Better information infrastructure would not have blocked the enactment; the projections existed and lost the argument. What it might have done is shorten the loop.

The Affordable Care Act showed how long the loop can stay open. Chapter 3 told the forecasting side: the CBO got total coverage gains roughly right while badly missing the composition, overestimating the exchanges and underestimating Medicaid. Voters who opposed the law over predicted exchange disruption and voters who supported it for its Medicaid reach both experienced something different from what they had been told, and the gap between perception and reality persisted for years. In February 2017, seven years after passage, a Morning Consult survey found that 35 percent of American adults did not know the Affordable Care Act and "Obamacare" were the same law .^[6] No calculator repairs that. It measures how far real voter knowledge sits from the idealized version, and it is a reminder that much of the perception problem is transmission, not computation.

So the honest summary: better voter information is necessary and nowhere near sufficient. It is one lever among identity, institutions, and strategy — and the only one of the four that infrastructure can move directly, rather than through the generational work of changing culture.

Limits no information repairs

Some constraints sit deeper than perception. Kenneth Arrow proved in 1951 that no ranked voting system can satisfy a short list of reasonable fairness conditions at once ;^[7] Gibbard and Satterthwaite later showed that any non-dictatorial voting system can be gamed by strategic voting. citation pending Better inputs repeal neither theorem. But the theorems constrain the aggregation mechanism, not the quality of what feeds it. An imperfect aggregator still delivers better outcomes from better-informed inputs; Arrow ruled out a perfect system, nothing more. If anything, the theorems sharpen the chapter's point: since the mechanism cannot be perfected, the inputs are where the improvement lives.

Three objections deserve direct answers. Maybe the preferences are the problem, not the perceptions. True, and orthogonal: voters who want harmful things and perceive accurately will get harmful things, which is a different failure from wanting good things and misperceiving. Better perception at least delivers people what they actually want, the precondition for holding them responsible for wanting it.

Maybe information won't reach the disengaged. Partly true: a non-voter will not open a policy calculator, and the citizens who would are already the relatively informed. But the interface is moving to where people already are. An AI assistant that volunteers a policy's household impact reaches past the self-selected, and upgrading the moderately informed still improves the signal.

Maybe calculated self-interest isn't citizenship. Fair, and the same engine computes poverty rates, inequality, and total budgetary cost as readily as one household's bottom line. The point is replacing perception with calculation, whatever a voter chooses to weigh.

The most radical proposal in this territory takes the split between wanting and knowing all the way to the constitution. Robin Hanson's futarchy would have democracies vote on values but bet on beliefs :^[8] elected representatives define what the country cares about — a national welfare metric, say a child-poverty rate — and prediction markets decide which policies achieve it. A legislature proposes a bill; markets price the welfare metric conditional on passage and on failure; if the market prices welfare higher with the bill than without it, the bill becomes law .^[9] No polity has adopted futarchy, and the reasons are visible from here. Deep pockets can lean on thin markets. Most policy questions never attract enough trading to price anything reliably. "Welfare" is exactly the contested term the mechanism pretends is settled. And democratic legitimacy depends partly on citizens feeling heard, which a price does not supply. But as a thought experiment, futarchy isolates the distinction this book is built on — values are the outcomes we want; beliefs are claims about which policies produce them. The first kind belongs to voters. The second kind can be checked.

Values are for humans; facts are for tools.

PolicyEngine works the beliefs side of that line and never crosses it. In Democrasim's vocabulary, it is an accuracy multiplier for whatever slice of a vote is policy-driven: it moves Sarah from "I sense this would hurt me" to "this changes my household's income by about this much," computed on her actual circumstances, free to anyone with a browser. The warrant runs deeper than "read the source code if you like" — the rules it applies are open encodings checked against reference calculators, chapter 10's machinery. And its limits are the ones this chapter has drawn. It does nothing for the identity-driven share of her vote, does not loosen her party's hold on her, does not repeal Arrow. It removes one source of noise from one channel. That is a smaller claim than "informed voters fix democracy," and it is the right size.

From toy to scoreboard

Democrasim can only toy with the perception problem, because its ground truth is invented: its voters misperceive effects that are themselves numbers the model made up. The version that terminates in something real belongs to chapter 13. Forecast a government metric — under the law as it stands, or conditional on a policy taking effect, like Medicaid call-center waits if a work-requirement deadline slips — publish the interval, and let the official number grade the forecast when it prints .^[10] That is the perception problem with a scoreboard attached: instead of asking whether voters guess policy impacts well, it asks whether anyone can state a policy-relevant number in advance and stand still while reality checks it. As I write, nothing on that docket has resolved, and its conditional entries will only ever resolve on the branch the world takes. The scoreboard has no scores yet; what it changes is the form of the question. A thought experiment is graded against numbers it invented. A forecast is graded by numbers nobody chose.

That reframes what all of this machinery is for. A tax calculator looks like a tool for people who want to trim their taxes. Built open, checked against the law, and pointed at an election, it becomes one component of informed self-government — an input to the accuracy term, offered to whatever share of the vote will take it. Democracy decides what we want; the tools in this book only sharpen the picture of what we will get. Which leaves the question the split saves for last. Facts are for tools and values are for humans — but values move, and the next chapter asks whether a simulation can see them moving before we can.