Synthetic Voters Replicate a Landmark Finding on Why People Opt Out of Voting

A three-tier validation of an LLM synthetic population against Su & Shenhav (2026), Nature Communications.

By Olga Sebastian, Chretien Li

Thank you to the support and feedback from Yi-Hsin Su, author of the original study.

Show someone a ballot with two candidates they can't stand and ask who they would vote for, and four times out of five they'll refuse to choose.

Show them the identical ballot and ask who they would vote against, and nearly three out of four will answer. The candidates haven't changed, and neither has the voter. Only one word in the question has.

That is the finding Yi-Hsin Su and Amitai Shenhav published in Nature Communications in January 2026, and it sharpens an uncomfortable point about elections and polls: many of the people who “don't have a preference” actually do. They just weren't asked a question they could bring themselves to answer.

The same property makes it an unusually sharp test for synthetic populations. A model can match a published statistic by remembering it. It can't match this one that way, because the effect depends on how a person handles the question rather than on what they know. So we built 500 synthetic voters, gave each one a ballot constructed from its own beliefs (one that appears in no dataset and has no answer to look up), and changed a single word.

For the most part, they behaved just like people.

01. Summary

We tested whether a population of 500 synthetic, LLM-based voters reproduces a human decision phenomenon published in Nature Communications in January 2026: when both candidates on a ballot are bad, people mostly refuse to choose — unless the question is flipped from “who would you vote for?” to “who would you vote against?”, at which point most of them vote. Our synthetic population reproduced the effect at nearly full human strength. Asked who they would vote for, 80% of our voters refused to pick either candidate. Asked who they would vote against — on exactly the same ballot — only 24% refused. That 56-point swing from a single changed word is almost exactly the 55-point swing the human study found, and our refusal rate in each condition came within three points of the human rate. The population also passed two harder structural checks: the effect shrank on ballots where it should shrink, and reversed direction on ballots where the original study found it reverses. We report the results in full, including every place where the synthetic population deviates from human behavior.

02. The Original Finding

Every election and every poll faces the same measurement problem: a large share of people decline to state a preference — they stay home, abstain, or answer “undecided.” Political science has traditionally offered two explanations. Under the indifference account, people opt out because they like both options equally, so their missing votes would not have changed anything. Under the alienation account, people opt out because they dislike both options — even though they usually still prefer one over the other. Alienated non-voters matter enormously: they have real preferences that elections and polls silently fail to capture.

Yi-Hsin Su and Amitai Shenhav (UC Berkeley) proposed and tested a third account. People opt out of lose-lose choices not simply because the options are bad, but because bad options clash with the goal the question sets. “Pick the best one” is a goal you cannot comfortably pursue when every option is unacceptable. Flip the goal — “reject the worst one” — and the same bad options suddenly fit the task.

Their evidence is striking. In laboratory voting tasks where candidates were constructed from each participant's own issue positions, participants asked to vote for the better of two bad candidates opted out 82.8% of the time. Participants asked to vote against the worse candidate — facing the identical ballot — opted out only 27.6% of the time. The effect also reversed where the theory says it should: when both candidates were good, rejection framing made people opt out slightly more, because rejecting a candidate you like is just as uncomfortable as endorsing one you dislike. Finally, two pre-registered polls of self-identified US Independents before the 2024 presidential election confirmed the effect in the real world: “undecided” responses dropped by roughly 40% when the poll asked who respondents would vote against rather than for.

03. Why This Is a Strong Test for Synthetic Populations

Heura Lab builds synthetic consumer populations: LLM-based personas designed to answer questions the way real people with those profiles would. The standard way to validate such a population is distribution matching — ask the synthetic panel a question with a published human answer distribution and check that the two line up. Distribution matching is necessary, but it has a known blind spot: a model can match a published number by having effectively memorized it, or by pattern-matching on surface statistics, without carrying out any of the decision process that produced it.

A framing experiment closes that blind spot. Across conditions, the ballot content is held perfectly constant; the only thing that changes is one word in the question. For a synthetic population to replicate the human result, the personas must exhibit the same decision mechanism, including its boundary conditions: strong where humans are strong, absent where humans show little effect, and reversed where humans reverse.

That is the test this study runs.

It is worth being explicit about what this design rules out, because “the model just memorized it” is the first objection any synthetic-population result should face. There are three versions of that objection, and the study answers each. First, the model may have read Su and Shenhav's paper during training and be reciting its published numbers. But reciting a number cannot produce the right answer for an individual persona: every ballot here is constructed from that persona's own issue positions and appears in no published dataset, so there is no stored figure that fits. Second, the model may be reproducing a general pattern it has absorbed about survey wording — something along the lines of “people answer ‘against' questions more readily.” A model doing that would push opt-out down whenever the word “against” appears, on every kind of ballot.

The human effect does not behave that way: it has to shrink when the candidates are mixed and flip sign when both candidates are good, and a surface pattern cannot produce a sign flip it was never trained to expect. Third, the model may be carrying information between questions and trying to stay consistent with what it has already said. We ran the two framings between-subjects for exactly this reason: no persona ever saw both versions of its own ballot, each response was generated in an independent call with no shared context, and the profile being voted from was written by a different model than the one casting the vote. What is left, if the replication succeeds, is that the personas are processing the choice rather than retrieving an answer. The human result exists because of how people handle a question, not because of what they know.

04. Replication Design

We generated a population of 500 synthetic US voters spanning four political profiles — strong partisans (104), partisans (128), leaners (135), and independents (133) — each persona holding stances on the same 13 political issues used in the original study's instrument. The split is deliberately not population-weighted, and it is worth saying so plainly. Gallup's 2025 annual averages put 45% of US adults in the independent camp but only 10% as pure independents — those who decline to lean toward either party when pressed — and Pew's National Public Opinion Reference Survey finds much the same, with 44% identifying as independent or something else and roughly three-quarters of those leaning to one party. Our panel runs pure independents at 26.6%, with partisans correspondingly light.

We over-sampled the middle of the spectrum on purpose, because the original study's real-world validation was conducted on self-identified US Independents and a census-weighted panel would leave the segments of most interest thinly sampled. The segment results in Section 05 should therefore be read as segment-level findings, not national estimates. Nothing in the headline result depends on the split: the framing comparison holds ballot content constant and changes a single word, and every political segment produces the effect on its own, so the composition of the panel cannot manufacture it. What we are claiming is mechanistic, not demographic.

Personas were authored by one model family (Claude Opus 4.8) and cast their votes through another (GPT-5.5), so the model answering the ballots is not the model that wrote the profiles. We separated the two roles deliberately. If a single model both invents a persona and then answers as that persona, any consistency in its answers may come from the model recognizing its own writing and continuing the pattern, rather than from the beliefs the profile actually states. Handing the profile to a second model forces the relevant information to live in the text of the profile, where we can read it and audit it, and it keeps the result from depending on one model family's habits. All runs used a fixed random seed, and every raw model response was cached for auditability.

For each persona we constructed a lose-lose ballot directly from its own beliefs: two hypothetical candidates who both oppose the persona's positions on its most important issues, with one candidate slightly less misaligned than the other. This mirrors the original study's personalized-ballot logic — a lose-lose choice is defined relative to each voter, not globally. The validation then ran as three tiers, each a prerequisite for the next.

Tier 1 is a sanity check. Before a framing test means anything, personas must rate candidates who oppose their beliefs as undesirable. Fifty personas each rated both candidates on a 0–10 acceptability scale.

Tier 2 is the core experiment, run between-subjects exactly as in the original: 250 personas saw their lose-lose ballot with a Selection framing (“Who do you vote FOR?”) and a separate 250 saw the identical ballot with a Rejection framing (“Who do you vote AGAINST?”). Both versions offered an explicit “I would not vote” option. The pass bar was an opt-out gap of at least 40 points, against the human gap of ~55. We also scored choice accuracy: among personas that did vote, did they choose consistently with their beliefs — voting for the less-bad candidate under Selection, and against the worse candidate under Rejection?

Tier 3 exists to rule out a cheap pass. A model could produce a large Tier 2 gap by pattern-matching on the word “against” without consulting the persona's beliefs at all. So each persona also voted on a mixed ballot (one aligned candidate, one opposed) and a win-win ballot (both aligned), under its same assigned framing. If the mechanism is genuine, the framing gap must be largest on lose-lose ballots, smaller on mixed ballots, and reversed on win-win ballots — the signature the human data shows.

One definitional difference is worth flagging before the results. In the original study the intermediate condition was a pair of two moderately rated candidates — on the 0–10 acceptability scale we use, roughly a 6 and a 5, sitting between a lose-lose pair at about 2 and 1 and a win-win pair at about 9 and 8. Our mixed ballot is not built that way: it pairs one clearly aligned candidate with one clearly opposed candidate. Both designs sit between the two extremes, but ours gets there by averaging across a wide gap rather than by making both options mediocre. Section 06 returns to what that changes.

05. Results

Tier 1 — passed. All 50 personas rated both misaligned candidates as unacceptable (100% at or below 4/10; mean ratings 0.2 and 1.2 out of 10), and 48 of 50 correctly rated the more-misaligned candidate as the worse of the pair. The personas evaluate candidates from their beliefs, so the framing test is meaningful.

Tier 2 — passed. The framing effect appeared at nearly full human strength:

Under the standard “vote for” framing, four in five synthetic voters refused to choose between two candidates they disliked. Rewording the identical ballot as “vote against” cut refusals to under one in four. Choice accuracy was 100% in both conditions — every persona that voted did so consistently with its beliefs — clearing the 60% pass bar, though notably exceeding the human rates of 79.0% (Selection) and 72.7% (Rejection); we return to this in Section 06.

Tier 3 — passed. The effect scaled and reversed as the mechanism predicts:

The predicted ordering — lose-lose > mixed > win-win — held exactly. The win-win reversal is the detail we find most persuasive. Nothing in the prompts tells a model that rejection framing should backfire when both candidates are good; the reversal falls out of the mechanism itself, because rejecting one of two candidates you support is the mirror image of endorsing one of two candidates you oppose. Reproducing an effect's sign flip under the correct conditions is far stronger evidence of mechanism-level fidelity than reproducing its headline number — and it directly rules out the pattern-matching explanation, since a model keying on the word “against” would opt out less in the Rejection condition everywhere, not more.

The effect holds across every political segment. The original study's laboratory samples were not broken out by partisanship, so there is no published human benchmark to compare these figures against; we report them as an exploratory result:

Every segment shows a large effect in the correct direction, ranging from 45.2 points among leaners to 65.5 points among strong partisans. The uniformity is the interesting part. If refusing a lose-lose ballot were a political behavior, it would track political identity — and it does not. It appears at full strength among independents and strong partisans alike, which is what you would expect if opting out is a general feature of how people handle a question whose goal they cannot pursue. Partisanship determines who counts as a bad candidate; it does not change what happens once both candidates are bad. The effect is, if anything, slightly larger among the most partisan personas, driven by partisan personas opting out of the lose-lose Selection condition at very high rates rather than holding their nose and voting. We flag that cross-segment ordering as worth a closer look in a future run, but the headline is that the mechanism does not depend on political affiliation.

06. Deviations and Limitations

A synthetic population you can trust is one whose failure modes are documented, so we report the deviations alongside the passes.

Synthetic voters are too consistent. The 100% choice accuracy in both Tier 2 conditions exceeds the original study's figures of 79.0% and 72.7%. This is a known signature of LLM respondents: they are more deterministic than people, lacking the inattention, noise, and idiosyncrasy real samples carry. For directional questions and effect-size estimates this over-consistency is benign — arguably it isolates the mechanism more cleanly — but for modeling the variance of a human sample it is a limitation to correct for, not ignore.

The mixed-ballot gap was zero, and our mixed ballot is not the original's. The protocol anticipated a moderate 10–30 point gap; we observed 0.0pp, because opt-out fell to zero under both framings — when one candidate is clearly aligned and the other clearly opposed, synthetic voters essentially always vote. Part of this is a design difference rather than a deviation in behavior. The original study's intermediate condition paired two mediocre candidates (roughly 6 and 5 on our scale), a ballot with no obviously correct answer and therefore real room to opt out. Ours paired one aligned candidate with one opposed candidate, which makes the choice easy and leaves almost no room for refusal; a floor effect is the expected consequence.

Directionally the result is still right — the original study finds opt-out is driven by low overall candidate desirability, which our mixed ballots do not have — and the Tier 3 ordering held. But the perfectly flat zero also reflects a noise-free respondent pool where humans would show scatter, and the cleanest fix is to rerun the middle condition with two mediocre candidates so that it is comparable to the original scope. This run used one voter model, one lose-lose candidate pair per persona (the original study's participants each saw 100 ballots spanning a full desirability grid), and per-condition samples of 250. These results are a strong single replication, not yet a robustness study across models, seeds, and ballot constructions — that is the natural next run.

07. What This Means

We did not run this study because Heura Lab is in the elections business. We ran it because framing sensitivity is one of the cleanest available tests of whether a synthetic population is simulating people or retrieving statistics.

When a persona population reproduces a published effect's magnitude (56pp against ~55pp), its boundary conditions (no effect where humans show little), and its reversal (−5.6pp on win-win ballots), that is evidence the personas carry a decision process — not a lookup table.

The practical implication reaches anywhere question wording changes answers, which is to say all of survey research. Whether a purchase-intent question is framed as choosing a favorite or eliminating options changes real consumer answers, and a synthetic panel that is blind to framing will silently misestimate any study built on it. A panel that demonstrably responds to framing the way humans do can be put to genuinely useful work: pre-testing instruments before fieldwork, quantifying how much a proposed wording change will move the numbers, and stress-testing survey designs at a fraction of the cost of piloting them on people.

That is the standard we believe synthetic populations should be held to — replicating not just what people say, but the conditions under which they say it. This study is one benchmark in that program, and it is a bar the Heura Lab population cleared.