AI voter personas shifted an election estimate toward Democrats—but accuracy is unclear

A new study found that richer artificial voter profiles changed simulated House-election responses in a plausible direction, while producing a larger Democratic margin than the polling average.

Abstract editorial illustration accompanying the article: AI voter personas shifted an election estimate toward Democrats—but accuracy is unclear

Editorial illustration — not from the study.

The study presents a framework for simulating pre-election survey answers with large language models. It gives each artificial respondent a biography, memories of important events and a pattern of news exposure, then models how that respondent would answer questions about turnout and candidate choice.

Applied to one weekly wave of the 2026 US House elections, the framework produced national and district estimates. Compared with profiles based only on demographics, the richer profiles shifted results toward Democrats, matching the direction of contemporary polling. Profiles that constructed judgments from assigned news produced a larger Democratic margin than profiles that did not receive news, although the estimates had substantial uncertainty.

What the artificial voters did

The researchers compared a demographics-only group of artificial respondents with personae enriched using synthetic biographies, memories of salient events and conditional news exposure. The enriched personae shifted simulated election responses toward Democrats. Their estimated Democratic margin was 18.7 percentage points, compared with 8.7 points in the polling average. Personae assigned to construct judgments from news produced a larger Democratic margin than those that retrieved no news. The study did not determine which individual elements caused the changes.

Why the comparison matters

The work offers a structured way to study how artificial respondents answer survey questions, rather than treating large language models as simple stand-ins for people. It also shows why plausible-looking shifts are not enough to establish that synthetic respondents can forecast elections: in this test, the estimated Democratic margin was substantially larger than the polling average, and the researchers did not establish predictive accuracy conclusively.

Evidence and caveats

The evidence comes from a computational comparison using one weekly wave of the 2026 US House elections, with national and district estimates generated through multilevel regression and post-stratification. The study compares different types of artificial personae, not actual respondents in a randomized survey. It does not isolate the effects of every enrichment component, and uncertainty was substantial for some comparisons. The authors identify continuing concerns about leftward bias, calibrating uncertainty and benchmarking the approach during an election campaign.

Tuna and billfish fishing moved offshore mainly in the past 200 years

People have fished for tuna for about 44,000 years and billfish for about 10,000 years, according to a global review of archaeological evidence. But sustained, intensive fishing in offshore and oceanic waters appears to have developed mainly during the past 200 years.

6 hours ago·4 views

Abstract editorial illustration accompanying the article: Coastal residents weigh flood damage, costs and family ties when considering a move

Society & Economics

Coastal residents weigh flood damage, costs and family ties when considering a move

People considering whether to leave flood-prone coastal communities weigh disastrous flooding and rising costs alongside crime, healthcare and proximity to family and friends. In a survey of 182 respondents, most said they would not relocate, while others said they would move immediately or in the future.

14 hours ago·6 views

Society & Economics

Managers accepted reflective coaching from humans but wanted answers from AI

A study of eight managers found that they preferred human coaches for reflective questioning, even though they welcomed that approach from a person. When the same coaching protocol came from an AI system, they wanted it to provide answers instead.

21 hours ago·16 views