The People Picked to Represent You Probably Don’t

Picture a citizens' assembly convened to advise the government on climate policy. A retired teacher, a young nurse, a small business owner. Diverse faces, different backgrounds, different zip codes. Exactly the kind of mix the organizers were aiming for.

Unfortunately, none of them particularly care about climate change.

That uncomfortable problem sits at the heart of a new study from the University of Southern California. When researchers asked over a thousand UK residents to choose which strangers they felt would best represent them, they found that in a general citizens' assembly - one that addresses a broad range of issues rather than a single topic - political alignment was the single strongest predictor of who felt like a good representative, outweighing age, gender, ethnicity, and region combined. In a climate-specific assembly, concern about climate change dominated above everything else when political alignment was not shown. Demographics barely moved the needle.

The study's authors include Evi Micha, WiSE Gabilan Assistant Professor, and Yusuf Hakan Kalayci, a former PhD student, both in the Thomas Lord Department of Computer Science at the USC Viterbi School of Engineering and the USC Stevens School of Computing and AI.

A democratic experiment gone global

Citizens' assemblies have become one of the most fashionable tools in modern democracy. Ireland used one to break a decades-long deadlock on abortion law. France convened one on climate. In Ostbelgien, the German-speaking region of Belgium, the model has been made permanent, with a rotating panel enshrined in law that sits alongside parliament. Support for randomly selected panels has been documented across fifteen European countries, and tech companies including OpenAI have begun running similar democratic input programs to shape how their AI systems should behave.

The standard method for building these panels is borrowed from polling: recruit the right proportions of men and women, different age groups, ethnicities, and regions. The assumption is that demographic diversity produces representative diversity.

To test it, the team - working in partnership with the Sortition Foundation - showed each of more than a thousand UK participants pairs of fictional candidate profiles covering nine attributes: age, gender, ethnicity, region, disability status, education level, area type, political alignment, and climate concern. Participants chose which person they felt would better represent them. By randomizing which attributes matched or differed from each respondent, the researchers could isolate precisely how much each factor drove choices.

They also drew simulated panels from their survey respondents to test what happens when an important attribute is left out of the selection criteria entirely. "A panel can meet every demographic quota and still fail to reflect the range of views on an issue people care about," Kalayci said. "When we left climate concern out, balancing demographics and political affiliation gave us about the same balance on climate concern as simply drawing names out of a hat. Including climate concern level directly made a substantial difference."

Not a case for scrapping demographics

The findings do not mean demographics should be thrown out. When Micha's team trained machine learning models on the data, they performed best when all attributes were included, not just political alignment. Knowing someone's age, gender, and ethnicity still added real predictive power.

"We don't advertise giving up on gender or race and just going with political alignment," Micha said. "We really need to discover the correct attributes to take into account."

The study also surfaced a pointed finding about who is served by the current system. Women placed more weight on gender match than men. Young people aged 16 to 29 cared far more about age similarity than older cohorts.

"People who seem to be more marginalised in the current democratic system are the ones asking for more representation from someone who looks like them," Micha said. "That was something surprising that we didn't expect to see."

The survey's open-ended responses pointed beyond the fixed profile attributes. Some participants flagged work experience, financial circumstances, family responsibilities, and connections to their community. "Beyond political alignment, some people also cared about how a representative would listen, consider evidence, and engage with different views," Kalayci said. "These comments suggest that feeling represented can also depend on how someone approaches a discussion and the experiences they bring to it."

Teaching a computer what representation feels like

The team's longer-term vision is a continuously updated public platform where practitioners designing an assembly on any topic could first learn which attributes people actually use to judge representatives for that issue.

"We want to keep learning from people as topics evolve," Micha said. "AI regulation is a new topic that didn't exist years ago. When we want to create an assembly tailored to that, what are the attributes we want to take into account?"

The study, currently under peer review, is already collecting data in Australia. Whether the findings travel across political cultures remains to be seen. But the core challenge they raise is universal: who do people actually feel represents them? That question has been answered for decades with demographic spreadsheets. The data is starting to suggest the answer is more complicated, and more interesting, than that.

Published on October 26th, 2026

Last updated on October 26th, 2026