How Algorithmic Stratification Drafts Citizens' Assemblies to Match Census Demographics

Algorithmic stratification filters civic lottery responses by computing thousands of valid panel combinations and selecting the one that maximizes the minimum probability of any single volunteer being chosen. This mathematical correction is required because citizens do not volunteer at equal rates, meaning a purely random draw would create a skewed assembly.[1][4]

When governments convene citizens' assemblies to tackle complex policy issues, they rely on a two-stage lottery known as sortition. First, thousands of invitations are mailed to randomly selected addresses, and then a final panel of roughly 100 people is drawn from those who respond.[1][3]

"Ideally, a citizens' assembly acts as a microcosm of society," said Ariel Procaccia, Gordon McKay Professor of Computer Science at Harvard University. "Whether this goal is realized in practice, however, depends on exactly how assembly members are chosen."

The problem lies in the response rate. Data from recent civic lotteries, such as the Scottish Government's 2026 People's Panel, shows an average response rate of just 4.7 percent. More importantly, that small slice of volunteers is never a perfect cross-section of the public.[2]

Highly educated, wealthier, and older citizens consistently respond to civic invitations at much higher rates than younger, lower-income, or marginalized demographics. If organizers simply pulled 100 names out of a hat from the volunteer pool, the resulting assembly would look nothing like the actual census data.[1][4]

The Mathematics of Demographic Parity

To fix this distortion, organizations like the Sortition Foundation use open-source stratification algorithms. These algorithms apply strict demographic quotas, matching the final panel to the exact age, gender, geography, and income distribution of the broader population.[1][3]

Applying quotas to a skewed volunteer pool creates a mathematical paradox. If a demographic group is heavily overrepresented in the volunteer pool, an individual from that group must have their selection probability artificially lowered to prevent them from dominating the final assembly.[1][4]

Conversely, if a marginalized group provides barely enough volunteers to meet their census quota, the algorithm must select almost all of them. This means their individual probability of selection approaches 100 percent, actively overriding the raw volunteer numbers to enforce demographic reality.[1]

In 2021, researchers from Harvard University and Carnegie Mellon University published a landmark paper in Nature detailing the Leximin algorithm. This open-source code is now the global standard for selecting citizens' assemblies, used in more than 40 major panels worldwide.[1]

Computing the Fairest Possible Draw

The Leximin algorithm solves the fairness paradox by maximizing the minimum probability of selection. Instead of a simple greedy quota fill, the software computes a vast distribution of all possible 100-person panels that perfectly satisfy the census constraints.[1]

Once the algorithm maps out every valid panel combination, it evaluates the selection odds for every single volunteer across that entire distribution. It then selects the specific distribution that gives the most mathematically disadvantaged volunteer the highest possible chance of being included.[1]

This approach guarantees end-to-end fairness. While it cannot give everyone an equal chance—because the initial volunteer pool is unequal—it ensures that no individual is mathematically locked out of the process just because of their demographic intersections.[1][4]

The code powering these selections, such as the Panelot and StratifySelect systems, is entirely open-source. This transparency allows independent data scientists and civil society groups to audit the mathematics, ensuring the lottery is not being manipulated by the convening government.[3]

Engineering the Initial Invitation Pool

However, the algorithm cannot conjure volunteers out of thin air. If a specific demographic group provides zero responses to the initial mailout, the software cannot fulfill that census quota, and the stratification process fails.[4]

To prevent this, public administrators now engineer the first stage of the lottery. Rather than mailing invitations completely at random, they over-sample demographics known to have low response rates, ensuring the algorithm has enough raw material to work with.[2][4]

For the 2026 Scottish People's Panel, organizers sent 8,000 invitations. To guarantee sufficient participation from lower-income citizens, 20 percent of those letters were specifically targeted at addresses within the most deprived postcodes in the country.[2]

This targeted mailing strategy provides the stratification algorithm with a viable pool of marginalized volunteers. Once the responses are collected, the software takes over, filtering the skewed enthusiasm into a perfectly balanced microcosm of the public.[2][4]

The Future of Algorithmic Representation

The use of algorithmic sortition is expanding rapidly. When the Nuffield Council on Bioethics convened a Citizens' Jury on assisted dying in England, they used these exact algorithms to recruit 34 jury members, creating a mini-public broadly representative of the national population.

By removing human bias from the final selection, algorithmic stratification protects the legitimacy of the assembly. Participants arrive knowing they were chosen by a verifiable, mathematically fair process rather than by political appointment or self-selection.[1][4]

As open-source sortition tools become more sophisticated, they are beginning to incorporate intersectional diversity metrics. Newer algorithms ensure that panels are not just balanced by isolated categories like age or gender, but by complex intersections like young, rural women.[1][3]

Ultimately, the success of a citizens' assembly rests on its ability to genuinely reflect the public. Algorithmic stratification provides the mathematical scaffolding required to turn a messy, unequal volunteer pool into a precise demographic mirror.[4]

The computational complexity of this task is significant. Finding a single valid panel that meets a dozen intersecting demographic quotas is difficult; computing the optimal distribution across thousands of valid panels requires advanced linear programming techniques.[1]

Yet, this complexity is exactly what makes the process robust. By relying on peer-reviewed, open-source algorithms, democratic innovators can prove that the final assembly is the fairest possible representation of the community, free from electoral gerrymandering or partisan influence.[4]

For citizens participating in these panels, the algorithmic selection is largely invisible. They simply receive a letter, register their interest online or via a freephone service, and wait to see if the lottery selects their name.[2][4]

Behind the scenes, however, that lottery is a triumph of civic technology. It demonstrates how advanced mathematics and open-source software can be harnessed to strengthen democratic institutions, ensuring that every voice has a mathematically protected chance to be heard.[4]