Democracy's Experiments: Citizens' Assemblies, Referendums, Participatory Budgets and Digital Democracy

Follows: Democracy: The Many Flavors · How Voting Systems Shape Outcomes

Democracy: The Many Flavors sorted democracies into direct, representative and deliberative kinds, and into presidential and parliamentary designs. How Voting Systems Shape Outcomes showed that the same ballots, counted under different rules, can produce different winners. Both pieces took one thing for granted: that the core of democracy is ordinary people choosing professional politicians, who then decide.

This piece is about the alternatives. Some are very old. In classical Athens, most officials were picked by lottery and election was considered the less democratic method. Others are new: a citizens' assembly in Dublin, neighbourhood budget meetings in a Brazilian city, a Swiss constitution that lets 100,000 signatures force a national vote, a Taiwanese website that shows people where they agree. Each is an answer to a different worry about elected government. One is that politicians don't look like the people they serve. Another is that voters rarely get to think a question through, or that the poor are outvoted on what affects them most. A fourth is that crowds online can only shout.

None of these experiments has replaced elections anywhere, and their supporters disagree about whether one should. The aim here is to describe each one: how it works, what happened, what the best studies found, and what its serious critics say. Part 2 is a working simulator, because choosing people by lot rests on mathematics anyone can test.

Part 1 — Athens: when the lottery was the democracy

The most detailed description of Athenian government that survives is the Constitution of Athens, a treatise attributed to Aristotle. Its later parts describe the system as it worked in the philosopher's own time, the fourth century BCE, and the word that keeps recurring is "lot".

"All the magistrates that are concerned with the ordinary routine of administration are elected by lot," it says, with a handful of exceptions: the Military Treasurer, the officials in charge of the festival fund, the Superintendent of Springs, and "all military officers", who were "elected by vote" (Constitution of Athens, Part 43, Kenyon translation). The Council of Five Hundred, which prepared business for the citizen Assembly, was "elected by lot, fifty from each tribe". Juries were drawn by lot using an elaborate apparatus of tickets, chests, staves and vases, from among citizens over thirty (Part 63). And no one could hold most civilian offices twice: "none of the others more than once, except the membership of the Council, which may be held twice" (Part 62).

Together those rules had a distinctive logic. The big decisions were taken by the Assembly, which any citizen could attend. The work of preparing and carrying out those decisions was done by people picked at random and rotated quickly, so that a large share of citizens would at some point serve. Election was kept for jobs where skill plainly mattered, above all commanding armies and fleets.

Aristotle drew the theoretical conclusion in the Politics. Discussing how to blend democracy and oligarchy into a stable "free state", he notes that "it seems correspondent to the nature of a democracy, that the magistrates should be chosen by lot, but an aristocracy by vote" (Politics, Book IV, ch. 9, Ellis translation).

"It seems correspondent to the nature of a democracy, that the magistrates should be chosen by lot, but an aristocracy by vote." — Aristotle, Politics IV.9

To a modern reader that sounds backwards, since we treat elections as the very definition of democracy. The Greek reasoning was that an election is a contest of reputation. People vote for those who are already prominent, wealthy, well connected or good at speaking. Over time that produces rule by the "best" (aristoi), or at least by those who can persuade others they are. A lottery gives every eligible person the same chance. The Stanford Encyclopedia of Philosophy's entry on democracy puts the modern version of the point this way: under sortition "citizens have an equal chance of being selected", so there is "equality at a crucial stage of the decision-making process".

Two limits keep this from being a fairy tale. First, the eligible pool was small. Citizenship was "open to all who are of citizen birth by both parents", enrolled at eighteen after the demesmen checked that each candidate was "free born" (Part 42), and full citizenship in Greek city-states excluded women, slaves and foreigners (SEP, "Aristotle's Political Theory"). Athenian sortition was equality among a minority of the people who lived there. Second, the lot was never used alone. It sat inside a system that also had an open assembly, elected generals, public audits of officials when they left office, and courts that could punish them.

After Athens the lottery didn't vanish, but it survived in a supporting role. The Republic of Venice, for example, chose its doge through a famously elaborate sequence of rounds that mixed drawing lots with voting, a system economists still study for how it combined chance with supermajority rule (Molinari, "How the Republic of Venice Chose Its Doge"). Courts in many countries still draw juries at random. But when the modern representative republics were built in the eighteenth century, they chose elections, and the lot largely dropped out of political practice.

Its revival began in the late twentieth century with a sharper version of the Athenian question. If you want a body that thinks like the public, not like professional politicians, why not choose it the way a pollster chooses a sample? First, though, it's worth seeing exactly what a lottery can and can't do.

Part 2 — The simulator: can a lottery mirror a people?

A citizens' assembly claims to be a miniature of the public. That claim rests on the same mathematics as an opinion poll: if you draw people at random, the sample's make-up will usually be close to the population's, and the margin of error shrinks as the sample grows. But real assemblies can't simply conscript people. They send invitations to randomly chosen households, and only some of those invited say yes. The rest of this part lets you see what that changes.

The simulator below uses an invented town of 30,000 adults. Each one has six traits: age group, gender, urban or rural home, whether they have a university degree, whether they're already active in civic life (they go to meetings and join groups), and whether they support a fictional proposal to turn the central car park into a park. In this town, support for the proposal is a little higher among the young, the urban, graduates and the civically active, and overall it sits just under half. All the numbers, including how willing different people are to accept an invitation, are made up for teaching. They're not estimates of any real place.

Try it in order. With pure lottery and an assembly of 99, draw a few single assemblies. Each one is a little off: one might have 22% young people, the next 33%. Then draw 1,000. The middle 95% of results for a trait held by about half the town spans roughly ten points either side, which is exactly what the textbook formula predicts: 1.96 × √(0.5 × 0.5 ÷ 99) ≈ 0.098. Move the size up to 500 and that range shrinks to about four points. A lottery is fair on average but noisy for small bodies, and the noise falls only with the square root of the size, so to halve it you need four times as many people. Watch the count of assemblies with a majority for the proposal: because the town is so close to evenly split, even a perfect lottery of 99 produces a pro-proposal majority in roughly four draws out of ten.

Now switch to invite at random, volunteers serve. Nobody is excluded and invitations go out by lottery, but the civically active accept three times as often (and, in this invented town, graduates somewhat more and the young somewhat less). The assemblies now have far too many graduates and active citizens, too few young people, and, because those groups lean towards the proposal, a majority in favour of something the town as a whole narrowly opposes. Drawing more assemblies doesn't help; the whole range has shifted. This is bias, not noise, and it's the central practical problem of modern sortition. Bailey Flanigan and colleagues, writing in Nature in 2021, put it plainly: organisers want a panel that is representative and want everyone to have an equal chance, but these goals "are in tension owing to differential participation rates across subpopulations".

Finally, try volunteers, filled by quotas. This is roughly what real assemblies do: invite at random, then choose among those who say yes so that the panel matches the population on a few measured traits. Here the quotas cover age, gender, home area and degree, split into 24 groups. Those four traits now come out right every time. Being active in civic life, which wasn't in the quotas, is still badly over-represented, and support for the proposal barely moves. Quotas fix what you measure. They can't fix a trait you didn't measure, or the opinions that go with it.

Real organisers know this. Ireland's 2016–18 Assembly was recruited by a polling company at random sampling points, with quotas for age, gender, social class and region, and it didn't accept volunteers who applied directly. Flanigan's team designed algorithms that pick from the volunteer pool to meet the quotas while keeping each person's chance of selection as equal as possible; by 2021 their implementation had been used to select more than 40 assemblies. Some assemblies also pay members, to draw in people who would otherwise decline. Whether such fixes are enough has to be asked of each assembly.

Part 3 — Deliberative polls: what a random sample thinks after it has studied

An ordinary opinion poll measures what people think right now, usually about questions they haven't thought much about. In 1988 the political scientist James Fishkin proposed a different instrument. Take a proper random sample, survey it, then bring it together for a weekend. Give everyone balanced briefing materials that partisans on each side have checked. Put them in small groups with trained moderators, let them question panels of competing experts, and then survey them again. The difference between the two surveys estimates what the public would think if it had a real chance to consider the question. Fishkin called it a Deliberative Poll, a name that is now a registered trademark.

The Deliberative Polling method as described by Stanford's Deliberative Democracy Lab.

The first was held with a national British sample on crime, and the results were published by Robert Luskin, Fishkin and Roger Jowell in the British Journal of Political Science in 2002. According to Stanford's Deliberative Democracy Lab, the method has since "stimulated more than 150 Deliberative Polling® experiments in over 50 countries and jurisdictions". In 2017 Mongolia went as far as passing a Law on Deliberative Polling.

The best-known practical success came from an unlikely place: Texas electricity companies. Between 1996 and 1998, eight Texas utilities ran Deliberative Polls of their customers about how to meet future electricity needs. A report for the US National Renewable Energy Laboratory describes what happened. "Customers changed their opinions substantially" after the town meetings, and the results "were unanticipated by either the utilities or their regulators", both of which then grew more interested in renewables and efficiency. In 1999 the state legislature included a renewable portfolio standard in its restructuring law, and the report counts "more than 1,000 MW of new renewables capacity" developed since the polls. The polls weren't the only cause, but the report's point was that they gave decision-makers a measurement of informed public opinion they hadn't had.

The results "were unanticipated by either the utilities or their regulators." — NREL report on the Texas Deliberative Polls, 1996–98

Two findings from the research stand out. One concerns polarisation. The "America in One Room" project brought more than 500 registered voters from across the United States together for a long weekend on five major issues, and a separate control group answered the same questions without deliberating. Writing in the American Political Science Review in 2021, Fishkin and colleagues reported that the deliberators "showed large, de-polarizing changes in their policy attitudes and large decreases in affective polarization", the dislike of the other side.

The other is subtler and concerns the mathematics of voting. How Voting Systems Shape Outcomes showed that counting rules matter most when preferences are tangled. Social choice theory adds that when preferences are tangled enough, majority voting can cycle, with A beating B, B beating C and C beating A. Christian List, Luskin, Fishkin and Iain McLean tested whether deliberation helps. Comparing before and after, they found preferences moved closer to "single-peakedness", a structure in which people line up along one shared dimension and majority cycles can't occur. Interestingly, this happened without people simply agreeing more. Deliberation seemed to clarify what the disagreement was about.

The limits are real too. A Deliberative Poll is a research instrument and an advisory one. It tells decision-makers what an informed public would think, but it has no power to make them listen. It's also expensive, since participants must be brought together, housed and paid. The "after" opinion depends on the briefing materials and the experts, so the method's credibility rests on their fairness. And people who agree to spend a weekend discussing politics may differ from those who don't, which is exactly the problem the simulator shows.

Part 4 — Ireland: an assembly, then a referendum

If one case persuaded governments that citizens' assemblies could matter, it was Ireland. Ireland's constitution can only be changed by referendum, and several of its provisions touched questions the political parties found hard to handle.

The first step was a Convention on the Constitution, which met from December 2012 to March 2014. It had a chair, 66 citizens chosen at random, and 33 politicians. Three of its proposals went to referendums. Allowing same-sex marriage was approved in May 2015 by 62% to 38%. Lowering the minimum age for the presidency from 35 to 21 was rejected on the same day by 73% to 27%. Removing the offence of blasphemy was approved in October 2018 by about 65%.

The second step was the Citizens' Assembly of 2016–18, set up by a resolution of the Oireachtas (parliament) in July 2016. It had 99 citizen members and a chair, Ms Justice Mary Laffoy, a Supreme Court judge. The members, and 99 substitutes, were recruited by the polling company RED C at random sampling points around the country, with quotas based on the census for age, gender, social class and region. Direct applications weren't accepted. Over the life of the Assembly, 53 members were replaced and 152 people served in total. After the January 2018 meeting, seven members were found to have been recruited improperly, in breach of the procedure, and were removed. The episode didn't change any recommendation, but it showed how much an assembly's legitimacy depends on the integrity of its selection.

Parliament gave the Assembly five topics: the Eighth Amendment (the constitutional provision on abortion), how best to respond to an ageing population, fixed-term parliaments, the manner in which referendums are held, and climate change. The Eighth Amendment came first. The Assembly received more than 12,000 public submissions on it and heard invited experts. The whole Assembly cost €1,505,960.90 up to June 2018.

The Assembly considered the question from November 2016 to April 2017. Its final votes were more specific, and more divided, than the headlines suggested:

Selected ballot results from the Citizens' Assembly's final report on the Eighth Amendment (2017).

Parliament and government then took up the report, and a constitutional amendment was put to the people. On 25 May 2018 it passed with 66.4% voting yes, on a turnout of about 64%. Supporters of the assembly model point to this as a case where a randomly chosen body worked through a question politicians had avoided for decades and reached conclusions the electorate then broadly endorsed. Some scholars describe Ireland as having "systematized" constitutional deliberation, by linking assembly, parliament and referendum in sequence (Farrell, Suiter and Harris, Irish Political Studies, 2019).

The Assembly's members, and 99 substitutes, were recruited at random sampling points, and direct applications were not accepted.

The fair reading includes three cautions. First, the Assembly didn't choose its own topics; parliament did, and parliament and government decided what to do with the answers. It was influential, but it was advisory. Second, anyone who disagreed with the outcome could fairly ask whether the choice of experts and the framing of the ballots shaped it, as they shape any deliberation. Defenders answer that the random sampling and the published record of the sessions make the process unusually open to exactly that scrutiny. On a question as contested as abortion, people's view of the process often tracks their view of the result, and this piece doesn't settle either. Third, the pattern doesn't always hold. A later Citizens' Assembly on Gender Equality, set up in 2019, reported in 2021. In March 2024 two constitutional amendments on family and care, worded by the government after the Assembly and a parliamentary committee had reported, were both rejected, by about 68% and 74%, on a turnout of about 44%. An assembly can open a question, but voters still decide, and the wording that reaches the ballot may not be the assembly's.

Ireland has kept using the model, with later assemblies on how Dublin is governed, on biodiversity loss and on drug use. Elsewhere the record is mixed. In British Columbia, Ontario and the Netherlands, citizens' assemblies on electoral reform each proposed a new voting system, and none was adopted; a book-length study of the three asks directly why the proposals were rejected (Fournier and colleagues, When Citizens Decide, 2011). France's Citizens' Convention for Climate brought 150 people drawn by lot together from October 2019, with a mandate to find ways of cutting greenhouse gas emissions by at least 40% by 2030 compared with 1990. The president committed to submitting its proposals "without a filter" to a referendum, to parliament or to direct implementation. A study of what followed found that the citizens' empowerment during the Convention "vanished during the joint elaboration of the law", letting "certain political and economic interests" impose their vision (Galván Labrador and Zografos, 2023). And in the German-speaking Community of Belgium, a randomly selected citizens' body works with the regional parliament permanently, which raises a new worry: that a permanent body becomes a target for capture by established elites (Junius, 2025).

Part 5 — Porto Alegre: a city that let its neighbourhoods write the budget

Citizens' assemblies ask a small random sample to think on everyone's behalf. Participatory budgeting takes the opposite route: it opens a real decision to anyone who turns up.

It began in Porto Alegre, in southern Brazil, after a newly elected left-wing city administration took office in 1989. The city had grown fast between 1950 and 1990 and couldn't keep up with demand for public services. Poor settlements with insecure land titles, known locally as vilas, spread around the edges, in contrast to the wealthier core. The new government proposed to let residents decide directly how to spend a share of the city's capital budget, the money not already committed to salaries and running costs.

The process, as described in a 2018 World Resources Institute case study by Rebecca Abers and colleagues, ran as an annual cycle. The city was divided into 16 (later 17) budget regions. Early each year, open assemblies in each neighbourhood identified proposals: pave this street, extend that sewer, build a health post. Regional assemblies heard the government report on the previous year's work and elected delegates, in proportion to the number of people present, to a Regional Delegate Forum. Over several months that forum ranked the neighbourhoods' proposals. The regional assemblies also elected two members each to a Municipal Budget Council, made up only of civil-society representatives, which decided how capital funds would be split among departments and regions, and then which projects each department would carry out, "rigorously following the ordered lists" from the regions. From 1994, parallel thematic forums handled citywide issues such as transport, education and economic development.

Participation grew from fewer than a thousand people in 1989–90 to about 8,000 by 1992 and more than 20,000 a year after that (Participedia). The participants weren't the usual suspects. In 1998, 30.3% of participants in the regional and thematic assemblies earned less than twice the minimum wage, against 11.4% of the city's population. The WRI study also found that the very poorest took part less than that figure might suggest.

Figures from Abers et al., Porto Alegre: Participatory Budgeting and the Challenge of Sustaining Transformative Change, World Resources Institute, 2018. About 300 km of streets were also paved between 1988 and 2002.

The table shows both halves of the story. In its first fifteen years the process went with a dramatic improvement in basic services in poor neighbourhoods and a flowering of local organisation. Correlation isn't proof: Brazil's economy and other city policies changed too. But the case study's authors judge that the budget process redirected investment towards the poor.

The second half is decline. The same administration held the mayoralty from 1989 to 2004. Later mayors kept participatory budgeting but, as the WRI authors put it, gave it "less priority". The share of investment decided through the process had already fallen, from roughly 15% of the budget in 1991–94 to 8–9% in 1995–2000, and it kept falling. Completion rates halved. In March 2017 the city suspended the process for two years. Across Brazil, Carla Bezerra and Murilo Junqueira argue, participatory budgeting declined as tighter fiscal laws left cities with less money to invest, which made the process less worthwhile.

A budget process is only as strong as the money on the table and the government's willingness to honour the lists.

Meanwhile the idea spread. By 2013, according to the WRI study, more than 2,700 governments worldwide used some form of participatory budgeting. Many of these versions are much smaller than Porto Alegre's, often a few percent of a district's budget for local projects. The best large-scale evidence comes from Brazil itself. Michael Touchton and Brian Wampler built a database of Brazil's largest cities over twenty years and found that participatory budgeting programmes were "strongly associated with increases in health care spending, increases in civil society organizations, and decreases in infant mortality rates", with the connection strengthening the longer the programmes lasted (Comparative Political Studies, 2014). That's an association across cities, not a controlled experiment, and cities that adopt such programmes may differ in other ways, but it's consistent with the Porto Alegre story.

Defenders see proof that ordinary residents, including poor ones, can make sound spending decisions when the stakes are real. Sceptics note that it worked best when one committed government championed it, that it handled a modest share of the budget, and that it relied on unpaid volunteers at evening meetings, who burn out. Both readings fit the facts. Either way, participatory budgeting is fragile in a way an election isn't: no one has to abolish it, it can simply be starved.

Part 6 — Switzerland: a country that votes several times a year

If Porto Alegre is the classic case of participation in a city, Switzerland is the classic case of direct democracy in a nation. Its federal constitution gives citizens three tools.

Federal Constitution of the Swiss Confederation, Articles 138–142. Six cantons count as half-cantons in the cantonal majority.

The double majority matters. A constitutional change has to win most voters nationally and most cantons, so small rural cantons can block a change that big urban ones want. The optional referendum gives any well-organised group a veto threat over new laws. Defenders of the system argue that this pushes the Swiss parliament to build broad coalitions in advance, so that no group has a reason to collect signatures.

The volume is extraordinary. A count of the Swissvotes database, maintained at the University of Bern, gives 706 federal votes decided between 1848 and September 2026. Since 2000 there have been about three federal voting days a year, never more than four, with roughly nine questions a year in all. Of the 706 votes, 244 were on popular initiatives, and only 26 of those, about one in nine, were accepted. A failed initiative can still move policy, because parliament may put a counter-proposal on the same ballot, and voters then choose between the two. Average turnout across all federal votes has been about 48%, and since 2000 about 47%.

Low turnout is the most common criticism, and the most commonly misunderstood. Swiss voters are called to the polls several times a year on several questions at once. Clau Dermont's research shows that most citizens vote selectively, taking part in some ballots and not others, so the share who vote at least sometimes is much higher than the turnout on any one day (Swiss Political Science Review, 2016). Defenders call this a division of labour; critics reply that it hands each decision to whoever cares most about that issue.

Swiss men refused women the federal vote in 1959 and granted it in 1971. In one canton, a federal court had to finish the job in 1990.

The sharpest criticism is about minorities and rights, and the Swiss record supplies evidence on both sides. Direct democracy gave women the federal vote only in 1971, decades after most of Europe. In 1959 male voters rejected women's suffrage, with 33.1% in favour and a turnout of 66.7%. On 7 February 1971 they accepted it with 65.7%. At cantonal level, the open-air assembly (Landsgemeinde) of Appenzell Innerrhoden voted against women's suffrage once more on 29 April 1990. On 27 November that year the Federal Supreme Court ruled that the canton's exclusion of women breached the federal constitution's guarantee of equality (BGE 116 Ia 359). Here a court, not a vote, extended the franchise.

Critics see that history, and votes such as the 2009 initiative banning the construction of minarets, accepted by 57.5%, as evidence that direct majorities can bear down on groups too small to win a vote. Defenders reply that women's suffrage was eventually won by the same direct-democratic process, that representative parliaments elsewhere had their own long records of excluding minorities, and that the regular discussion of every major question gives Swiss citizens an unusually deep familiarity with their own laws. Both sides cite the same events.

One more feature is easy to miss. More than half of all federal voting days since 1848 (187 of 332) have put more than one question to voters, often on unrelated subjects. Voters can accept one proposal and reject another on the same afternoon, so a Swiss citizen's record looks less like loyalty to one camp than a long series of separate judgements. Supporters call that the system's deepest virtue; critics say it asks too much of busy people.

Part 7 — vTaiwan and Pol.is: mapping a million opinions

The newest experiments live online. The internet promised to make mass participation cheap, but its commonest forums, comment sections and social media, reward outrage and drown quiet agreement in noise. The Taiwanese platform vTaiwan is the best-documented attempt to design something different.

vTaiwan began in 2014, after the protests that year known as the Sunflower Movement, when civic technologists from a community called g0v ("gov-zero") began working with government ministries on open consultations. A 2018 study by Yu-Tang Hsiao and colleagues describes a process of four stages: proposal, opinion, reflection and legislation. Each issue has a facilitator, and moving from one stage to the next "is decided by consensus from the vTaiwan community". The goal is "rough consensus" among stakeholders, including officials, companies and citizens, before a regulation is written.

The key tool is Pol.is, open-source software (licensed AGPL3) built for this kind of conversation. Participants see short statements (under 140 characters) written by other participants and vote agree, disagree or pass. They can add new statements of their own, but there are no threaded replies, so there's no comment war to join. The software groups people by how they vote, shows a live map of the opinion clusters, and highlights statements that win support across clusters. Its makers say it has been used with groups from 40 people to 40,000.

The best-known case concerned the ride-hailing service UberX, which had started operating in Taiwan while officials considered it an unlicensed taxi company. It was vTaiwan's first use of Pol.is. In its month-long Pol.is survey, according to Hsiao and colleagues, 925 people took part, adding 145 comments and casting 31,115 votes. (vTaiwan's own website gives a larger figure, more than 4,000 participants, probably counting everyone involved at any stage; the paper's figures are the more precise.) Four opinion groups emerged, then merged into two: one of 45% of participants, most of whom agreed that the ministry should "actively outlaw unlicensed passenger vehicles", and one of 55% who welcomed the disruption of the taxi trade. Yet one statement won 95% support from both:

"The government should leverage this opportunity to challenge the taxi industry to improve their management & quality control system, so that drivers & riders would enjoy the same quality service as UberX's." — statement from the vTaiwan UberX consultation, 2015

Those common-ground statements became the agenda for a live-streamed meeting of stakeholders, which agreed that Uber had to meet legal requirements on insurance and driver training. The vTaiwan site reports that across more than 28 cases, about 80% led to some decisive government action. That figure is self-reported.

The study's authors are also frank about the limits. "Very few cases are proposed by citizens." Because vTaiwan is funded by the government, it "struggles to retain autonomy in the selection of topics", and "if a government authority refuses to discuss a public issue, the sensitive topic will not go through the vTaiwan process". Not every consultation becomes law: a draft bill on selling alcohol online was rejected by the legislature. And the people who join such a consultation choose themselves, so Pol.is maps the opinions of those who show up, not of the public. That's the volunteer problem from the simulator again, in digital form.

Pol.is is open source, so anyone can run a consultation with it. Its core idea is to surface agreement across groups instead of ranking the loudest posts. Whether that helps a society reason together, or just smooths over conflicts that ought to be argued out, is one of the open questions in this field. A statement that 95% of people endorse may be wise, or it may be so vague that nobody has to give anything up.

Part 8 — What the evidence says, and the strongest objections

Step back and the experiments share a pattern. Each one fixes a particular weakness of elections and creates a weakness of its own.

The strongest objection to sortition comes from democratic theory itself. The philosopher Cristina Lafont, in Democracy without Shortcuts (2019), argues that asking citizens to defer to a randomly chosen assembly is a "shortcut" that bypasses everyone else. As the Stanford Encyclopedia summarises the objection, sortition "ignores citizens' interests in being part of the process of collective self-governance" and offers them only "an equal chance" to be part of it. Ninety-nine people meeting over several weekends may think harder than millions of voters, but they aren't those millions, and the millions didn't choose them. On this view, assemblies are valuable as a way to improve public debate, not as a substitute for it.

A second objection concerns accountability. An elected representative can be thrown out at the next election, but an assembly member serves once and goes home. That is precisely the point for defenders such as the philosopher Alex Guerrero, who argues that sortition can avoid "capture of the political process by elites", since nobody needs to raise money to be chosen by lot. It is precisely the problem for critics, who ask who answers to the public when an assembly gets something wrong. A related point from the same encyclopedia entry: a lottery can't choose a party or a programme, a "conception of how all the interests in society are to fit together", so assemblies are better at single questions than at governing as a whole.

A third objection is practical, and the cases support it: the power of whoever sits around the assembly. Every assembly described here had its topic chosen by politicians, its experts chosen by organisers and its recommendations filtered by governments. France's climate convention shows what can happen at the last stage. The vTaiwan study shows it at the first, and the Ostbelgien research worries about it throughout. A team that observed the French convention closely found that it built strong consensus among the 150 but struggled to make its proposals resonate with the wider public (Giraudet and colleagues, 2022).

Each experiment fixes a weakness of elections and brings in a weakness of its own.

What do citizens themselves want? One careful study ran a conjoint experiment with a representative sample of Germans who had never taken part in such a forum. Saskia Goldberg and André Bächtiger found that citizens in general "challenge notions of the strong empowerment" of citizen forums. They prefer forums "limited to policy advice" that "collaborate with legacy institutions". Disaffected citizens were somewhat more open to empowering them, but mostly indifferent (British Journal of Political Science, 2022). If that finding generalises, the future of these experiments lies in partnership with elected institutions, not in replacing them.

The case for them is also strong. Random samples that deliberate do change their minds, and in the best-studied cases they become less hostile to each other. Participatory budgeting, where it has been sustained and funded, has gone with real gains for poor neighbourhoods. Direct democracy gives Swiss citizens a final say that no parliament can take away. Well-designed online tools can find agreement that comment sections bury. And Ireland suggests that an assembly can help a society face a question its politicians couldn't.

There's also an honest summary of what isn't known. Much of the evidence comes from prominent case studies, and cases tend to become prominent by succeeding or failing dramatically. Controlled comparisons, like the control group in America in One Room, are the exception. The OECD has surveyed this wave of processes in its report Innovative Citizen Participation and New Democratic Institutions: Catching the Deliberative Wave (2020), but a survey of cases isn't the same as a measurement of effects. Whether assemblies raise trust in government across a whole population, not just among participants, is still an open research question.

The Athenians would have recognised the debate. They used the lot where any citizen could do the job, the vote where they wanted expertise, and an open assembly for decisions that bound everyone. The live question today isn't whether lotteries, referendums, budget assemblies or online maps beat elections. It's which questions each one suits, and how they can be combined so each covers the others' blind spots.