Future Democracies

Democratic innovation means adding something to representative democracy rather than replacing it: a citizens' assembly drawn by lot, a participatory budget, a deliberative poll, an online consultation platform, a change to the rule that converts votes into seats. The claim made for all of them is roughly the same — that ordinary people, given time, information and good facilitation, produce better and more legitimate decisions than the ordinary political process does. The claim is usually justified by a second claim, that democracy is in measured decline and something must be done.

This brief takes both claims seriously enough to check them, and finds that the second is a live measurement dispute rather than a premise, and that the first is true of the deliberating room and unproven everywhere outside it. The awkward question is not does deliberation change the people in the room — it demonstrably does — but what happens to what they decide. Those are different questions with very different evidence bases, and the gap between them is the subject.

The organising finding, stated at the top because everything below is an instance of it: every measured failure in this field is a failure of transmission, and almost every proposed remedy addresses production. Better facilitation, longer deliberation, better recruitment, AI-assisted consensus drafting and permanent standing bodies all improve what comes out of the room. None of them touches what an elected government does next, or how it words the question when it puts the result to voters — which, on the 2024 evidence, is where the outcome is actually decided.

Where this brief stops. This slot owns the input side of democratic authority: who deliberates, how they are selected, what a body recommends, whether recommendations travel, whether participation changes minds, whether an electoral rule changes who wins, and whether the measured state of democracy justifies any of it. Digital Constitutional Systems owns the meta-rules and their entrenchment — amendment procedures, rights entrenchment and its enforceability, judicial review of coded decisions, and the legal authority of a coded rule against its natural-language original. The two slots meet on the same objects and split by question. Ireland's 2024 referendum defeats as evidence about whether assembly recommendations survive transmission belong here; Ireland's constitutional requirement of a referendum to amend, as an amendment rule that constrains what any deliberative output can achieve, belongs there. Taiwan's Join platform as a participation instrument belongs here; the statutory obligation on agencies to reply to it, as an entrenched procedural right, belongs there. AI inside the administrative state belongs to AI-Assisted Governance; AI used inside a deliberative process appears here.

Frontier The premise most of this field rests on is a live measurement dispute between three peer-reviewed papers, and none of them won. The headline decline figures come from expert-coded indices. V-Dem's Democracy Report 2026 reports 44 countries autocratising, containing 41% of world population, against 18 democratising at 5%; 92 autocracies against 87 democracies at the end of 2025; liberal democracies down to 31 from 45 in 2009; and democracy for the average global citizen, population-weighted, back to 1978 levels. Freedom House's 2025 edition reports 60 countries declining against 34 improving and a nineteenth consecutive year of net decline. Both instruments are judgement aggregations. Neither is a count of events.

Frontier The serious challenge found little decline, and it has severe data problems of its own. Little and Meng assembled low-subjectivity indicators — incumbent-party loss rates, winners' vote and seat shares, multiparty-competition share, constitutional executive constraints, term-limit evasion, journalists jailed and murdered — and found little evidence of global democratic decline over the past decade, proposing time-varying coder bias as the explanation. The reply from the V-Dem team is an audit and it lands: the median country-year in that index is missing two indicators, 43% are missing a quarter or more, and 11% are missing half or more, and not at random, because less democratic states systematically lack data. The face-validity failures follow directly. China scores a perfect 1.0 from 1982 to 2017, including 1989, against a V-Dem electoral democracy index of 0.10 for that year; Turkey scores 0.95 in 1980, the year of its military coup, higher than in 1987–90, 1999–2001, or any year since 2007.

Established The index-makers' own concession is the most useful thing in the exchange. They argue the objective–subjective distinction is “a false dichotomy: seemingly ‘objective’ measures often have considerable elements of subjectivity baked into them,” and demonstrate it: coding party seat shares requires judgement about independents and diverging sources, and the NELDA election-coding project's own inter-rater agreement runs from 58% to 98%, averaging 83%. An indicator with 83% inter-rater agreement is not obviously more objective than an expert rating with a published confidence interval.

Established The bias mechanism has been tested independently and was not found. Nineteen waves of expert survey against parallel public survey: frequent participants rate democracy the same as infrequent ones; among 682 political scientists surveyed in October 2022, the 11% who tweeted about democracy weekly rated United States democracy 66/100 against 65 for non-users and 68 for occasional users; of 544 respondents, 89% had never been invited to code for V-Dem, and those who had rated democracy higher in twelve of thirteen countries; and experts rated United States democracy about ten points higher than the general public across 2017 to 2023.

Established What survives all three papers is a narrow, robust and inconvenient claim: electoral competition has not measurably declined in a decade. Incumbent loss rates have been roughly constant since the late 1990s and the share of elections with real multiparty competition shows no decline. Even V-Dem's own audit of itself reports only 32 countries with significant negative movement on its electoral democracy index between 2012 and 2022, against 13 positive and more than 100 showing no significant change — a much narrower statement than the report's headline. The dispute is entirely about the things that sit above elections: executive constraints, judicial independence, media freedom, civil society. Those are what you need experts to see, and they are exactly what you cannot count.

Established Participation is well documented. Influence is largely uncollected. The OECD's database records 80,622 citizens randomly selected cumulatively between 1979 and 2023, 160 new processes between 2021 and 2023 engaging 11,812 citizens, across 34 countries with 96% of cases inside OECD members, and institutionalised standing bodies rising from 22 in 2020 to 41 in 2023. That is a real census. Handwave But the database records design and inputs, not verified outturn. The one influence figure offered — that authorities respond and implement “at least half of the inputs received” in most cases — is self-reported by the commissioning bodies and published by the organisation that promotes the practice. It has the evidentiary standing of a vendor case study, and any sentence of the form “X% of assembly recommendations are implemented” remains unsupported.

Established Exactly one rigorous multi-recommendation tracking study exists, and it is about France. The Convention Citoyenne pour le Climat produced 149 measures; an academic team tracked every one and found 20% implemented in full, 51% in modified form, 29% not at all. Frontier A separate proposal-by-proposal audit by a French environmental investigative outlet, coding each measure as fully implemented, watered down, or abandoned, put full adoption at 15 of 149 — 10%. This brief does not resolve the divergence and both numbers belong in any honest account. The gap is a coding judgement about how much dilution a measure survives, and the entire 51% “modified” band is where the judgement bites. Note the standing on both sides: the academic study was commissioned and funded by a climate foundation and published through a network constituted to advance climate assemblies; the 10% figure is an advocacy-adjacent outlet's own analysis. Use the numbers and discard the adjectives.

Established The load-bearing finding inside the tracking study is not the percentage. It is that government “has been selective in prioritising recommendations based on incentives rather than regulation and has been more responsive to those recommendations that were more detailed”. The filter is systematic and points toward the status quo. It is not random attrition. And it operated under the strongest political commitment ever offered to a mini-public: the promise in April 2019 that what emerged would be submitted “without filter” to parliamentary vote, referendum or direct regulatory application. The convention's own members scored the government's response 3.3 out of 10 in February 2021. Selective compliance under a maximal promise is a harder problem than neglect, because there is no obvious institutional fix left to propose.

Established Ireland is two different institutions, and 2024 inverted the case. The Convention on the Constitution (2012–14) was 66 citizens plus 33 politicians, budget €1.367 million, and produced the marriage-equality recommendation; only 3 of 40 of its recommendations reached a referendum, and of the two put to voters in May 2015 marriage equality passed 62.07% while the presidential age of candidacy was rejected 73.06% to 26.94%. The Citizens' Assembly (2016–18) was 99 citizens with politicians deliberately excluded, budget €2.355 million, and handled abortion. Conflating them makes the case look twice as strong as it is. Established Then, on 8 March 2024, two amendments descended from the Citizens' Assembly on Gender Equality were put to referendum and both were destroyed — the Family amendment rejected 67.69% to 32.31% and the Care amendment rejected 73.93% to 26.07%, on a turnout of 44.36%. The second is the largest referendum defeat in the history of the state.

Established And the 2024 defeats were a failure of the transmission chain, not of the deliberation. The chain is documented and it is long: the Assembly on Gender Equality sat from 2019 and reported in June 2021; an Oireachtas committee formed in December 2021 and reported in December 2022; government announced its intention in March 2023; bills passed in January 2024. Along the way, gender equality was explicitly dropped from the wording in December 2023, to avoid placing gender discrimination above other kinds, producing two narrower amendments than the Assembly had envisaged. Exit polling blamed the wording. The assembly deliberated on one thing and the electorate voted on another. Ireland's record now reads: two referendums won, three lost, most recommendations never reaching a ballot at all, and four subsequent assemblies — Gender Equality, Biodiversity Loss, Dublin Mayor and Drugs Use — with no published implementation accounting for any of them.

Established Participatory budgeting is the strongest sub-literature and it is not about Porto Alegre. The quantitative work is cross-municipal panel analysis across Brazil: participatory-budgeting municipalities matched expenditure more closely to popular preferences, moved a larger share of budget into sanitation and health, and showed measurable reductions in infant mortality, with effects strengthening the longer a programme ran. Frontier The identification problem is serious and the authors flag it: adoption correlates strongly with Workers' Party incumbency, and the literature has not separated the mechanism from the coalition that adopts it. Established The field's own leading researchers state that reliable data is rare and impact evaluation is at a very early stage — a concession that carries weight precisely because they are the practice's principal advocates.

Established The legitimacy effect on non-participants is real, small, and conditional on the thing governments least reliably do. A pre-registered factorial experiment on 1,309 Irish adults found that mini-public involvement raised perceived fairness by 0.7 points — about 27% of a standard deviation — but that when recommendations were ignored, the acceptance effect fell to non-significance. A separate vignette experiment on 4,225 Australians put the procedural-legitimacy gain at Glass's Δ = 0.079. Frontier An effect of eight hundredths of a standard deviation is roughly a fortieth of the malaise this practice is prescribed for.

Established The comparator wins, and this is rarely stated. In the same experiment, referendums produced roughly twice the legitimacy gain of mini-publics — +1.47 against +0.7 on fairness. Effects in both studies concentrate almost entirely among low-trust citizens; high-trust respondents show no boost at all.

Established The spillover literature is thin and its own systematic reviewers say so. A systematic review of 60 eligible studies published between 1999 and 2018 concludes that the evidence for most spillover effects onto non-participants “remains tentative because the relevant body of empirical evidence is still small,” and that it is ambiguous to what degree small-scale forums matter for mass democracy at all. A separate 2025 review of 121 articles plus country cases in Belgium, Ireland and Sweden identifies gaps in research on mini-publics' impact on justice, accountability, expertise and resources — and pitches its recommendations at political coupling and communication, which is to say at the transmission chain rather than the deliberation.

Frontier The open question is whether the in-room effect can be made to scale. A national deliberative-polling field experiment with a control group found large de-polarising shifts in policy attitudes and large decreases in affective polarisation. Standing caveat that applies simultaneously with its peer review: the authors run the laboratory that owns and licenses the method. Whether an effect on 500 people transmits to the 50 million who were not there is the whole problem, and the honest answer on the systematic-review evidence is that it transmits weakly.

Frontier The genuinely open institutional question is whether a duty to report closes the loop, and Brussels is the only place testing it. The Brussels deliberative committees seat 45 randomly selected citizens with 15 MPs from the relevant standing committee, written into parliamentary regulations in 2019. Topics come from citizens or MPs; the Bureau selects and must justify rejecting a citizen-proposed topic; and within six to nine months government and parliament must report what they did about each recommendation, after which the citizens reconvene to assess the follow-up and report to the plenary. Five committees have run, on 5G, homelessness, citizen participation in crises, urban biodiversity and work-linked training. Handwave What has been published from all this is satisfaction: 95% of MPs said citizens contributed constructively and 82% of citizens reported better understanding of the political system. No implementation tally has been published by the one institution designed to produce one, and this brief records that absence rather than filling it.

Frontier Whether AI can improve deliberation is now an empirical question with one large result and one unanswered question inside it. A 2024 study of an AI mediation system ran over 5,000 United Kingdom participants, with 5,734 reported in its methods, and found AI-generated group statements preferred to those written by human mediators, producing higher group agreement and endorsement, with minority perspectives reportedly retained rather than averaged away; a separate demographically representative virtual citizens' assembly deliberated across three weekly sessions. Handwave What the accessible reporting does not establish is whether anyone's opinion changed. The measured quantity is endorsement of consensus statements, which is a measure of drafting quality. Treating it as a measure of agreement is the error waiting to be made, and this brief declines to make it.

Frontier Whether the recruitment bias is correctable at national scale is open, and it is the most tractable question in the field. An outreach-based selection method has been demonstrated to move reported federal-election turnout among selected members from 96% down to 77%, which is to say down to the population rate. That is a real correction on a real bias. What is untested is whether it survives being run at the scale of a national assembly, at acceptable cost, and whether a body recruited that way deliberates as productively as one recruited from volunteers — because the trait being removed is enthusiasm for the exercise. Nobody has yet run the same assembly twice with the two recruitment methods and compared what came out. That experiment is cheap, and its absence after four decades is the field's most conspicuous incuriosity about its own foundations.

Frontier What merely sounds open: whether deliberation “works”. It does, on every measure taken inside the room, and has for decades — participants become better informed, shift positions on evidence, and de-polarise. Repeating that question is how the field avoids the one it cannot answer. Similarly settled is whether ordinary citizens can handle technical material; whether facilitation can be professionalised; and whether the logistics are affordable. None of these is a frontier. The frontier is the six inches between a recommendation and a statute.

Frontier Liquid democracy has two real deployments and no governmental one. The best-measured is a party platform: 13,836 registered users, 499,009 votes, 6,517 initiatives and 14,964 delegations over 1,200 days. Its headline finding cuts against the standard critique — super-voters held high theoretical power but voted with the majority, and removing all delegations would have changed only about one outcome in six. Handwave But the platform ran inside a party that subsequently collapsed electorally, and the only other deployment with published data ran inside a single employer. Claims that liquid democracy is ready to scale have no outturn base.

Established The first bottleneck is transmission, and it binds before anything else because everything upstream of it already works. Deliberation changes participants. Assemblies produce recommendations — Ireland's Drugs Use assembly carried all 36 of its recommendations, including a comprehensive health-led response to possession for personal use. Platforms collect proposals — Join took 13,853. And then: 289 of 13,853 Join proposals crossed the threshold to compel a reply; 15 of 149 French measures were fully adopted on one coding and 30 of 149 on the other; two of two Irish referendums descended from an assembly were lost in 2024 on wording the assembly had not written; and zero published implementation tallies exist for the Brussels committees or for any Irish assembly since 2021.

Established The second is durability, and it kills the best-evidenced cases specifically. Porto Alegre grew participation from under 1,000 a year to roughly 40,000, had a quantitative literature crediting it with better health spending and lower infant mortality — and was suspended in 2017 by an incoming administration citing lack of resources, having already been undercut by federal transfers that carried no participation requirement. Ostbelgien, designed explicitly for permanence, has walked its own design back: a target of one to three assemblies a year became roughly one; the citizen-initiated agenda route was abolished because the signature threshold was rarely met; and parliament now selects at least one topic itself. Democratic innovations do not usually die of not working. They die of a change of administration, a budget line, or a capacity ceiling.

Established The third is recruitment, and it is a bottleneck on legitimacy rather than on numbers. A 7% response rate is survivable as a logistics problem and corrosive as a representativeness problem, because the 96%-versus-77% turnout gap says the filter selects on exactly the trait the exercise claims to be correcting for. Every argument for sortition's superiority over election rests on the body being a microcosm. It is a microcosm of the letter-answering public.

Frontier The fourth is that the evidence base cannot currently distinguish success from failure. Sixty studies of spillover yield “tentative” evidence. A framework for categorising what mini-publics actually recommend was still being pilot-coded in late 2025, on 20% of 130 recommendations from ten Polish assemblies. The field's flagship census records design and self-reported response. A practice cannot be optimised against a measurement that does not exist, and four decades in, it does not exist.

Frontier The fifth is political, and 2024 showed it has a ceiling. Electoral-system reform was rejected at the ballot box in five American states by an average of 58.5% opposition, constitutionally prohibited in Missouri by 68.4%, and survived repeal in Alaska by 743 votes — while being approved in Washington, D.C. by 72.9%. Whatever the technical merits, the public appetite for procedural reform is not uniformly there, and where reform is put directly to voters it now frequently loses.

Established What is not a bottleneck, stated plainly. Deliberative quality is not a bottleneck; it is the field's solved problem. Recruitment logistics are not a bottleneck; the multiplier is known and affordable. Technology is not a bottleneck; the platforms are ordinary web software and the AI mediation tooling now works better than human mediators on the drafting task. Public willingness is not the bottleneck either — survey work finds around two-thirds support for governments being obliged to implement assembly recommendations. What binds is that no jurisdiction has adopted such an obligation, and the two that came closest produced satisfaction statistics and a broken promise respectively.

Established Nothing here waits on a research result. Sortition is a lottery, facilitation is a trained profession, and the platforms are ordinary web software. Even the newest technical input — AI-assisted consensus drafting — already outperforms human mediators on the task it was tested for. This is a field whose capabilities exceed its institutional uptake by a wide margin, which is unusual in this corpus and which is why the dependency row below carries no scientific prerequisites.

Established What the field waits on is institutional, and all of it is within existing legal competence. A statutory basis that survives a change of government, which is what the durability record says is missing. A duty to respond that is enforceable rather than declaratory, since the experimental evidence is that honouring recommendations is what carries the legitimacy effect. A funded standing secretariat with capacity to run more than one process a year, the absence of which halved the flagship permanent body's own frequency. And independent tracking of recommendations to their legislative fate, done properly exactly once anywhere.

Frontier It depends, weakly and newly, on recruitment methods that do not select for civic disposition. The 96%-versus-77% turnout gap is a dependency in the sense that every legitimacy argument for sortition assumes it away. An outreach-based method closes it; whether that method scales to national assemblies at acceptable cost is untested.

Established What depends on this is narrower than the field claims. No brief on this map is gated by a deliberative result. The honest statement of the enabling relation is that a second channel of democratic authority would matter most on questions where electoral incentives are worst — constitutional change, long horizons, distributional trade-offs, and the governance of technologies whose costs land after the next election. AI Governance is the live instance: citizens' assemblies are increasingly proposed there as the legitimating device, which means this brief's transmission finding is load-bearing for that one.

Established The decisive experiment is cheap, obvious and almost never run: independently track every recommendation from a body to its legislative fate. It has been done properly once, for France, and a second outlet's independent coding of the same 149 measures produced a materially different headline number — which is itself the most useful methodological result in the field, because it shows the coding rule is where the disagreement lives. Doing this for ten assemblies across five countries, with a published and contested coding scheme, would settle more than another decade of case studies.

Established Brussels is the natural experiment already running, and its readout is overdue. Five deliberative committees have completed, each carrying a duty on government and parliament to report within six to nine months on every recommendation, and a reconvening of the citizens to judge that report. The instrument to measure implementation exists, has been exercised five times, and has produced no published tally. Publishing one would be the single most valuable document this field could produce, and it requires no new institution, no new funding and no new consent.

Frontier Randomise the response rather than the deliberation. The legitimacy experiments already show that honouring recommendations is what carries the effect, so the untested variable is the government's behaviour, not the citizens'. A design in which a pre-committed authority acts on a randomly chosen subset of equally-ranked recommendations would identify the legitimacy return to compliance directly, and no such study was found.

Frontier Measure the counterfactual nobody measures. What does a comparable committee of elected members produce on the same question, with the same evidence and the same time? Every claim of deliberative superiority is a comparative claim and the comparison has not been run.

Established The recruitment experiment has already produced a negative result worth recording. An outreach-based selection method, run against conventional lottery recruitment, moved reported federal-election turnout among selected members from 96% to 77%. That is not a null result; it is a demonstration that the standard method has a measurable and correctable bias, published by a practitioner organisation whose commercial interest runs the other way.

Frontier Run the same assembly twice with two recruitment methods. The bias is measured, the correction exists, and nobody has tested what the correction costs in deliberative output. If a body recruited to match population turnout produces materially different recommendations from one recruited by standard lottery, every published assembly result acquires a caveat. If it does not, the representativeness objection loses most of its force. Either answer is worth more than another case study, and the design is a single duplicated assembly.

Frontier Publish the vote margins. Ireland's Drugs Use assembly carried all 36 recommendations and released no vote counts. Margins are the cheapest available signal of which recommendations a deliberating body actually converged on and which it merely tolerated, and their systematic absence makes it impossible to tell whether governments filter toward the assembly's own weak preferences or against its strong ones. The convening bodies already hold this data.

Established And 2024 ran the largest natural experiment this subject has had on the technology side. The prediction that generative AI would corrupt a global election year was tested against the United Kingdom, French and European elections and failed: 16 viral cases in the United Kingdom, 11 across the European Union and France, and no evidence of impact on results. Recording that as a negative result rather than quietly moving on is the point of running it.

Established The engineering is procedural, and the first stage of it leaks badly. Stratified random selection begins with a mass mailing, and roughly 7% of invitations are answered. The working arithmetic given by practitioners is that seating 100 participants requires contacting about 4,000 people and expecting about 280 responses to select from, a contact-to-participant multiplier of 40 to 200. Stratification then corrects the resulting pool for age, gender, education and region.

Established What stratification cannot correct for is the disposition to answer a letter from the government, and that turns out to be the trait that matters. Among conventionally recruited assembly members, 96% had voted in German federal elections. Under an outreach-based random selection method designed specifically to counter self-selection, the figure fell to 77% — which is approximately the actual national turnout. Standard sortition therefore produces a demographically representative sample of unusually civic people, over-representing habitual voters by close to twenty points on the single dimension most relevant to political judgement. This is the most under-reported number in the field and it is measurable at every assembly, because every organiser knows their own response rate.

Established The rest of the procedural stack is well understood and unevenly executed. Facilitation must be auditable rather than merely skilled. Voting procedures must be transparent: one participant in the Irish Assembly described its process as “very sloppy”, and no external observer could check. Deliberation needs time — the field's own practitioners put the minimum at four to five days. Frontier And the agenda-setting rule is the most consequential design choice available: the 2016–18 Irish Assembly took every topic from government; Ostbelgien abandoned its citizen-initiated route entirely; and Brussels keeps a citizen route but reserves selection to the parliamentary Bureau, which must at least justify a refusal.

Established Online platforms are a different engineering problem and their throughput is the number to look at. Taiwan's Join platform obliges the relevant agency to issue a reasoned reply once a proposal reaches 5,000 signatures. By June 2022 it had received 13,853 proposals, of which 289 met the threshold. That is 2.1% — the arithmetic is this brief's, the inputs are sourced. A statutory response duty triggered by one proposal in fifty is a real institution and a narrow one, and it is the only quantified example anywhere of a platform obligation actually binding an executive agency.

Established The comparison between the two Taiwanese systems is the useful engineering lesson. vTaiwan combined Pol.is opinion-clustering with in-person stakeholder meetings across four stages, run through a government innovation unit with policy-authorised staff required to attend as trained participation officers. On its best-documented case, the UberX regulation of 2015, more than 4,500 citizens voted on 145 opinions. It was also, in its own evaluators' words, complex and tech-intensive, disadvantaging non-tech-savvy participants and confined to digital-policy questions; drivers with lower digital literacy reported being marginalised in the very case it is most famous for. Join is cruder, has no deliberative phase worth the name, reaches an older and less technical population, and is the one with a legal duty attached. The cruder instrument with a statutory hook outlasted the sophisticated one without.

The nearest neighbour is Digital Constitutional Systems, and the relationship is complementary rather than overlapping. That brief's central finding is that every effective constitutional constraint on a digital system in the record was imposed by an ordinary court applying an instrument written decades earlier — four for the courts, zero for the code. This brief's central finding is the same shape from the other side: every effective conversion of public judgement into policy in the record ran through an ordinary elected institution. Read together they say that the entrenched institutions keep winning, and that the innovations which work are the ones that plug into them rather than around them.

AI Governance is adjacent in the strong sense, because citizens' assemblies are now routinely proposed there as the legitimating device for decisions nobody wants to make electorally. Anyone making that proposal is relying on a transmission mechanism this brief finds unmeasured, and should read the French record first.

AI-Assisted Governance shares this brief's characteristic failure mode: an apparatus that is measured on process compliance and not on outcome. Collective Intelligence meets the same aggregation problem as a design question, where this brief meets it as a constitutional one — the difference being that a design question has a best answer and a constitutional question has only a legitimate procedure.

Electoral-system reform is adjacent in a way this brief treats as substantive rather than as a neighbouring topic, because it is the alternative use of the same reforming energy and it has a longer and harder evidence record. Ranked-choice voting and proportional representation change who wins rather than who is consulted, which makes their effects easier to identify and their politics harder — a point 2024 made forcefully, when five American states rejected ranked-choice voting outright.

Outside the map: comparative electoral-system research, where the causal evidence is thinner than the advocacy on both sides; survey methodology, which supplies the instruments the democracy indices are built from and the critiques they are attacked with; the philosophy of democratic legitimacy, which is where the argument about empowered mini-publics is genuinely being conducted; and research integrity, because this is a field in which a large majority of the accessible literature is written by people who are paid to promote the practice, and the ethical and institutional sections treat that as substance rather than as an aside.

Established The defining institutional fact is that no jurisdiction has adopted an enforceable duty to act on a deliberative recommendation, and the two that came closest failed in opposite ways. Brussels has the strongest procedural duty in existence: 45 randomly selected citizens sitting with 15 MPs from the relevant standing committee, written into the regulations of two parliaments in 2019, with topics selectable by citizens and the Bureau obliged to justify a refusal, a duty on government and parliament to report on every recommendation within six to nine months, and a reconvening of the citizens to assess that report. Five committees have run. What has been published is that 95% of MPs found citizens constructive and 82% of citizens better understood the political system. Handwave No implementation tally has been published by the institution built to produce one. France, at the other extreme, offered a political commitment rather than a procedural one — the promise that the output would go forward without filter — and delivered 10 to 20% full adoption and a member satisfaction score of 3.3 out of 10. A procedural duty produced no measurement; a political promise produced systematic filtering. Neither produced compliance.

Frontier The institutional design that has actually persisted is the one with the weakest formal powers. Ostbelgien's standing citizens' council sets agendas and monitors follow-up, attached to a parliament obliged only to justify a refusal — and it has walked back its own assembly frequency and abolished its citizen-initiated agenda route. Whether a weak duty that survives beats a strong duty that gets repealed is an open question and the most practically important one in this brief. The Brussels evidence suggests a third possibility neither option anticipated: a duty that survives, is exercised, and is never measured.

Established Who convenes determines what gets asked, and it is almost never the citizens. The 2016–18 Irish Assembly took every topic from government. Ostbelgien abolished its citizen route because the signature threshold was rarely met, and parliament now selects at least one topic itself. Brussels reserves selection to the parliamentary Bureau. Taiwan's Join is the exception and quantifies the cost of a genuinely open route: 13,853 proposals to June 2022, of which 289 — 2.1% — crossed the 5,000-signature threshold that compels an agency reply.

Established The institution that does not exist is an independent evaluator. There is no audit office, statutory reviewer, or standing academic consortium that tracks deliberative recommendations to their legislative fate the way public spending is tracked to its outputs. The one properly tracked case was a foundation-funded academic project on a single convention in a single country, and its headline is contested by a second independent coding. Meanwhile a framework merely for categorising what mini-publics recommend was still being pilot-tested in late 2025, on 20% of 130 recommendations from ten Polish assemblies. Four decades in, the field lacks the measurement instrument that would tell it whether it works.

Frontier Public appetite is not the constraint, which is itself an institutional finding. Survey work across four large European democracies finds around a third supporting institutionalisation of deliberative democracy and around two-thirds supporting an obligation on government to implement recommendations. That is a wide gap between what publics say they want and what any legislature has enacted, and it locates the blockage precisely: in the body that would have to bind itself.

Established The one institution built in this period that acquired real power did so by contesting an election. A party founded on 8 May 2025 around digital-democracy tooling won 11 of 465 seats on 6.66% on 8 February 2026, having taken one seat of 125 on 2.6% seven months earlier, and now ships open-source political-finance visualisation and AI-assisted plain-language legislative tracking — one component of which was adopted by a rival party. It has also had an ordinary political scandal, losing two likely seats over a candidate's undisclosed prior employment. That mixture of real power and ordinary vulnerability is what an institution looks like, and no advisory body in this brief has either half of it.

Evidence quality is the first ethical question here, because a large majority of the accessible literature is produced by people paid to promote the practice. Established The case census is maintained by an intergovernmental body's open-government unit and reported on that unit's own staff blog. The best implementation study was funded by a climate foundation and published through a network constituted to advance climate assemblies. The flagship deliberative-polling evidence comes from the laboratory that owns and licenses the method. The most-cited overview of the global movement was written by the person who then ran the OECD's deliberative democracy work and subsequently founded an organisation to promote it. The clearest description of the recruitment problem comes from a practitioner whose employer sells recruitment services — a concession against interest, and weighted accordingly. None of this makes the work wrong. It means the field has very little independent evaluation and almost no audit-institution scrutiny of the kind that exists for ordinary public spending.

Frontier The sharpest objection is not that assemblies fail but that they might substitute for public reasoning rather than constitute it. The philosophical case against empowered mini-publics is that a citizen who did not deliberate has no reason to expect the body's decisions to track her interests and values — the same structural position as a voter choosing representatives by tossing a coin, which is formally what sortition is. The standard proposed is that all those subject to a rule can see themselves as its authors, with ongoing control over substantive content rather than merely formal authorisation. The reply is that sortition produces inclusive equality because only chance distinguishes selector from selected, that public justification supplies accountability without electoral sanction, and that participants return to ordinary life. The reply also concedes the strongest form of the worry: informed participants risk ceasing to be a representative sample of the citizenry at large.

Established The recruitment evidence turns that philosophical concession into a measurement. A body whose members report 96% federal-election turnout against a national 77% was not representative before it deliberated. The empirical result and the philosophical objection are the same finding arrived at from opposite directions, and neither literature cites the other.

Established And a filtered assembly may be worse than none. The experimental result is that the acceptance benefit disappears when recommendations are ignored; the French evidence is that governments filter systematically toward incentives and away from regulation. An institution that reliably raises expectations and reliably disappoints them on the regulatory questions that matter most is not obviously a net contribution to legitimacy. Ireland's 2024 defeats are the strongest available instance: a long transmission chain altered the question, the electorate rejected the altered question by roughly three to one, and the assembly's name is attached to the defeat.

Frontier Public money and public accountability, stated as an unresolved obligation. Ireland's two constitutional bodies cost €1.367 million and €2.355 million; Paris runs a participatory budget of €100 million a year alongside its permanent assembly; the OECD counts 41 institutionalised processes as of 2023. No jurisdiction has published a recommendation-by-recommendation implementation accounting for its own bodies, including the one jurisdiction whose parliamentary regulations require a report on every recommendation within six to nine months. This brief's position is that a publicly funded institution which is not evaluated is an institution whose value is being asserted rather than demonstrated, and that the obligation to publish that accounting falls on the convening authorities rather than on academics who happen to obtain funding.

Frontier The premise is contested, which is awkward for the prescription. Democratic innovation is almost always justified by democratic backsliding, and backsliding is a measurement dispute with three peer-reviewed papers and no resolution. The indices report 44 countries autocratising, a nineteenth consecutive year of net decline, and democracy back at 1978 levels population-weighted. The challenge finds little decline on low-subjectivity indicators and has severe missing-data problems. The bias mechanism has been tested and not found. Established What all parties concede is that electoral competition has not substantially declined — which sits awkwardly under an argument for innovation, since the residual dispute is about executive constraints, judicial independence and media freedom, none of which a citizens' assembly touches.

Established The general principle this case illustrates is about the location of the binding constraint, and it recurs across this corpus. A capability can be demonstrated, refined, and made cheap, and remain unused for decades because nothing downstream of it is obliged to act. Deliberative democracy has solved the production of good, representative, well-informed public judgement — a genuinely hard problem it is reasonable to be proud of solving. It has not solved, and has barely begun to measure, the problem of getting that judgement to bind anyone. Four decades and eighty thousand randomly selected citizens have produced one properly tracked implementation study, and its headline number is contested by a factor of two.

Speculative If deliberative bodies do become permanent constitutional furniture, the interesting consequence is not better policy. It is a second source of democratic authority alongside election, with no settled account of what happens when the two disagree. The legitimacy objection is not a technicality here: a citizen who did not deliberate has no particular reason to expect a hundred strangers who did to track her interests, and the honest reply from the practice's defenders is that participants remain ordinary people who return to ordinary life. Both positions are held by serious people and neither has been tested against a case where an empowered assembly overrode an elected majority, because no such case exists.

Established A quieter significance, which the 2025–26 record makes visible. The route from democratic-innovation technique to actual state power that was demonstrated in this period ran through an ordinary ballot: a party founded in May 2025 holding 11 of 465 seats by February 2026, shipping transparency software as its product. Nothing in the deliberative record matched that, and the asymmetry is a finding about where authority lives rather than about which tools are better.

Established What already happened, because this timeline usually starts too late. The practice is four decades old: the OECD's census counts randomly selected participants back to 1979 and totals 80,622 of them to 2023. Porto Alegre's participatory budget ran from 1989 and was suspended in 2017. Ireland's Convention on the Constitution sat 2012–14 and its Citizens' Assembly 2016–18. vTaiwan launched in 2014 and stopped driving major decisions after 2018. Ostbelgien's permanent council was created in 2019, the same year Brussels wrote deliberative committees into parliamentary regulations. The French convention reported in 2020.

Established 2024 was the year the exemplars broke, and it is now the reference year for this subject. On 8 March, Ireland lost two assembly-descended referendums by 67.69% and 73.93%. In November, five American states rejected ranked-choice voting, Missouri prohibited it constitutionally, and Alaska retained it by 743 votes. And across three major European elections the predicted AI-disinformation catastrophe did not occur.

Established 2025 to 2026: the interesting movement was electoral, not deliberative. A Japanese engineer who came fifth in the July 2024 Tokyo governor's race with 154,638 votes founded a party on 8 May 2025, won one seat of 125 on 2.6% that July, and took 11 of 465 seats on 6.66% — 3,813,749 votes — on 8 February 2026, shipping open-source political-finance and legislative-tracking software as its actual product. The digital-democracy toolkit reached a national legislature by winning an ordinary election.

Frontier Next: whoever publishes an implementation series first sets the terms of the field. Brussels has the mechanism and has not used it. The OECD has the census and does not collect the variable. Ireland has four unreported assemblies. There is no technical or financial obstacle to any of these, which is why no date attaches to them and why their absence is a choice rather than a constraint.

Speculative Within twenty-five years the practice forks, and the two branches are distinguishable now. Either implementation tracking becomes routine — Brussels uses the mechanism it already wrote down, the OECD adds the variable to its census, Ireland reports its four outstanding assemblies — and the field acquires the evidence base it currently lacks; or the practice settles as a legitimation device, commissioned when a government wants cover on a hard question and shelved when the answer is inconvenient. Frontier On present evidence the second is the better bet, and the evidence has strengthened since this brief last said so: the exemplar case lost both its referendums by margins of 68 and 74 points after its question was altered in transmission, the convention whose recommendations were promised unfiltered delivered 10 to 20%, and nobody who could publish an implementation series has chosen to. The fork is not technical. It turns on whether any commissioning government wants the number.

Speculative Longer run: the plausible equilibrium is a standing consultative tier on constitutional and long-horizon questions, where electoral incentives are worst, rather than general-purpose decision-making. That is roughly what Ostbelgien and Brussels already are, and it is what survives contact with both the legitimacy objection and the durability record.

Handwave Any projection of assemblies replacing legislatures. No such institution exists anywhere, no jurisdiction has adopted an enforceable duty to implement, and the one case where a head of state promised to transmit recommendations unfiltered delivered 10 to 20% and a member satisfaction score of 3.3 out of 10. This brief declines to supply a date.

• Depends on Nothing on this map. No result produced by another brief here is on this topic's critical path. Sortition is a lottery, the platforms are ordinary software, and the newest technical input — AI consensus drafting — already exceeds human performance on its measured task. The constraints are institutional throughout and are recorded below.
• Requires (not on this map) A statutory basis that outlives the administration which created it — the constraint that killed the best-evidenced case, where two decades of measured results ended with a change of mayor. A duty to respond that is enforceable rather than declaratory: Brussels has the strongest such duty in existence, requiring a report on every recommendation within six to nine months, and has published satisfaction figures instead of an implementation tally. Independent tracking of recommendations to their legislative fate, done properly exactly once anywhere, and contested even there between a 10% and a 20% headline. A recruitment method that does not select for civic disposition, since standard lottery recruitment yields members with 96% federal-election turnout against a national 77%. And a funded standing secretariat: the flagship permanent body halved its own assembly frequency on capacity grounds. None is a research result and all five are why this field's central claim remains unevaluated.
• Enables A second channel of democratic authority on questions where electoral incentives are worst — constitutional change, long horizons, distributional trade-offs. No typed enabling edge is claimed: briefs that assume functioning democratic institutions assume them generically rather than depending on this route. The nearest thing to a real edge runs to AI Governance, where assemblies are increasingly proposed as the legitimating device and where this brief's transmission finding therefore bites.
• Adjacent Digital Constitutional Systems, which owns the amendment rules that determine what any deliberative output can achieve; AI-Assisted Governance, which shares the problem of an accountability apparatus measured on process rather than outcome; Collective Intelligence, which treats aggregation as a design question where this brief meets it as a constitutional one; and outside the map, comparative electoral-system research, survey methodology, and the experimental political science that supplies nearly all the causal evidence here.

“Ireland proves citizens' assemblies deliver constitutional change.” Established The 2024 record inverts the case. Two amendments descended from the Citizens' Assembly on Gender Equality were defeated on 8 March 2024 by 67.69% and 73.93%, on a 44.36% turnout, the second being the largest referendum defeat in the state's history. Exit polling blamed the wording — and the wording had been altered along a chain running from a 2021 assembly report through a 2022 parliamentary committee to a 2024 bill, with gender equality explicitly dropped in December 2023. Add the earlier record — only 3 of 40 Convention recommendations reaching a referendum, and the presidential age of candidacy rejected 73.06% to 26.94% in 2015 — and Ireland shows two wins, three losses, and four subsequent assemblies with no published implementation accounting. The correct reading is not that assemblies fail. It is that the assembly is not the part of the chain that determines the outcome.

“Ireland's marriage-equality referendum came from the Citizens' Assembly.” Established It came from the earlier Convention on the Constitution, a mixed body containing 33 politicians alongside 66 citizens. The Assembly that excluded politicians handled abortion. The two are routinely merged, which doubles the apparent strength of the assembly-to-referendum story, and the Assembly's own website compounds the error by stating that the second 2015 referendum also passed. It did not. A convening body is wrong about its own outcome on its own site, which is the cleanest possible illustration of why self-description is not a standing.

“Governments just ignore assemblies.” Established Too pessimistic, and what actually happens is worse. The French government did not ignore the Convention Citoyenne; it filtered it, selectively prioritising recommendations based on incentives rather than regulation and responding more to recommendations that were more detailed. It did so having promised in April 2019 that the output would go forward “without filter”. The result was 20% fully implemented on the academic tracking and 10% on an independent proposal-by-proposal audit, with the convention's own members scoring the response 3.3 out of 10. Selective compliance under the strongest promise ever offered is a harder problem than neglect, because there is no institutional fix left to propose.

“X% of assembly recommendations get implemented.” Handwave There is no such statistic and the two figures that exist for the one properly tracked case disagree by a factor of two. This brief does not resolve the 10%-versus-20% divergence, and states both, because the gap is a coding judgement about how much dilution a measure survives rather than a dispute about facts. There is also no “OECD deliberative wave dataset” showing implementation rates: there is a 2020 report covering 289 cases to October 2019 and a live database whose 2023 update records 34 countries and 80,622 cumulative participants — different objects with different denominators, and neither containing an independently verified implementation variable.

“Sortition produces a representative sample of the public.” Established It produces a demographically stratified sample of people who answer official letters. Response rates run around 7%. Among conventionally recruited assembly members, 96% had voted in German federal elections, against a national turnout near 77%; an outreach-based recruitment method closed the gap. A body over-representing habitual voters by roughly twenty points is not a microcosm of the public on the single dimension most relevant to political judgement, and every argument for sortition's superiority over election assumes it is.

“Democracy is measurably declining.” Frontier It is measurably declining on expert-coded indices, which is a different sentence. The best-known challenge to those indices has severe data problems of its own — a median of two missing indicators per country-year, China scoring a perfect 1.0 from 1982 to 2017 including 1989, Turkey scoring 0.95 in its 1980 coup year. The pessimism-bias mechanism it posits has been tested independently and not found. And the index-makers' own concession is that the objective–subjective distinction is a false dichotomy, with a leading “objective” election-coding project running at 83% average inter-rater agreement. Established What survives all of it: electoral competition has not measurably declined in a decade, and the dispute is entirely about the things above elections, which are the things you need experts to see.

“AI-generated disinformation is transforming elections.” Established It did not, in 2024, in the elections that were systematically reviewed. A structured review of the United Kingdom, French and European elections found 16 viral AI-disinformation cases in the United Kingdom and 11 across the European Union and France combined, with exposure concentrated among users already holding aligned beliefs, and no evidence that AI-enabled disinformation or deepfakes meaningfully impacted results. The durable mechanism argument is that AI removes the cost of generating persuasive content while the binding constraint on influence operations is distribution — exposure costs and audience sizes — and that constraint did not move. The harms that did occur were real and of a different kind: deepfakes inciting hate and death threats against candidates, deepfake pornography of women politicians, generalised confusion about authenticity, and politicians using unlabelled AI in their own campaigns. The accompanying warning is worth carrying whole — that the AI panic may distract from voter disenfranchisement and attacks on election administration, which are cheaper, older and better evidenced.

“vTaiwan is the model for digital democracy.” Established vTaiwan has not driven a major decision since 2018. Its own co-creator says it could have been more effective; the reasons given are a difficult interface, loss of public interest, no mandate on government to adopt anything, legislators not taking it seriously, a scope confined to digital policy, and confusion with the parallel Join platform. An independent foundation assessment calls it complex and tech-intensive and states explicitly that Taiwan's approach does not serve as a transferable model. Frontier The sources conflict and this brief does not resolve the conflict: the nonprofit that maintains the underlying opinion-clustering software — an interested party — describes vTaiwan as still active, claims 200,000-plus cumulative participants, and states that more than half of Taiwan's population has interacted with Join. The reporting of decline and the maintainer's account of health cannot both be right, and no source consulted reconciles them.

“Digital democracy means consulting citizens between elections.” Established The most consequential digital-democracy result of 2025 to 2026 was an ordinary electoral one. A party founded on 8 May 2025 by an engineer who had come fifth in the Tokyo governor's race took one seat of 125 on 2.6% in July 2025 and 11 of 465 seats on 6.66% in February 2026, shipping open-source political-finance visualisation and plain-language legislative tracking as its actual product, one of which was adopted by a rival party. Contesting elections turned out to be the faster route into a legislature than being appointed to advise one.

“Ranked-choice voting is spreading.” Established In November 2024 it was rejected in five states by an average of 58.5% opposition — Idaho's measure failing 70.0% to 30.0% — constitutionally prohibited in Missouri by 68.4% to 31.6%, and retained in Alaska by 743 votes after a recount. It was also approved in Washington, D.C. by 72.9% to 27.1%. Frontier The evidence underneath is genuinely mixed: no aggregate citywide turnout increase but a nine-point increase among young voters; campaign-civility perceptions that improve overall but may not extend to front-runners; a nine-point rise in minority candidate share in one study against a candidate-count effect that diminishes in later cycles in another. It is a reform with real but modest measured effects that has recently been losing at the ballot box.

“Participatory budgeting's evidence is the Porto Alegre story.” Established It is cross-municipal panel econometrics across Brazil, and Porto Alegre itself suspended the practice in 2017. The North American record is a different and much smaller thing: New York City's 2014–15 cycle put 58,095 participants in charge of $31.9 million across 179 neighbourhood assemblies, with 57% of participatory-budget voters identifying as people of colour against 47% of local election voters — a genuine inclusion result attached to a sum that is a small fraction of a percent of the city's budget.