Beyond "Which System Wins": Designing Social Systems with AI, and What the Evidence Allows

1Key Findings

• Hayek's 1945 argument was already a claim about system design: the economic problem is using knowledge that exists only in scattered, local form, and prices are a way of communicating it [4]. Any proposal for an AI-designed economy has to answer this argument, not a capitalism-versus-socialism slogan.
• A 2023 peer-reviewed paper argues that big data and AI do not solve the calculation problem, because the problem is about which institutions exist, not about computing power. That claim is disputed by a cyber-communist literature arguing that computation does make planning feasible [8].
• Project Cybersyn (Chile, 1970–1973) was an early, real attempt to run part of a national economy as a cybernetic system. Its historian finds that its technical structure reflected the government's socialist ideology. Design did not escape ideology; it encoded it [9].
• Arrow's theorem shows that, except in trivial cases, no procedure for turning individual preferences into a social ranking meets a short list of reasonable conditions. Arrow himself treated those conditions as value judgments [10]. The known escapes all involve either more information (grades, comparable utilities) or restricting which preferences are allowed. Some evidence suggests deliberation can produce such a restriction [10].
• The AI Economist used reinforcement learning to design tax schedules in a simulated economy, optimizing a chosen objective (equality multiplied by productivity). The preprint reports that its learned policy improved that trade-off by 16% over a prominent analytical tax framework [2].
• DeepMind's Democratic AI trained a system to design a redistribution rule for an economic game, using human majority votes as the objective. The AI-designed rule won majority support over strong human-designed baselines [1].
• The Habermas Machine (Science, 2024) used a language model to write group statements on divisive UK policy questions. Participants preferred its statements to those written by human mediators, and group positions moved toward common ground [3].
• LLM stand-ins for citizens are not yet reliable for testing designs: ChatGPT matched average survey answers but showed less variation than real respondents, gave different regression coefficients, and produced significantly different results from the same prompt over 3 months [6].
• The planning side disputes the Austrian claim of victory: a 2024 paper argues the in-principle case against planning quietly depends on what computing technology can do, and so needs re-examining [13].
• Mechanism design (2007 Nobel) is the existing formal field closest to the thesis: it compares rules and institutions, with markets as one mechanism among many [11]. Ostrom's work documents governance that fits neither market nor state [12].
• Scott's history of high modernism shows designed social orders failing across the political spectrum when they override local, practical knowledge [15]. This is the strongest warning for an AI design programme, because it empowers exactly the planners and technocrats Scott describes.
• Community Notes is the best-measured deployed bridging mechanism. Notes chosen for cross-viewpoint approval cut agreement with misleading tweets by about 26% and likes and retweets by 25–34%, but by design the mechanism needs some agreement across sides in order to work [16].
• The claim that the ideological debate is over was made in 1960 by Daniel Bell and was overtaken within a decade. His test for any proposed utopia (destination, route, cost, who pays) remains a usable design specification [14].
• The tribal side of politics is real and measurable. US affective polarization rose from 22.64 degrees in 1978 to 40.87 degrees in 2016 on the ANES feeling thermometer. But it is "largely distinct from the ideological divide", and ideological sorting explains only about 5% of its variance [7].

2Details

1. Is designing the system instead of picking a side a new idea? The precedent of the calculation debate and Cybersyn

The way the question is framed (the old ideological debate is obsolete because we can now design systems precisely) has a direct historical parallel. The parallel matters because it shows how such claims have fared before.

The debate was always partly about information design. In 1920 Ludwig von Mises argued that economic calculation is impossible in a socialist system without private property and exchange in the means of production. He reasoned that without market exchange there are no money prices, and so no common numbers to do capital accounting with [8]. Friedrich Hayek's 1945 essay The Use of Knowledge in Society turned this into a general claim about how systems process information. He argued that the central economic problem is not allocating given resources with known data. It is using "the knowledge of the particular circumstances of time and place", which no single mind holds in full. He treated the price system as a mechanism for communicating that dispersed knowledge, so each person can act on far more than they know directly [4]. This is a systems-engineering argument made in the 1940s, and it is the strongest version of the case that you cannot simply design an economy from the centre.

The claim that computers will solve it has a long and disputed record. Lambert and Fegley, in the Journal of Economic Behavior & Organization (February 2023), review the debate's side-arguments and the newer cyber-communist literature. That literature includes Cockshott and Cottrell's 1993 proposal for planning in labour-time units, and later work arguing that modern computing makes such calculation feasible [8]. Their position is that "advances in computing power are irrelevant to Mises' most fundamental argument". They argue the calculation problem is about institutional prerequisites, namely property and monetary exchange, not processing speed, and that Mises' argument "remains unrefuted" with respect to big data and AI [8]. The other side of the dispute is put by Dapprich (University of Potsdam) and Greenwood (University of Westminster) in the Erasmus Journal for Philosophy and Economics (July 2024). They note that Austrian economists claim to have won the debate by arguing that the calculation problem is epistemological, not computational. Their reply is that this theoretical claim depends, more than its proponents admit, on what fast-changing computing technology can and cannot do. They argue for re-examining the Austrian conclusions using Mises' own distinction between supply-side and demand-side calculation, and treat recent cybersocialist proposals as opening several separate research avenues, including how calculation, incentives and innovation interact [13]. Dapprich's own work includes a simulation of a socialist economy that plans optimally using consumer feedback [13]. Assessment: Both papers are peer-reviewed and each comes from inside its own tradition. Lambert and Fegley rest on an in-principle argument about institutions; Dapprich and Greenwood argue that the in-principle argument quietly depends on empirical assumptions about computing. This report read the Dapprich and Greenwood abstract and author information, not the full paper. The fair statement is that whether AI dissolves the calculation problem is actively disputed, and that the dispute has shifted from logic toward questions that could in principle be tested. Neither side has shown its case with a working system at national scale.

The one serious attempt to build it. Eden Medina's history of Project Cybersyn (Journal of Latin American Studies, 2006) describes an early computer network built under Salvador Allende (1970–1973) to regulate the nationalised sector and "manage the transition of Chile's economy from capitalism to socialism" [9]. Under British cybernetician Stafford Beer, a Chilean team built cybernetic models of factories and a network to send economic data quickly between government and factory floor [9]. Medina's article also "examines how its structure reflected the socialist ideology of the Allende government" [9]. That is the key lesson for this thesis. The most ambitious design-over-ideology project in history did not stand outside ideology. Its architecture carried the government's political commitments.

The warning from high modernism. James C. Scott's Seeing Like a State (Yale University Press, 1998) is the most cited case history of designed social orders failing. In an excerpt published by the Schumacher Center, Scott traces many of the worst episodes of twentieth-century state development to a combination of elements. The first is high modernism, the aspiration to administratively order nature and society; the third is a weakened civil society unable to resist the plans [15]. Two points matter for this thesis. First, Scott insists high modernism was not the property of one ideology: "As a faith, it was shared by many across a wide spectrum of political ideologies." Its carriers were engineers, planners, technocrats, administrators, architects and scientists [15]. That is exactly the group that an AI-enabled design programme would empower. Second, his central objection is about knowledge, close to Hayek's. Practical, local knowledge, which he calls metis, resists being reduced to formal rules, and "Formal order, to be more explicit, is always and to some degree parasitic on informal processes, which the formal scheme does not recognize, without which it could not exist, and which it alone cannot create or maintain." [15]. Scott is also explicit that he is "emphatically not making a blanket case against either bureaucratic planning or high-modernist ideology", only against an imperial, hegemonic planning mentality that excludes local knowledge [15]. The editors of the excerpt note that, despite Scott's own disclaimers, his argument corroborates the Austrian case against central planning [15].

Inference: The thesis needs to separate itself from two earlier positions: technocratic planning (the computer replaces the market) and market fundamentalism (the price system is the only workable design). The defensible new position is neither. It is that markets and plans are both mechanisms whose properties can be stated, compared and partly tested, and that AI changes the cost of doing the testing.

2. What formal theory says about designing collective choice: there is no neutral aggregator

If ideologies are replaced by designed systems, those systems still have to turn many people's preferences into one decision. Social choice theory is the mathematics of that step, and its central result applies directly to any theory of everything for social organization.

Arrow's impossibility theorem. The Stanford Encyclopedia of Philosophy entry (Michael Morreau, revised December 2025) states the result. Take more than two alternatives. Then no procedure for deriving a social ranking from individual rankings satisfies all of: unrestricted domain (it handles any preferences), social ordering (no cycles), weak Pareto (if everyone prefers x to y, so does society), non-dictatorship, and independence of irrelevant alternatives [10]. The alternatives can be "policies, public projects, candidates in an election, distributions of income" [10]. The theorem covers choosing between whole economic arrangements as much as between candidates. Arrow received the 1972 Nobel Prize in economics for his contributions [10].

Two points from the entry matter for the thesis. First, Arrow regarded his conditions as value judgments, not neutral axioms, and the entry notes that it is sometimes undesirable for a procedure to satisfy all of them [10]. Any aggregation rule an AI uses therefore embodies a political choice about which conditions to give up. Second, proofs of the Gibbard–Satterthwaite theorem link violations of the independence condition to vulnerability to strategic voting [10]. Designed systems face strategic users, not passive data.

The escape routes, and what they cost. The entry reports that Arrow's theorem has been described as being "about the impossibility of trying to do too much with too little information" [10]. The escapes come in two kinds:

• Restricting the domain. If preferences are single-peaked along a shared dimension, majority rule with an odd number of voters yields a consistent ranking, and the median voter's ideal point wins (Duncan Black, 1948) [10]. Miller (1992) and List and Dryzek (2003) argued that deliberation can push preferences into this more aggregable shape, and "List et al. (2013) present empirical evidence that deliberation sometimes does have this effect" [10]. This is the theoretical basis for deliberation tools like those in section 3: they may work by changing the structure of disagreement, not by finding a hidden consensus.
• Using more information. Letting voters grade options rather than rank them allows rules such as median grading, or Balinski and Laraki's majority judgment, that escape the impossibility. This works only if people share a common grading language; when their grading thresholds differ, a close relative of Arrow's result returns (Morreau 2016) [10]. In a 2012 interview Arrow himself acknowledged that grading changes the nature of voting [10]. Cardinal utilities alone are not enough. Sen showed that, in addition, "utilities have to be interpersonally comparable" [10]. Classical utilitarianism and Rawls's difference principle are two rules that become available once utilities are comparable [10].

Inference: This is the most important constraint on the thesis. An AI can search a huge space of designs, but every design must still be scored against a social objective, and choosing that objective (utilitarian sum, Rawlsian maximin, median grade, majority vote) is the political question the old ideologies were answering, in compressed form. Conservatism, liberalism and socialism can be read, roughly, as bundles of answers about which conditions to relax, whose utilities count and how much, and how far to trust local knowledge over central models. The debate does not disappear. It becomes a set of explicit parameters.

3. What AI system-design experiments have actually demonstrated

Since 2020 a handful of peer-reviewed or widely cited experiments have moved designing the system from theory into working code. Each shows a real capability, and each shows where the human value choice re-enters.

The AI Economist (Salesforce Research; preprint 2020, later in Science Advances). Zheng and colleagues built a two-level reinforcement-learning setup. Simulated worker agents learn to gather resources, trade and build in a gridworld economy, while a planner agent learns tax schedules that respond to how the workers adapt [2]. The researchers set the planner's objective as the product of equality and productivity. The preprint reports that the learned policy improved this trade-off by 16% over a tax framework derived from Emmanuel Saez's optimal-tax theory. The authors also tested policies with human participants recruited online [2]. What it shows: policy can be optimized against agents who adapt to the policy, which partly addresses the old objection that people change behaviour when rules change. What it does not show: that the result transfers to a real economy. The simulated economy is small and stylized, and the objective is one value choice among many.

Democratic AI, or human-centred mechanism design (DeepMind; Nature Human Behaviour, 2022). Koster and colleagues used an online public-investment game in which players contribute to a shared pot and a rule decides how returns are redistributed [1]. Instead of fixing a social welfare function, they made human preference the objective. A deep reinforcement-learning agent designed redistribution rules, and real players voted between the AI's rule and strong baselines, including strict egalitarian, libertarian and liberal-egalitarian schemes. The AI-designed rule won the majority vote [1]. It redistributed in proportion to players' contributions relative to their means, which corrected the starting wealth imbalance while discouraging free-riding [1]. What it shows: an AI can find a rule that beats the familiar ideological templates in a head-to-head vote, which is close to the thesis's idea of design replacing tribal contest. What it does not show: that majority approval is the right objective. Majority vote is itself an aggregation rule subject to section 2's constraints, and the authors acknowledge the risk of a design that a majority likes while a minority bears its costs [1].

The Habermas Machine (Google DeepMind; Science, October 2024). Tessler and colleagues fine-tuned a large language model to act as a mediator. It reads individual participants' written opinions on a divisive policy question, drafts a group statement, collects critiques, and revises [3]. Across experiments with more than 5,000 UK participants, the AI-written statements were preferred to those written by human mediators 56% of the time, and participants' positions converged toward a common view after the process [3]. The authors also ran a virtual citizens' assembly with a demographically representative sample, and report that the statements incorporated minority critiques rather than simply restating the majority view [3]. What it shows: LLMs can do the drafting work of deliberation at scale, the step that matches section 2's escape route of deliberation restructuring preferences. What it does not show: real policy outcomes. The process produced statements, not adopted laws.

Community Notes, originally Birdwatch (deployed on Twitter/X). This is the best-measured deployed bridging mechanism. Wojcik and colleagues at Twitter (arXiv, October 2022) describe a matrix-factorization algorithm that rates a crowd-written note highly only when raters from different viewpoints, as inferred from their past ratings, find it helpful. It does this by heavily regularizing each note's intercept, so a note scores well only if it is rated helpful by raters with a diversity of factor vectors [16]. In a survey experiment with Twitter users (Wave 3, April 2022), notes chosen by the algorithm reduced the likelihood of agreeing with the substance of a potentially misleading tweet by about 26% (N=6,046), with no statistically significant difference in effect between Democrats and Republicans [16]. Of respondents, 64% found the selected notes helpful, but the split by party was wide: 78.30% of Democrats and 48.22% of Republicans rated them at least somewhat helpful [16]. In live A/B tests between July 27 and September 27, 2022, in which 20% of US users were eligible to see notes, people exposed to notes were 25–34% less likely to like or retweet the annotated tweets [16]. The authors name the limit that matters most for this thesis: "By construction, the algorithm needs some cross-partisan agreement to function." On topics where no such agreement exists, the mechanism may find nothing to surface [16]. What it shows: a mechanism designed to reward agreement across groups, rather than majority size, works at platform scale and affects behaviour. What it does not show: that it reduces partisan animus, or that it works where there is no shared ground to find. The paper is written by the platform's own staff, so independent replication matters.

Polis and vTaiwan (deployed). The Computational Democracy Project's case study describes vTaiwan, a process launched by Taiwan's civic-tech community. It used Polis, an open-source tool that clusters participants by how they vote on short statements and surfaces statements that win support across clusters, as input to national regulation, with ride-hailing (Uber) regulation as the best-known case [5]. This is the clearest real-world example of a deployed bridging system: its design rewards cross-group agreement rather than majority size. What it shows: such mechanisms can feed real legislation. What it does not show: this report found no rigorous causal evaluation of how much they changed outcomes compared with ordinary consultation.

Inference across the four cases: Each system has three layers: (a) a representation of people (simulated agents, game players, written opinions, statement votes), (b) an optimizer (reinforcement learning, LLM drafting, clustering), and (c) an objective chosen outside the system (a welfare formula, a majority vote, endorsement, cross-cluster agreement). AI has made layer (b) cheap and powerful. Layers (a) and (c) are where the old political questions now live. A theory of everything for designed social systems would be a theory of how to choose and legitimate (c), and how to make (a) faithful.

The formal toolkits behind these experiments. Three bodies of work supply the theory these systems draw on, whether or not they cite it.

• Mechanism design. The 2007 economics Nobel went to Leonid Hurwicz, Eric Maskin and Roger Myerson for laying the foundations of mechanism design theory [11]. The Nobel committee's popular account describes it as the theory of designing institutions, from markets and auctions to voting procedures and public-goods provision, when participants hold private information and act in their own interest. It treats markets as one mechanism among many, a framing that can show when markets perform well and when other arrangements do better [11]. Hurwicz founded the field, Myerson developed the revelation principle that simplifies the search for good mechanisms, and Maskin developed implementation theory on when a mechanism can achieve a given social goal [11]. This is the closest existing formal version of what the thesis is reaching for: it does not ask whether capitalism or socialism wins, but which rules produce which outcomes given what people know and want. Koster et al. explicitly frame Democratic AI as mechanism design [1].
• Polycentric governance. Elinor Ostrom's Nobel lecture, published in the American Economic Review in June 2010 as Beyond Markets and States: Polycentric Governance of Complex Economic Systems, summarizes decades of research on how communities govern shared resources and public goods through institutional arrangements at multiple levels. That research explains outcomes that do not fit a two-way choice between the market and the state [12]. Her work is the strongest empirical evidence that the capitalism-or-socialism framing leaves out much of how people actually organize. Coverage note: this report read the journal's record of the article, not the full lecture text.
• Quadratic mechanisms. Buterin, Hitzig and Weyl's A flexible design for funding public goods (Management Science, 2019) extends quadratic voting into a mechanism for funding public goods without a fixed organizational structure, and claims approximately optimal provision [17]. It is an example of a new, explicitly designed alternative to both private markets and one-person-one-vote. Coverage note: the only version read here was an index page whose abstract is labelled as AI-generated, so this report relies on it for nothing beyond what the mechanism is.

4. Can AI simulate a society well enough to test designs before deploying them?

A natural extension of the thesis is to test candidate systems on simulated populations of LLM citizens before trying them on real people. The best-measured evidence says this is not yet reliable for inference.

Bisbee, Clinton, Dorff, Kenkel and Larson (Political Analysis, October 2024) prompted ChatGPT to adopt demographic personas and give feeling-thermometer scores for 11 sociopolitical groups. They then compared the answers with the 2016–2020 American National Election Study [6]. The averages lined up closely with the real survey. But "there is less variation in responses than in the real surveys", regression coefficients often differed significantly from those estimated on ANES data, small changes in prompt wording shifted the distribution, and "the same prompt yields significantly different results over a 3-month period" [6]. The authors conclude that their findings raise serious concerns about the quality, reliability and reproducibility of LLM-generated synthetic survey data [6].

For system design this failure mode is serious. A mechanism's behaviour depends heavily on the spread and correlation of preferences, that is, on minorities, outliers and strategic actors, not on averages. Arrow-type problems arise precisely from heterogeneity (section 2). A synthetic population that is too uniform will make a design look more consensual and stable than it is. The AI Economist's simulated workers (section 3) share this limit: they are reward-maximizers in a gridworld, not people [2].

Coverage note: Larger generative-agent simulations of real individuals (Park et al., 2024) report better fidelity, but this report could not access that paper and does not rely on it.

5. Is the ideological contest really in the rearview mirror? What the tribalism evidence says

Given your decision to treat the left/right fight as the wrong frame, the empirical question is what that fight actually consists of. Partisanship research gives a precise and somewhat awkward answer: much of the team-sports behaviour is not about systems at all.

Iyengar, Lelkes, Levendusky, Malhotra and Westwood's review in the Annual Review of Political Science (2019) separates two things. Ideological polarization means differences in policy positions. Affective polarization means dislike and distrust of the other party's people [7]. On the ANES feeling thermometer (0–100), the gap between how partisans rate their own party and the other party rose "from 22.64 degrees in 1978 to 40.87 degrees in 2016". The rise was driven by growing hostility to the other side rather than more warmth toward one's own, and the gap actually fell between 2012 and 2016 [7]. The authors argue affective polarization is "largely distinct from the ideological divide", that extreme issue positions are not required for it, and that "in some settings, affective polarization can increase while ideological divisions shrink" [7]. Ideological sorting accounts "for only about 5% of the variance in affective polarization", and over the past 40 years affective polarization rose by about the same amount among people with the most and the least ideologically consistent views [7].

The behavioural evidence supports the team-sports image directly. In a scholarship-selection experiment, 79.2% of Democrats picked the Democratic applicant and 80% of Republicans picked the Republican one, and the probability of choosing a better-qualified out-party applicant (4.0 vs 3.5 GPA) was never above 30% [7]. On an implicit association test, the partisan bias gap (D-score difference 0.50) was larger than the racial one (0.18) [7]. In 1960 only 4–5% of people were upset at the idea of their child marrying someone from the other party; by 2010 the figure was one-third of Democrats and one-half of Republicans. The review compares this with the 17 to 20% who would not want their child to marry a fan of a rival baseball team [7]. The authors trace the cause to partisanship working as a social identity (in-group/out-group dynamics in the Tajfel tradition), intensified by sorting, partisan media and campaigns. People were 50–150% more affectively polarized on election day than a year earlier [7].

Two findings bear directly on whether better design can defuse this. First, interventions that work in surveys are fragile. Correcting misperceptions about who belongs to the other party reduces animus, and priming shared national identity helps, but the authors warn that many sensible strategies will fail, and real-world transfer is unproven [7]. Second, on small-group discussion programmes between Democrats and Republicans, the authors write: "But there have been no systematic evaluations of these efforts, and it seems questionable whether they are scalable." [7]. That is the gap the section 3 tools are trying to fill, with AI providing the scale.

Inference: This evidence both supports and complicates the thesis. It supports the view that the public fight is largely tribal, more like team loyalty than a reasoned comparison of systems. But it also means the fight would not end if a better-designed system were offered, because the fight was never mainly about system design. A design thesis therefore has to include the identity layer as a design target: interfaces and processes, such as bridging algorithms and AI mediation, that lower the salience of team identity, not just better mechanisms. This evidence is also largely American, and the authors themselves flag how thin the comparative literature is [7].

6. Has "the ideological debate is over" been claimed before?

Yes, and the earlier claim is instructive. Daniel Bell's The End of Ideology (1960) argued that by the end of the 1950s, among Western intellectuals, the old ideological passions were spent [14]. His diagnosis of ideology is close to your own: ideologists are terrible simplifiers, and ideology spares people from confronting individual issues on their merits: they turn to an ideological "vending machine, and out comes the prepared formulae" [14]. His proposed replacement was also close to a design programme. He did not abandon utopia, but demanded that it be empirical rather than a matter of faith. A utopia, he wrote, must specify where one wants to go, how to get there, what the enterprise costs, and who is to pay, with a justification for that choice [14]. He also held that while the old debates were meaningless, some old verities were not: free speech, a free press, the right of opposition and free inquiry [14].

Within a decade the New Left, which Bell named in the same conclusion as having passion and energy but little definition of the future [14], had brought ideological conflict back to the centre of Western politics. Inference: Bell's claim was a description of a mood among intellectuals, not of the public, and it underestimated how much people want politics to supply identity and passion. That is the same gap section 5's polarization evidence exposes. A modern version of the thesis can still learn from Bell. His test (say where you want to go, how you get there, what it costs and who pays) is a good specification for any designed system, and it is one AI tools can now help fill in quantitatively. But any claim that the debate is over should expect the same rebuttal Bell received. Coverage note: Francis Fukuyama's 1989 The End of History?, the other major claim of this kind, was not read for this report.

3Beyond this question

• A sharper thesis. The evidence supports this claim: the old ideologies are compressed, implicit answers to design questions (whose knowledge counts, how preferences are aggregated, who bears risk), and AI makes it possible to state those answers explicitly and test them against each other. It does not support the claim that ideology is over. That framing has an awkward history (Medina shows Cybersyn's design carried its ideology [9]), and Arrow's result means the objective cannot be designed away [10].
• A useful vocabulary. Every designed social system can be described by its representation of people, its optimizer and its objective (section 3). The live research frontier is the first and third, not the second.
• The calculation debate becomes testable. Hayek's dispersed-knowledge argument [4] and the cybersocialist counter-claims [8] are now partly empirical questions about what AI systems can and cannot learn from data. That is a stronger footing than another round of in-principle argument.
• Legitimacy is a design variable. Democratic AI used votes as its objective [1], and the Habermas Machine optimized for endorsement [3]. Both treat public acceptance as part of the target, which is a different philosophy from the AI Economist's fixed welfare formula [2]. Choosing between those approaches is itself the modern version of the old split between democratic and technocratic thinking.

4Open questions

• No deployed AI-designed economic mechanism at scale has been evaluated. Every AI design result here is a simulation, a lab game, an online experiment, or a statement-drafting exercise [1][2][3]. Polis/vTaiwan is deployed but has no rigorous causal evaluation in the sources read [5].
• Whether computation dissolves the calculation problem is unsettled. The Austrian side claims it does not [8]; the cybersocialist side was not read in full here. A thesis should read Cockshott and Cottrell (1993) and Dapprich and Greenwood (2024) directly.
• Faithful synthetic populations. LLM respondents currently understate variation and drift over time [6]. Until that is solved, testing designs on simulated citizens will make them look more consensual than they are.
• Who chooses the objective, and how is that choice made legitimate? No source here offers a settled procedure, and Arrow's theorem says there is no neutral one [10].
• Does AI mediation reduce affective, tribal polarization, or only produce agreeable text? The Habermas Machine measured convergence on statements [3]. No study read here measured downstream effects on out-party animus of the kind Iyengar et al. track [7].
• Do bridging mechanisms fail exactly where they are most needed? Community Notes' designers state the algorithm needs some cross-partisan agreement to work [16]. How often the most divisive questions offer no such agreement has not been measured in the sources read.
• Can AI-assisted design respect Scott's metis? No study read here tests whether an AI-designed mechanism preserves or overrides the informal local practices that Scott argues formal orders depend on [15].
• Coverage gaps in this report. Read only in part or by abstract: Ostrom's 2010 lecture (journal record only) [12], Dapprich and Greenwood (abstract only) [13], Scott (a published excerpt, not the book) [15], and quadratic funding (an index page) [17]. Not read: Cockshott and Cottrell's Towards a New Socialism, Fukuyama's The End of History?, Weyl and Tang's Plurality, and Park et al.'s generative-agent simulations.

5Sources

1. Human-centred mechanism design with Democratic AI — Koster et al., Nature Human Behaviour, 2022
2. The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies — Zheng et al., Salesforce Research, arXiv preprint, 2020 (later published in Science Advances, 2022)
3. AI can help humans find common ground in democratic deliberation — Tessler et al., Science, October 2024 (copy hosted by rivista.ai)
4. The Use of Knowledge in Society — F. A. Hayek, American Economic Review, 1945 (Library of Economics and Liberty edition)
5. vTaiwan case study — The Computational Democracy Project
6. Synthetic Replacements for Human Survey Data? The Perils of Large Language Models — Bisbee, Clinton, Dorff, Kenkel and Larson, Political Analysis 32(4), October 2024 (IDEAS/RePEc record)
7. The Origins and Consequences of Affective Polarization in the United States — Iyengar, Lelkes, Levendusky, Malhotra and Westwood, Annual Review of Political Science 22, 2019
8. Economic Calculation in Light of Advances in Big Data and Artificial Intelligence — Lambert and Fegley, Journal of Economic Behavior & Organization 206, February 2023
9. Designing Freedom, Regulating a Nation: Socialist Cybernetics in Allende's Chile — Eden Medina, Journal of Latin American Studies 38(3), August 2006
10. Arrow's Theorem — Michael Morreau, Stanford Encyclopedia of Philosophy, first published 2014, revised December 2025
11. The Prize in Economic Sciences 2007 — Popular information — Royal Swedish Academy of Sciences / NobelPrize.org
12. Beyond Markets and States: Polycentric Governance of Complex Economic Systems — Elinor Ostrom, American Economic Review 100(3), June 2010 (journal record)
13. Cybersocialism and the Future of the Socialist Calculation Debate — Dapprich and Greenwood, Erasmus Journal for Philosophy and Economics 17(1), July 2024
14. The End of Ideology, Conclusion (excerpt) — Daniel Bell, 1960; excerpt hosted by Al Filreis, University of Pennsylvania
15. Excerpt from Seeing Like a State — James C. Scott, Yale University Press 1998; excerpt published by the Schumacher Center for a New Economics
16. Birdwatch: Crowd Wisdom and Bridging Algorithms can Inform Understanding and Reduce the Spread of Misinformation — Wojcik et al., Twitter, arXiv, October 2022
17. A flexible design for funding public goods — Buterin, Hitzig and Weyl, Management Science 65(11), 2019 (index page by Pablo Stafforini)