{"path":"research/ai-augmented-cooperation-infrastructure.md","content":"# AI-Augmented Cooperation Infrastructure\n\nResearch compiled April 6, 2026. Eight angles on how AI, structured deliberation, and epistemic infrastructure intersect — and what has been proven to work empirically.\n\n---\n\n## 1. Cooperative AI Foundation (Dafoe et al.)\n\n**The institution.** The Cooperative AI Foundation (CAIF) is a UK registered charity (#1201294) backed by a **$15 million philanthropic commitment from Macroscopic Ventures**. Board: Allan Dafoe (DeepMind/GovAI), Audrey Tang (Taiwan's first digital minister), Thore Graepel (DeepMind), Gillian Hadfield (Johns Hopkins), Jesse Clifton (Macroscopic Ventures). Mission: \"support research that will improve the cooperative intelligence of advanced AI for the benefit of all.\" Prioritizes grants by importance, neglectedness, and tractability.\n\n**The framework.** The foundational paper (\"Open Problems in Cooperative AI,\" Dafoe et al., NeurIPS 2020) identifies four cooperative capabilities:\n\n1. **Understanding** — agents must comprehend \"other agents, their beliefs, incentives, and capabilities\"\n2. **Communication** — building shared language while \"overcoming mistrust and deception\"\n3. **Commitment** — creating cooperative arrangements that overcome incentives to defect\n4. **Institutions** — social structures from informal norms to formal legal systems that promote cooperation\n\nThe framework explicitly addresses **mixed-motive dynamics** — situations where groups have conflicting interests but potential mutual benefits. Pure common-interest scenarios (everyone wants the same thing) are trivial; the hard problem is cooperation under partial conflict.\n\n**Key insight for Deliberus**: Dafoe emphasizes that \"cooperative competence can be used to exclude others\" — the same capabilities enabling cooperation also enable collusion, manipulation, and coercion. This dual-use concern mirrors Deliberus's analysis-vs-attunement dialectic: analytical tools for decomposing arguments could also be weaponized for sophisticated manipulation. The foundation explicitly funds research on measuring and mitigating these risks.\n\n**Grants awarded** (17 projects, selected highlights):\n- **FOCAL Lab** (Conitzer, CMU) — $500K for game theory foundations of cooperative AI\n- **Policy Aggregation** (Procaccia, Harvard) — $233K for aligning AI with groups using \"voting and Nash welfare\" — directly relevant to Deliberus's aggregation challenge\n- **Cooperation and Negotiation in LLMs** (Jacques/Levine, UW/Berkeley) — $450K studying LLM cooperative capabilities including deception and agent modeling\n- **AI for Humanitarian Crisis Negotiation** (Doshi Velez, Harvard) — $214K on multi-party negotiations using coalition game frameworks\n- **Emergent Norms for Sustainable Cooperation** (Kleiman-Weiner, UW) — $294K on how cooperative norms emerge through communication in LLM and RL agents\n- **AI Coercive Capabilities** (Hatz, Uppsala) — 640K SEK measuring coercive capabilities vs. cooperation benefits\n\n**Funding area directly matching Deliberus**: \"AI for Facilitating Human Cooperation\" — seeks proposals developing AI tools that help resolve major cooperation challenges in mixed-motive settings. Prioritizes collective decision-making tools, negotiation/conflict resolution AI, and institutional design. Climate, armed conflict, democratic systems, and AI governance are priority domains. Out of scope: general productivity tools, social media, healthcare unless addressing fundamental cooperation challenges.\n\n**Relevance**: Deliberus's persistent argument graphs + bridging detection + structured decomposition fit squarely in the \"AI for Facilitating Human Cooperation\" grant area. The worldview-filter UX is institutional design for epistemic cooperation. The convergence thesis maps to the foundation's interest in whether cooperation scales with capability.\n\n---\n\n## 2. The Procaccia/Konya Peacebuilding Paper (2025)\n\n**Citation**: Konya, Thorburn, Almasri, Leshem, Procaccia, Schirch, Bakker. \"Using Collective Dialogues and AI to Find Common Ground Between Israeli and Palestinian Peacebuilders.\" arXiv:2503.01769, ACM FAccT 2025.\n\n**The experiment.** April-July 2024, during the most severe phase of the Israel-Gaza conflict. ~100 peacebuilders in uninational phase, ~120 in joint phase. Three groups: Israeli Jews, Palestinian citizens of Israel, Palestinians from West Bank/Gaza.\n\n**Methodology** (four phases):\n1. **Collective dialogue**: Anonymous responses to prompts + agreement/pair-choice votes. Matrix completion converts sparse votes into complete participant vote matrices\n2. **Bridging-based ranking**: Two metrics — max-min agreement across groups, and the Community Notes latent factor algorithm — to surface cross-group common ground\n3. **LLM distillation**: GPT-4 extracts unique ideas from bridging statements, generates collective statements preserving original language and cultural nuances\n4. **Final validation**: Participants vote on collective statements (5-point Likert + ranking). Equal-power metrics ensure each group has equivalent influence regardless of numbers\n\n**Results:**\n- Five demands to world leaders: **at least 90% agreement** across all sides\n- \"An immediate ceasefire and release of all hostages\" — universally ranked first\n- Five messages to residents: **at least 84% agreement** across all sides\n- Six shared values discovered: peace, equality, human life, independence, safety, prosperity\n\n**Critical design features:**\n- **Linguistic integrity**: Hebrew and Palestinian Arabic monolingually in uninational phases; trilingual (Hebrew/Arabic/English) in joint phase. LLM prompts tested in native languages for optimal performance\n- **Equal-power metrics**: Max-min agreement (minimum support across ALL groups), Dowdall scores averaged independently per group, instant runoff voting with equal-power elimination\n- **Trust architecture**: Uninational phase precedes joint dialogue — groups clarify their own positions before engaging across conflict lines\n\n**Unexpected finding**: \"Red lines\" showed mutual alignment — Palestinian outgroup red lines (e.g., \"call for forced displacement\") corresponded to Israeli ingroup red lines (e.g., \"talk about deporting Palestinians\"). The groups were closer than expected on what they considered unacceptable.\n\n**Why this is an existence proof**: The paper demonstrates that AI-assisted structured deliberation can produce substantive agreement (84-90%) among parties in the most intractable conflict imaginable, during active hostilities. The authors note this was \"unlikely\" by prevailing expectations. If this works for Israeli-Palestinian peacebuilders during a war, it works.\n\n**Limitations**: Self-selected peacebuilder sample (not general population), no controlled experiment (real-world constraints), case study not research design, potential translation artifacts.\n\n**Relevance to Deliberus**: The bridging-based ranking + LLM distillation pipeline is remarkably close to Deliberus's extraction + auto-connect + bridging detection architecture. Deliberus adds persistent argument structure (WHY people agree, not just THAT they agree) and recursive decomposition (the claims beneath the bridging statements). The Procaccia/Konya work validates the outer loop; Deliberus could provide the inner structure.\n\n---\n\n## 3. Computational Social Choice\n\n**Arrow's impossibility theorem** (1951): No rank-order voting system can simultaneously satisfy unanimity, independence of irrelevant alternatives, and non-dictatorship. This means ANY aggregation mechanism for ranking preferences will violate at least one desirable property. The theorem shapes all subsequent social choice theory.\n\n**The deliberative escape from Arrow**: Dryzek and List (\"Social Choice Theory and Deliberative Democracy: A Reconciliation,\" British Journal of Political Science 33(1), 2003) argue that deliberation can reduce the severity of Arrow's impossibility by structurally altering the preference profile. Deliberation that produces **single-peaked preferences** (where voters agree on the underlying dimension of disagreement) eliminates the cycling problem. Median voter theorem then applies cleanly. The argument: deliberation doesn't change the voting rule — it changes the INPUT to the voting rule, making consistent aggregation possible.\n\n**Adachi, Chung, and Kurihara** (2024, AJPS) prove \"The Impossibility of Deliberation-Consistent Social Choice\" — formalizing that even deliberation cannot fully escape impossibility results if consistency with group reasoning is required. The tension between aggregative and deliberative ideals is mathematically fundamental.\n\n**Schulze method**: Condorcet-consistent ranked voting used by Wikimedia Foundation, Debian, Gentoo, and many Pirate Parties. Finds the \"strongest path\" between all candidate pairs in pairwise comparisons. Properties: Condorcet winner selection, monotonicity, independence of clones, reversal symmetry. Adopted specifically because it handles nuanced multi-option decisions better than simple majority. Used at scale by Debian (1000+ developers) and Wikimedia. Computational complexity: O(C^3) for C candidates, with recent work on large-scale efficiency (Csar, Lackner, Pichler).\n\n**What social choice function best serves deliberation?** The research suggests this is the wrong question. Deliberation and aggregation serve different functions:\n- **Aggregation** (voting) takes fixed preferences and combines them — subject to Arrow constraints\n- **Deliberation** transforms preferences through reason-giving — can produce the preference structures that make good aggregation possible\n\nThe most promising approach: use deliberation to structure the problem space (decompose, clarify, identify shared values), THEN apply appropriate aggregation to residual disagreements. Schulze or similar Condorcet methods for multi-option ranking; bridging metrics for identifying cross-group consensus; QEM gradual semantics for argument strength.\n\n**Relevance to Deliberus**: Deliberus's architecture implicitly addresses Arrow by operating at the deliberation layer, not the aggregation layer. Decomposing claims into atomic subclaims + discovering shared value premises via the convergence thesis is precisely the preference-structuring that Dryzek and List argue makes aggregation tractable. The QBAF/QEM strength scoring provides continuous gradual semantics rather than binary vote aggregation — sidestepping some impossibility results.\n\n---\n\n## 4. AI Mediation Research\n\n### The Habermas Machine (Tessler, Evans, Bakker et al., Science 2024; expanded arXiv 2601.05904, Jan 2026)\n\nGoogle DeepMind's landmark experiment. n=2,110 UK participants. 15-minute deliberation protocol per question: participants submit opinions on controversial topics, receive AI-generated group statements, rank them, provide critiques, rank revised statements.\n\n**Key findings:**\n- AI-generated statements \"garnered levels of endorsement from the participants exceeding that of human mediators\"\n- Multiple rounds led to greater approval; deliberation \"left groups less divided on the issue than they were prior to deliberation\"\n- The system \"fairly represented majority and minority viewpoints\" proportionally\n- Participants \"felt they had an impact on the final outcome\" — perceived legitimacy maintained\n\n**Limitation**: Operates on opinion statements, not structured arguments. No decomposition, no scheme detection, no persistent graph. It's consensus synthesis, not reasoning infrastructure.\n\n### AI Reflectors (Revel & Penigaud, 2025; \"AI-Enhanced Deliberative Democracy and the Future of the Collective Will,\" arXiv:2503.05830; also \"AI-Facilitated Collective Judgements,\" SSRN)\n\nRevel (Harvard) and Penigaud (Yale) provide a taxonomic analysis of computational frameworks for finding common ground. They unpack design choices behind longstanding and new approaches, examining potential for a \"collective will\" concept.\n\nTheir companion paper at the AI Objectives Institute emphasizes the **augment-don't-automate principle**: \"AI should support but not replace the human part of deliberation: the collective reasoning and thinking together.\" Two essential dimensions:\n- **Explainability**: Participants must understand and articulate reasoning behind AI outputs\n- **Steerability**: Communities need capacity to \"intervene, correct, or overwrite AI outputs\"\n\nThey introduce the concept of **opinion commitment** — the spectrum from surface preferences to deeply-held views. Higher-commitment outputs (policy recommendations) require proportionally higher human agency. A system that generates profound ethical conclusions needs more robust human oversight than one that sorts restaurant preferences.\n\n### DCI: From Debate to Deliberation (Prakash, March 2026; arXiv:2603.11781)\n\nPrakash's paper directly challenges multi-agent debate (Irving et al.) as the dominant LLM reasoning approach. Core argument: debate discards disagreements, lacks convergence guarantees, and scales poorly. DCI introduces **14 typed epistemic acts** (assert, challenge, refine, synthesize, etc.) within an interaction grammar. Four archetyped delegates operate through phased sessions using shared workspaces and a convergence algorithm.\n\n**Empirical results:**\n- On hidden-profile tasks (fragmented knowledge requiring synthesis): DCI scores **9.56** — the study's highest score\n- On non-routine tasks: DCI outperforms debate by **+0.95 points** (significant)\n- On routine tasks: DCI scores only 5.39 (underperforms baselines) — structure adds overhead for simple problems\n- Cost: DCI consumes **~62x tokens** of single-agent approaches — justified only when decision transparency matters\n\n**Key output**: DCI produces \"decision packets\" containing selected options, residual objections, minority reports, and reopening conditions — structured artifacts that unstructured debate cannot generate.\n\n**Relevance to Deliberus**: DCI's typed epistemic acts parallel Walton's argument schemes. The decision packet structure (with minority reports and reopening conditions) maps to Deliberus's sorry markers + CQ layer. The finding that structure helps on hidden-profile tasks but hurts on routine tasks validates Deliberus's progressive disclosure approach — show structure when it matters, hide it when it doesn't.\n\n### AI Negotiation Enhancement (Tech and Social Cohesion, Jan 2026)\n\n\"Negotiators using AI achieve 48% better outcomes\" — empirical finding from recent mediation research. The mechanism: AI helps negotiators explore option spaces more systematically, not by replacing human judgment but by expanding the consideration set.\n\n---\n\n## 5. Digital Public Goods and Epistemic Infrastructure\n\n### Decidim\n\n**Adoption**: 490 active instances, 32 countries, 309 institutions, 925,152 participants, 100,129 proposals, 120,554 comments, 11,215 meetings. Major users: Barcelona, NYC, Helsinki, Lyon, plus universities, NGOs, cooperatives. Open-source (Ruby on Rails), born from Barcelona's 15M movement (2015).\n\n**What it does well**: Participatory budgeting, public consultations, citizen assemblies. Strong institutional adoption in European municipalities. Robust governance structure with an elected \"meta-decidim\" community.\n\n**What it misses**: No argument structure. No reasoning quality assessment. No bridging detection. Proposals are flat text, voted on without decomposition. It's a participation platform, not a deliberation platform — it collects opinions but doesn't analyze their logical structure or find hidden consensus.\n\n### Consul (Consul Democracy)\n\nOpen-source (also Ruby on Rails), developed by Madrid City government. Used by 100+ cities worldwide. Similar to Decidim: proposals, debates, polls, participatory budgets. A 2025 EU case study found it effective for basic participation but limited in deliberative depth.\n\n### Loomio\n\nBorn from Occupy Wellington (2011). Cooperative structure. Asynchronous deliberation with structured decision methods (consent, advice, consensus). Used by Enspiral network and various cooperatives. Key limitation identified by co-founder Richard Bartlett: the \"problem of deliberation\" — online text deliberation lacks the embodied trust-building of in-person dialogue. Loomio works well for small groups with existing trust; struggles to scale to strangers or adversaries.\n\n### Polis\n\n**The success story.** Over 10 million participants since 2012. Used in Taiwan (vTaiwan), UK, Finland, Singapore, Philippines, Austria (Climate Citizens Council). Uses PCA and k-means clustering to visualize opinion distributions, then identifies \"bridging statements\" supported across clusters. Eliminates reply buttons entirely — structurally prevents flame wars.\n\n**vTaiwan results**: Over a dozen bills influenced. Uber regulation: initial fierce conflict resolved, 95% agreement on \"passenger safety\" concern. 80% of vTaiwan consultations led to government action.\n\n**Why Polis succeeded where others struggled**: (1) No replies = no flame wars (structural, not moderation-based); (2) bridging algorithm makes invisible consensus visible; (3) participation cost is low (vote agree/disagree/pass, no writing required); (4) institutional embedding in Taiwan's governance.\n\n**What Polis misses**: No argument structure. Tells you THAT groups agree, not WHY. No decomposition into premises. No persistent reasoning graph. Bridging statements are opaque — you know they bridge, but not what logical moves make them bridge.\n\n### Noosphere (Gordon Brander)\n\nAmbitious protocol for decentralized knowledge — self-sovereign data and credible neutrality. **Shut down in 2024.** The vision: protocol-level infrastructure for collective thinking, not controlled by any stakeholder. The failure: building infrastructure without a concrete use case or user base.\n\n### deliberation.io (Stanford)\n\nOpen-source, open-science platform for democratic engagement at scale. Focused on facilitating deliberation with AI assistance. Stanford Digital Economy Lab. Published at 2025 academic conferences. Early stage but academically grounded.\n\n### The Epistemic Pipeline Problem\n\nThe Gitcoin collective intelligence research identifies a critical failure pattern: most deliberation tools skip the **sensemaking stage**, jumping directly to voting/allocation without building shared understanding. The correct pipeline:\n1. Information gathering\n2. Sensemaking / shared understanding (most tools skip this)\n3. Preference expression\n4. Decision / allocation\n\nDeliberus operates primarily at stage 2 — the most neglected and most valuable stage. The extraction pipeline + decomposition + concept clarification IS the sensemaking infrastructure that other tools assume already exists.\n\n### Global Digital Compact (UN, September 2024)\n\nAdopted by UN member states. Commits to treating digital technologies as \"digital public goods\" — open-source, open-data, open-standards. Specific commitments: advance responsible use of AI, protect human rights online, strengthen digital cooperation. Notably light on epistemic infrastructure specifically — focuses on access, safety, and governance rather than reasoning quality. The gap: institutional recognition of digital public goods without epistemic infrastructure to ensure those goods serve truth-seeking.\n\n### Why Most Failed and What Would Make One Succeed\n\n**Common failure modes:**\n1. **No argument structure** — all opinions are flat. Quality invisible. Bad arguments count the same as good ones\n2. **No persistent graph** — each consultation starts from zero. No accumulation of shared understanding\n3. **No bridging detection** — platforms surface popular views, not cross-group consensus\n4. **No decomposition** — bundled claims can't be separately evaluated\n5. **Institutional orphaning** — tools built without governance connection; outputs go nowhere\n6. **Scale collapse** — deliberation quality degrades with participant count (Loomio's problem)\n\n**What would succeed** (the Deliberus thesis): persistent argument graphs that accumulate across conversations, AI-driven decomposition that makes quality visible, bridging signals that surface hidden consensus, worldview filters that enable perspective-taking, and progressive disclosure that scales from simple to arbitrarily deep. The platform must be both a personal thinking tool (adoption driver) and a civilizational infrastructure (purpose driver).\n\n---\n\n## 6. Metacrisis and Sensemaking Solutions\n\n### Schmachtenberger's Framework\n\nThe metacrisis is not any single crisis but the **generator functions** that produce them all:\n\n- **Multipolar traps**: Individual rational actors pursuing self-interest produce collectively catastrophic outcomes. Each fisherman benefits from catching more fish; coordinated overfishing destroys the resource. \"No individual agent can unilaterally stop\" without competitive disadvantage\n- **Perverse incentives**: \"Almost every dollar in the global economy carries externalities\" — markets reward extraction while penalizing restraint\n- **Exponential technology amplification**: Advanced tools magnify competitive pressures, creating race dynamics where cutting safety corners becomes necessary for survival\n\nThree attractors: **catastrophe** (cascading failures), **dystopia** (comprehensive surveillance/control), or the **third attractor** — distributed coordination enabling \"agent-centric self-organization\" with localized resilience and global coordination capacity.\n\n### The Consilience Project\n\nFounded by Schmachtenberger as a publication of the Civilization Research Institute (CRI). Aimed to \"describe the state of our information commons and uncover the roots of the challenges facing open societies.\" Published collectively-authored articles on global risks, governance design, and cultural challenges.\n\n**Current status**: No longer in active production. \"Although The Consilience Project is no longer releasing new articles at a regular cadence, it will continue to publish related content at varying intervals.\" No explicit explanation for reduced activity — likely resource constraints. CRI continues as parent organization.\n\n**What it got right**: Diagnosis of information commons degradation, framing of the metacrisis as coordination failure, emphasis on sensemaking as prerequisite to solution.\n\n**What it missed**: Produced analysis and essays, not infrastructure. Published insights but didn't build tools that embody those insights. The sensemaking stayed in long-form prose, never became interactive, computable, or decomposable. It was a magazine about the need for better sensemaking, not a sensemaking tool.\n\n### Game B vs Game A\n\n**Game A**: The current civilizational operating system — rivalrous dynamics, zero-sum competition, resource extraction, short-term optimization. Produces the generator functions above.\n\n**Game B**: The hypothesized successor — anti-rivalrous dynamics where sharing and cooperation increase returns. Not a specific system but a design constraint: any viable Game B institution must make cooperation individually rational, not just morally aspirational.\n\n**Anti-rivalrous epistemic infrastructure** would be: a system where contributing knowledge makes YOUR understanding better (not just others'), where disagreement produces insight (not just conflict), where the more people use it the more valuable each person's contribution becomes. This is precisely the `@[simp]` flywheel in Deliberus — each extraction discovers connections to the existing graph, making the whole graph more valuable for everyone.\n\n**The key unsolved problem**: How do you bootstrap anti-rivalrous dynamics in a rivalrous world? Schmachtenberger's framework identifies the problem but not the adoption mechanism. Deliberus's answer: start as a personal thinking tool (rivalrous benefit — I want to understand this article better), accumulate into collective infrastructure (anti-rivalrous benefit — the graph gets better as more people use it).\n\n---\n\n## 7. AI Alignment Through Deliberation\n\n### Debate as Alignment (Irving et al., 2018; ongoing UK AISI work)\n\n**Core mechanism**: Two AI agents alternately present arguments about a problem's solution. A human judge selects which agent is more truthful. The key asymmetry: \"it is harder to lie than to refute a lie.\" An honest debater can pinpoint exactly where a dishonest argument fails. A dishonest debater must construct a globally consistent false narrative — substantially harder.\n\n**Judge enhancement without judge improvement**: Debate achieves \"complexity amplification\" — a polynomial-time judge can reach correct conclusions for PSPACE-complex problems through polynomially many debate steps. The judge doesn't become smarter; the adversarial structure forces agents to surface critical evidence.\n\n**Empirical results** (DeepMind, 2024): Tested across nine task domains with ~5 million model generation calls. \"Debate outperforms consultancy across all tasks\" when consultants are randomly assigned correct/incorrect positions. Critical scaling property: \"Stronger debaters lead to higher judge accuracy, including for a weaker judge\" — as systems become more capable, debate becomes MORE effective.\n\n**MNIST demonstration**: Agents competing to convince a sparse classifier by revealing pixels \"boosted the classifier's accuracy from 59.4% to 88.9% given 6 pixels\" — showing how structured argumentation reveals informative evidence.\n\n**UK AISI Safety Case** (Irving, Pfau, Hilton, May 2025): Sketch of a full alignment safety case based on debate. Addresses \"dodging systematic human errors in scalable oversight.\" Introduces Prover-Estimator Debate as a new protocol. Key finding: \"legibility to smaller LLMs transfers to legibility to humans\" — making AI reasoning legible to weaker models is a viable proxy for human oversight.\n\n**Remaining vulnerabilities**: LLM judges exhibit \"12 distinct bias types\" that could destabilize the honest equilibrium. \"Deception intensifies as the capability gap between weak and strong models increases.\" Non-verifiable domains (long-horizon planning, aesthetic judgment) remain unaddressed.\n\n### The CEV Connection\n\nYudkowsky's Coherent Extrapolated Volition (2004, MIRI) proposes aligning superintelligent AI to act according to \"our wish if we knew more, thought faster, were more the people we wished we were, had grown up farther together.\" This is explicitly a deliberative ideal — what humanity would conclude given perfect information and unlimited reasoning time.\n\n**The structural parallel**: CEV requires (1) aggregating diverse human values, (2) extrapolating them under improved reasoning, (3) finding coherence among the extrapolated versions. Deliberus's architecture addresses all three: (1) extraction captures diverse viewpoints, (2) decomposition + CQ generation improves the reasoning, (3) the convergence thesis claims values converge when decomposed far enough.\n\n**The practical gap**: CEV is a specification for a superintelligent AI. Deliberus is a tool for humans. But if the convergence thesis is empirically validated — if values DO converge when decomposed and examined — then Deliberus becomes an approximation of CEV using human intelligence augmented by AI, rather than requiring superintelligence. This is the alignment argument: structured deliberation infrastructure is \"judge enhancement infrastructure\" that could approximate what a fully aligned AI would compute.\n\n### Constitutional AI and Deliberative Alignment (Anthropic, 2023-ongoing)\n\nAnthropic's Constitutional AI replaces human evaluation with principle-based self-improvement, achieving \"Pareto improvement over standard RLHF on helpfulness vs. harmlessness.\" The \"constitution\" is itself a product of deliberation — principles extracted from human rights documents and ethical frameworks.\n\n**The connection to deliberation infrastructure**: If the constitution governing AI behavior should itself be deliberatively determined (as Anthropic acknowledges), then infrastructure for structured deliberation over ethical principles is upstream of AI alignment. Deliberus's decomposition of value premises into challengeable subclaims is precisely the mechanism needed to deliberate about what AI constitutions should contain.\n\n---\n\n## 8. Community Notes (Twitter/X)\n\n### How It Works\n\nCommunity Notes uses a **latent factor model** (matrix factorization) to identify and discount predictable voting patterns. Rather than ranking notes by upvote count:\n\n1. Users vote \"helpful\" or \"not helpful\" on notes attached to posts\n2. The algorithm discovers latent dimensions in voting behavior (typically political leaning)\n3. It identifies clusters of users who consistently vote the same way\n4. It promotes notes receiving support **despite** partisan leanings, not because of them\n\nThe critical mechanism: \"the algorithm boosts Notes which receive support despite, or disregarding, the political leaning of voting users.\" Cross-partisan agreement on a note's helpfulness is treated as evidence of genuine factual accuracy.\n\n### Effectiveness Evidence\n\n**PNAS study** (Slaughter et al., September 2025): Analyzed 40,078 posts from X (March-June 2023), of which 6,757 received helpful community notes.\n\nAfter a note is attached (48-hour window):\n- **Reposts**: -46.1% growth reduction\n- **Likes**: -44.1% growth reduction\n- **Structural virality**: -48.5% growth reduction\n- **Maximum cascade depth**: -39.9% growth reduction\n\nNotes fundamentally alter HOW content spreads: noted posts show \"less deep and less viral\" diffusion patterns — they specifically reduce person-to-person resharing chains.\n\n**Timing matters critically**: Notes within 12 hours: -49.6% repost reduction. Notes after 47+ hours: -6.2%. Speed is everything.\n\n**COVID note accuracy**: 97% accuracy according to medical professionals.\n\n**Deletion effect**: Notes increase probability that the original author deletes the post by 80%.\n\n### What It Gets Right\n\n1. **Structural bridging**: Cross-partisan agreement required — not majority vote. This is genuinely novel at scale\n2. **No editorial board**: Decentralized, no single entity decides what's true\n3. **Measurable impact**: Repost reduction, deletion increase, cascade disruption — real behavioral change\n4. **Open-source algorithm**: Fully transparent, auditable\n\n### What It Misses\n\n1. **No argument structure**: Binary \"helpful/not helpful\" — no decomposition of WHY a claim is wrong. A note might say \"this is misleading because X\" but the reasoning structure isn't formalized or reusable\n2. **Coverage crisis**: Only 29% of fact-checkable tweets received helpful notes. During three days before the 2024 election, fewer than 6% of notes reached \"helpful\" status. The system is too slow for the information environment\n3. **Speed problem**: Typical notes take 7+ hours to appear; some take 70 hours. Most misinformation damage occurs in the first hours\n4. **Partisan gaming vulnerability**: \"Groups of bad faith actors can skew the algorithm\" by targeting credible source classifications\n5. **Single-dimension assumption**: The latent factor model assumes the primary dimension is political. If the main split is expertise vs. non-expertise, the algorithm might \"actively work to disregard expertise in votes\"\n6. **No persistence**: Each note is an island. No accumulated understanding across related claims. No graph of related arguments. No memory\n7. **No why**: Tells you a claim has cross-partisan skepticism, not what the underlying reasoning structure looks like\n\n**The Deliberus extension**: Community Notes validates that bridging-based scoring works at scale. Deliberus adds the missing WHY — decomposing claims into premises, detecting argument schemes, generating critical questions, and building persistent graphs where bridging signals accumulate and become searchable. Community Notes is a checkpoint (\"this claim is contested\"); Deliberus is the investigation (\"here's what the argument looks like when you unfold it\").\n\n---\n\n## Cross-Cutting Synthesis\n\n### What Has Been Proven Empirically\n\n| Finding | Source | Strength |\n|---------|--------|----------|\n| AI-mediated deliberation exceeds human mediation quality | Habermas Machine (n=2,110) | Strong (RCT) |\n| 84-90% agreement achievable in Israeli-Palestinian conflict | Procaccia/Konya (n=~120) | Moderate (case study) |\n| Cross-partisan bridging reduces misinformation spread by ~46% | Community Notes (n=40,078 posts) | Strong (quasi-experimental) |\n| Bridging-based deliberation → 80% government action rate | Polis/vTaiwan | Moderate (institutional) |\n| Structured deliberation can produce single-peaked preferences | Dryzek & List (2003) | Theoretical + case evidence |\n| Typed epistemic acts outperform debate on hidden-profile tasks | DCI/Prakash (2026) | Moderate (benchmarks) |\n| Stronger AI debaters → higher judge accuracy | DeepMind debate study (2024) | Strong (large-scale empirical) |\n| AI negotiation assistance → 48% better outcomes | Mediation research | Moderate |\n\n### What Has Not Been Built\n\nDespite the above evidence, **no existing system combines**:\n1. Persistent argument graphs (knowledge accumulates)\n2. AI-driven decomposition (claims → atomic premises)\n3. Bridging detection at the reasoning level (WHY groups agree, not just THAT)\n4. Worldview filters (navigate the same graph from different perspectives)\n5. Gradual semantics (continuous strength scoring, not binary votes)\n6. Progressive disclosure (simple for casual users, deep for committed ones)\n7. Self-similar decomposition (the system's own representations are challengeable)\n\nPolis does (3) at the opinion level but not the reasoning level. Community Notes does (3) at the factual level but not the argument level. The Habermas Machine does consensus synthesis but not persistent accumulation. Decidim does institutional participation but none of (1-7). DCI does (2) and (5) for AI agents but not as human-facing infrastructure.\n\n### The Funding Landscape\n\nThree active funding sources directly relevant to Deliberus:\n\n1. **Cooperative AI Foundation** — $15M, grants $10K-$500K, \"AI for Facilitating Human Cooperation\" area. Deliberus fits the mixed-motive institutional design category\n2. **Coefficient Giving (formerly Open Philanthropy) Forecasting Fund** — was $8-10M, now closed for this specific RFP, but explicitly sought \"argument analyzers\" and \"arbitrators facilitating debate between opposing viewpoints.\" The fund is winding down but the program area (AI for sound reasoning) persists in their broader portfolio\n3. **Various EA/rationalist grant programs** — 80,000 Hours alignment research, LTFF, Survival and Flourishing Fund. Deliberus's AI safety angle (deliberation as alignment infrastructure) is a fit\n\n### The Key Insight\n\nThe empirical evidence shows that structured, bridging-based deliberation works — in active conflict zones, at national scale, on the most partisan social media platform. What doesn't exist is infrastructure that makes this CUMULATIVE. Every experiment starts from zero. The Procaccia/Konya peacebuilders generated brilliant consensus statements that exist as a PDF. Community Notes generates millions of factual corrections that exist as isolated annotations. Polis surfaces bridging statements that disappear when the consultation closes.\n\nDeliberus's distinctive contribution is the persistent, decomposable, accumulable graph — the argument infrastructure that turns each deliberation from an event into a building block. This is the difference between a sensemaking tool and sensemaking infrastructure.\n\n---\n\n## References\n\n- Dafoe, A., Hughes, E., Lanctot, M., et al. \"Open Problems in Cooperative AI.\" NeurIPS 2020. arXiv:2012.08630\n- Conitzer, V. & Oesterheld, C. \"Foundations of Cooperative AI.\" AAAI 2023\n- Konya, A., Thorburn, L., et al. \"Using Collective Dialogues and AI to Find Common Ground Between Israeli and Palestinian Peacebuilders.\" arXiv:2503.01769, ACM FAccT 2025\n- Tessler, M.H., Evans, G., Bakker, M.A., et al. \"Can AI Mediation Improve Democratic Deliberation?\" arXiv:2601.05904, January 2026 (expanded from Science 2024 publication)\n- Prakash, S. \"From Debate to Deliberation: Structured Collective Reasoning with Typed Epistemic Acts.\" arXiv:2603.11781, March 2026\n- Revel, M. & Penigaud, T. \"AI-Enhanced Deliberative Democracy and the Future of the Collective Will.\" arXiv:2503.05830, March 2025\n- Revel, M. & Penigaud, T. \"AI-Facilitated Collective Judgements.\" SSRN, January 2025\n- Irving, G., Christiano, P., Amodei, D. \"AI Safety via Debate.\" arXiv:1805.00899, May 2018\n- Irving, G., Pfau, J., Hilton, B. \"An Alignment Safety Case Sketch Based on Debate.\" Alignment Forum, May 2025\n- Brown-Cohen, J. & Irving, G. \"Prover-Estimator Debate: A New Scalable Oversight Protocol.\" Alignment Forum, June 2025\n- Slaughter, I., et al. \"Community Notes Reduce Engagement with and Diffusion of False Information Online.\" PNAS 122(38), September 2025\n- Stafford, T. \"The Algorithmic Heart of Community Notes.\" Reasonable People newsletter, January 2025\n- Stafford, T. \"Do Community Notes Work?\" LSE Impact Blog, January 2025\n- Arjmandi-Lari, Z., Mantzarlis, A., Stafford, T. \"Threats to the Sustainability of Community Notes on X.\" arXiv:2510.00650, October 2025\n- Dryzek, J.S. & List, C. \"Social Choice Theory and Deliberative Democracy: A Reconciliation.\" British Journal of Political Science 33(1), 2003\n- Adachi, T., Chung, H., Kurihara, T. \"The Impossibility of Deliberation-Consistent Social Choice.\" American Journal of Political Science 68(3), 2024\n- Schulze, M. \"The Schulze Method of Voting.\" arXiv:1804.02973\n- Yudkowsky, E. \"Coherent Extrapolated Volition.\" MIRI, 2004\n- Yang, J.C., Bachmann, F. \"Bridging Voting and Deliberation with Algorithms: Field Insights from vTaiwan and Kultur Komitee.\" arXiv:2502.05017, February 2025\n- Coefficient Giving (Open Philanthropy). \"Request for Proposals: AI for Forecasting and Sound Reasoning.\" November 2025 (closed February 2026)\n- Barandiaran, X.E., et al. \"Decidim: A Brief Overview.\" In: Decidim, a Technopolitical Network for Participatory Democracy. Springer, 2024\n- Vagnoni, S. & Rodriguez-Doncel, V. \"A Deliberation Knowledge Graph: Bridging Institutional and Civic Democratic Discourse.\" 2025\n- Gitcoin Research. \"The Metacrisis: Coordination Failure at Civilizational Scale.\" March 2026\n- Gitcoin Research. \"Collective Intelligence Infrastructure: Protocols for Thinking Together.\" March 2026\n"}