{"path":"research/alignment-infrastructure-red-team/human-side-of-alignment-theory-and-evidence-report-2026-11-02.md","content":"# Helper report, verbatim: the human side of alignment, theory and measured evidence (2026-11-02)\n\n*An `Explore` helper on Sonnet, briefed with the founder's claim (before anyone can be confident a machine\nis aligned with human values and intent, humans must be able to say what those are and agree on the system\nthat serves machine alignment; a public record of reasoning that holds disagreement is the layer for that),\nthe angles to cover, the angles already in the corpus (to skip) and a read-status rule. Saved unchanged the\nmoment it arrived, before it was read further. Its figures are claims until the main note cites them with\ntheir tag; nothing here is ruled. The synthesis is in `../alignment-infrastructure-red-team.md` § 8d.*\n\n---\n\nStatus tags: [full text], [part], [abstract only], [secondary]. Years for 2025 and 2026 items are stated.\n\nCORRECTIONS TO THE BRIEF\n1. I found no paper matching \"Hadfield and Koh 2025, normative infrastructure\". Hadfield and Koh, \"An Economy of AI Agents\" (arXiv 2509.01063, Sept 2025) is an economics chapter on agent identity, licensing and the firm [part]. The normative-infrastructure argument sits in Hadfield's talks and in Trivedi, Hadfield and Hadfield-Menell, \"Building AI for the Democratic Matrix\" (Knight Institute essay, forthcoming on her Feb 2026 CV; text not retrieved). \"The Agentic Web Requires New Normative Infrastructure\" (arXiv 2606.10711, 2026) is by Pattison, Kolt, Lazar and others, not Hadfield [metadata only].\n2. Meta's \"6,000 people, 32 countries\" belongs to the earlier metaverse forum, not the AI forums.\n3. Stanford \"People Shaping AI\" ran Nov 2025 (report Feb 2026). I found no percentages.\n\n1. SOCIAL CHOICE AND THE THEORY OF WHAT WE ALIGN TO\n- Conitzer and 11 others, 2024, ICML (arxiv.org/abs/2404.10271) [part]. Position paper, no experiments. Adds a social-welfare-function step to RLHF; asks who gives input (representative subsets; citizens' assemblies only as a sampling tool), what feedback form, which aggregation rule. Borda, instant runoff and ranked pairs disagree on the same data. Deliberation is not developed; benefits are stated as beliefs.\n- Mishra 2023 (arxiv.org/abs/2310.16048) [abstract only]. From impossibility results: no single voting rule can align RLHF models for everyone; wants transparent voting rules and group-specific agents.\n- Gabriel 2020, Minds and Machines (arxiv.org/abs/2001.09768) [abstract only]. The task is fair principles people can reflectively endorse despite moral disagreement. No empirical test.\n- Zhi-Xuan, Carroll, Franklin, Ashton 2024, Philosophical Studies (arxiv.org/abs/2408.16984) [part]. Target is normative standards for a social role, negotiated by stakeholders. This is your claim, stated as conjecture: they say agreeing on norms is probably much easier than agreeing on a metric that ranks all consequences. They leave the conditions for fair agreement out of scope and fall back on norms people would agree to (virtual bargaining) when real agreement fails.\n- Sorensen et al. 2024, ICML (arxiv.org/abs/2402.05070) [abstract only]. Three kinds of pluralistic model (Overton, steerable, distributional), three benchmark types; says standard alignment may reduce distributional pluralism. No numbers.\n- 2025 to 2026. Zhang, Milli et al. (Meta), Community Alignment Dataset (arxiv.org/abs/2507.09650, ICLR 2026): five countries, N=15,000, 233,319 comparisons; people vary more than 21 LLMs do [abstract only]. Wojtowicz, Si, Doshi-Velez, Procaccia (arxiv.org/abs/2608.24046, Aug 2026): alignment as welfare optimisation, tried on real preferences (kidney allocation, food distribution, LLM answers, trolley problems); no figures in abstract [abstract only]. Freedman (arxiv.org/abs/2605.01642, 2026): reward-model jury plus voting rule, simulated annotators only [abstract only].\n- Moral graph: Klingefjord, Lowe, Edelman 2024, OpenAI-funded (arxiv.org/abs/2404.10636) [part]. 500 US participants, 3 divisive prompts. People wrote values; others voted which value is wiser in a context. 89.1% felt well represented; 70.3% learned about others' values; 85 values, 100 edges; about 4% of conflicts left unresolved and plural. Outcome was felt fairness, not agreement on a rule.\n\n2. HADFIELD\n- FAR.AI talk, 5 Aug 2025 (far.ai/events/sessions/gillian-hadfield-alignment-is-social-lessons-from-human-alignment-for-ai) [part, page summary]. Preference extraction and constitutions are the wrong model; norms are equilibria of human systems; a constitution needs institutions with procedures to interpret it. Proposes normative competence, normative reasoning, digital institutions (classification repositories, real-time APIs) and citizen juries. Concedes that letting AI take part in institutions, as against fitting into them, is a separate open question.\n- Regulatory markets: Hadfield and Clark (arxiv.org/abs/2304.04914; v5 Feb 2026; Jurimetrics 65, 2026) [abstract only]. Governments license private regulators; goals public, methods by competition.\n- Hadfield and Koh 2025 (arxiv.org/abs/2509.01063) [part, sections 3.3 to 5]. Registration, liability versus legal personhood, licensing by public or private bodies (left open), records. They concede governments do not know whether or how to regulate, and that firms can game what they disclose.\n- Kolt et al. 2026, 17 listed authors including Hadfield, Gabriel, Lazar (arxiv.org/abs/2601.04175) [part]. Legal alignment is necessary, not sufficient; law is silent or contested on much; current techniques allow no meaningful public participation.\n- No measured agreement outcome. This line puts agreement inside existing institutions, not in a new public record.\n\n3. DEMOCRATIC INPUT\n- Collective Constitutional AI, Anthropic and CIP, Oct 2023 (Huang et al., arxiv.org/abs/2406.07814; anthropic.com/research/collective-constitutional-ai-aligning-a-language-model-with-public-input) [part]. 1,002 US adults screened for AI familiarity; 1,127 statements (275 after moderation); 38,252 votes; two opinion groups. 95 ideas kept by group-aware consensus (threshold 0.723); about 50% concept overlap with Anthropic's constitution. Public model less biased on BBQ in all nine dimensions; MMLU 72.3 vs 72.4; GSM8K 85.6 vs 85.2. Contested statements (collective good vs individual liberty) were dropped, so the process removed disagreement rather than keeping it. No production change stated. Anthropic's Jan 2026 constitution (anthropic.com/constitution; keyword check of the full page) never mentions Collective Constitutional AI or public input; it says feedback comes from Claude models [part].\n- OpenAI Democratic Inputs, Jan 2024 (openai.com/index/democratic-inputs-to-ai-grant-program-update) [part]. Ten teams, $100,000 each, nearly 1,000 applicants, 113 countries. Opinion shifted day to day; online recruits skewed pro-AI; one team's US guidelines, vaccines included, all had over 72% support among Democrats, Independents and Republicans, while a small group opposed to any restriction lost under majority rule. Lab change: a promise of an end-to-end process, which became the 2025 study below.\n- OpenAI collective alignment, 27 Aug 2025 (openai.com/index/collective-alignment-aug-2025-updates) [full text]. About 1,000 people in 19 countries ranked responses to pre-selected prompts; they matched a model applying the Model Spec about 80% of the time; gaps at political, sexual and pseudoscience content. OpenAI adopted some edits (for example, political content for broad audiences is allowed) and declined tailored political content and erotica. It says changes are less likely higher in the instruction hierarchy, did not re-check proposals with participants, and has not studied disagreement within the crowd. The only documented case I found of a lab editing a governing text from public input.\n- CIP (cip.org/blog/alignment-assemblies-nine-months-in; blog.cip.org/p/2025gdindex) [part]. Assemblies: 1,000 US participants each for OpenAI and Anthropic; the only lab change named is that Anthropic built a new chatbot; CIP concedes the processes carry almost no accountability. Global Dialogues (2025): about 1,000 per round, about 70 countries, 6,000+ over seven rounds; 58% trust chatbots over elected representatives; no lab named as user.\n- Meta Community Forums [part]. Oct 2023 (Stanford report, fsi9-prod.s3.us-west-1.amazonaws.com/s3fs-public/2024-03/meta_ai_final_report_2024-04_v28.pdf): 1,545 people (336 Brazil, 402 Germany, 414 Spain, 393 US) in four separate deliberations, 1,108 controls, 38 proposals, knowledge gains 3.2 to 20.5 points; US share saying AI has had a positive impact rose from 49.8% to 54.4% (NBC, secondary). Oct 2024: about 900 people, five countries, 1,033 controls (fsi.stanford.edu). Nov 2025 industry forum with Cohere, Meta, Oracle, PayPal, DoorDash and Microsoft: 503 people, US and India, qualitative findings; a 2026 round is described as finished, with no results found [secondary]. Meta says results inform decisions and implementation takes time; I saw no named product change.\n- Taiwan, 23 Mar 2024 (participedia.net/case/13316; flagged for missing citations) [secondary]. 447 participants, 44 groups; about 85% backed stronger platform self-regulation; 89% wanted AI-generated content detected (+2.6 points). No adopted policy documented.\n\n4. THE SKEPTICS\n- Achen and Bartels 2016: voters follow group identity and bend views, even facts, to fit [secondary, getabstract.com summary]. Collected input may be loyalty, not judgment.\n- Deliberation evidence cuts both ways. Fishkin, Siu, Diamond, Bradburn, APSR 2021 (doi.org/10.1017/S0003055421000642) [part]: 526 voters, one weekend; partisans moved closer on 22 of 26 polarised proposals (19 significant); other-party ratings rose about 13 to 14 points; some effects faded within a year. Schkade, Sunstein, Hastie 2007: 63 citizens, like-minded groups grew more extreme [secondary, hls.harvard.edu]. Texas utility polls 1996 to 98 are credited with a turn to wind; no controlled comparison [secondary, nationalcivicleague.org]. Mutz: cross-cutting exposure raises awareness of others' reasons but lowers participation [secondary].\n- Attributed, incentivised disagreement can fail to converge. Karger, Tetlock et al., XPT 2023 (forecastingresearch.org/xpt) [abstract; figures via EA Forum quote of report]: 80 experts, 89 superforecasters, months of debate paid for persuasion; AI-extinction-by-2100 stayed at 3% (experts) vs 0.38% (superforecasters); authors call convergence minimal.\n- Procedure may not buy acceptance. Skitka and Mullen 2008: moral conviction overrides procedural fairness; contested by Napier and Tyler [secondary]. Esaiasson et al. 2019, BJPolS: outcome favourability dominates acceptance [secondary]. This hits \"agree on the system, then accept its outputs\".\n- Moral disagreement among the informed (Rawls on burdens of judgment; Doris and Plakias 2008): not read this pass; background only.\n- Realist. Amodei, \"The Adolescence of Technology,\" Jan 2026 (darioamodei.com) [part]: with trillions of dollars a year at stake it is \"very difficult for human civilization to impose any restraints on it at all.\" Rost 2026 (arxiv.org/abs/2604.14070) [part]: no institution can compel a frontier lab to modify, delay or withdraw a deployment; the White House commitments were negotiated without public deliberation. Sloane et al. 2020 (arxiv.org/abs/2007.02423) warn of participation-washing [abstract only].\n\n5. LEGIBLE DISAGREEMENT AND RULES\nNo study I found tests this directly. Nearest:\n- List, Luskin, Fishkin, McLean 2013, J. Politics (researchonline.lse.ac.uk/id/eprint/46863) [abstract only]: after deliberation, preferences moved closer to single-peaked (sharing one issue dimension), which makes majority outcomes stable; substantive agreement generally did not rise. Agreement on the shape of the disagreement, not the answer.\n- IETF, RFC 7282 (Resnick 2014; rfc-editor.org/rfc/rfc7282) [part]: rough consensus when every objection has been understood and answered, not necessarily accommodated; chairs ask for the nature of objections. No outcome statistics.\n- Debian [part]: 2014 init vote, five named options with 5 to 13 seconders each; the winner said no general resolution was needed (debian.org/vote/2014/vote_003). 2022 firmware vote, six named options; a Social Contract change won 289 to 63 against a 3:1 bar (debian.org/vote/2022/vote_003.en.html).\n- Wikipedia policy process: no outcome study found.\n- My inference: every case where structure helped had a decider bound by the process (IETF chair, Debian constitution). Where the decider was not bound (Iceland's Parliament, the labs), the record changed nothing.\n\nCLOSING TABLE\n| Study or process | What was made legible | What agreement followed | What did not | Status |\n| Collective Constitutional AI 2023 | 1,002 voters, 38,252 votes, two groups | 95 consensus ideas became a constitution | Contested items dropped; no production change | [part] |\n| OpenAI collective alignment 2025 | Rankings plus reasons, about 1,000 people | About 80% match to Spec; some edits adopted | Tailored politics, erotica declined; crowd disagreement unstudied | [full text] |\n| OpenAI grants 2024 | Ten team prototypes | One team: over 72% cross-party support | Opinion drifted daily; no model change shown | [part] |\n| Meta forums 2023 to 2025 | Pre/post attitudes on 38 proposals | Knowledge gains; some opinion shifts | No named product change | [part] |\n| Taiwan assembly 2024 | 447 people, 44 groups | About 85% backed platform self-regulation | No policy found | [secondary] |\n| Moral graph 2024 | Values and \"wiser\" votes, 500 people | 89.1% felt represented | 4% of conflicts open; no rule adopted | [part] |\n| America in One Room 2019 | Balanced briefings, small groups | 22 of 26 proposals closer | Partial fade in a year; no rule | [part] |\n| XPT 2022 to 2023 | Named experts' reasoning, months | Little; 3% vs 0.38% on AI risk | Convergence despite paid persuasion | [abstract; part] |\n| List et al. 2013 | Pre/post preference orders | More single-peaked | Substantive agreement | [abstract only] |\n| vTaiwan UberX 2015 to 16 (Hsiao et al. 2018, osf.io/xyhft) | Pol.is: 925 people, 31,115 votes, groups of 45% and 55% | Rules in force 25 Oct 2016 | A liquor bill failed; authors include Tang | [part] |\n| Ireland Assembly 2016 to 18 (Irish Times: irishtimes.com/news/ireland/irish-news/citizens-assembly-backs-abortion-rights-in-wide-range-of-circumstances-1.3058170; referendum reports) | Graded ballots: 64% of 87 for no limit on reasons | Referendum 66.4% yes | Committee took a different constitutional route | [secondary] |\n| Iceland 2010 to 13 (Gylfason 2014, ifo.de/DocDL/cesifo1_wp5056.pdf) | 950-person forum, elected council | 67% yes at 49% turnout | Parliament never enacted; author sat on council | [part] |\n| British Columbia 2004 to 05 (Wikipedia 2005 BC referendum page; Ipsos) | Assembly's STV plan | 57.69% yes | Under 60% bar; 36% claimed fair knowledge or more | [secondary] |\n| IETF rough consensus | Objections stated and answered | Standards (no statistics) | No outcome data | [part] |\n| Debian votes 2014, 2022 | Named options, Condorcet | 2022 rule change, 289 to 63 | 2014: no change | [part] |\n| Cogitate 2025 (Nature; Birmingham record; Down To Earth on the 2023 letter) | Rival predictions preregistered, 256 people | Test run and published | 124 scientists called IIT pseudoscience, 2023 | [secondary] |\n\nOn the evidence here, a public, attributed, disagreement-preserving record has not been shown to help people agree on a rule, though the nearest studies show that structure (single-peaked preferences, answered objections, named options) can be agreed where a decider is bound to the process. The one lab that edited a governing text from public input did so at low levels of its hierarchy, and the one consensus-filtered constitution dropped contested items instead of keeping them. So the founder's claim is plausible but untested, and the sharpest test would be a record whose participants include a decision-maker who has committed in advance to be bound by it.\n"}