{"path":"research/the-group-premise-evidence.md","content":"# The group premise, for and against — evidence for a constellation tile\n\n**Date**: 2026-09-09 · **Type**: evidence pass for one wall sentence, on founder instruction during the constellation review (*\"I'd also like us to explore and research evidence both for and against this group premise, start with reading up in the corpus, then do a bit more online research.\"*). **Read-depth**: the corpus sources were re-read; the new outside sources are search-grade (abstracts and summaries; the Becker 2019 abstract read in full) — verify any figure before it reaches a funder or a public page.\n\n**The tile under review** (Branch 2, the medium cluster): *\"Thinking works best in groups. The graph is one.\"* It makes two claims: that groups think better than individuals, and that the graph is a group in the sense that matters. The evidence below bears on both.\n\n---\n\n## 1. What the corpus already held\n\n- **Mercier & Sperber** ([academic-foundations.md](../academic-foundations.md) § 7): individuals are hopelessly biased, groups remarkably effective — **and only when there is genuine disagreement**. The corpus's own open question there: how does Deliberus ensure diverse participation without unproductive conflict?\n- **DeliData (2023)**: 64% of groups found better solutions than any individual, and **43.8% of the successful groups contained no member who had solved the problem alone** ([non-zero-sum-economics-and-civilizational-cooperation.md](non-zero-sum-economics-and-civilizational-cooperation.md) § the cooperation evidence; [progressive-disclosure.md](progressive-disclosure.md)). Structured deliberation with visible reasoning, not voting.\n- **[collective-intelligence-and-epistemic-democracy.md](collective-intelligence-and-epistemic-democracy.md)**: Landemore's epistemic case for inclusive deliberation; Hong & Page's diversity theorem **with its critiques and what survives them**; the citizens'-assembly record (Ireland, France, Taiwan, Belgium, America in One Room — *\"among the strongest empirical evidence that structured deliberation reduces polarization, by exposing people to the actual reasoning behind opposing views\"*); Surowiecki's four conditions (diversity, independence, decentralization, aggregation). Its § 1 presents Woolley's collective-intelligence factor as *real and measurable*; see § 3 below for why that now carries a caveat.\n- **Sunstein's four failures of deliberating groups** — amplification, cascades, polarization, hidden profiles (the project's always-loaded instructions, § Academic Foundations → Collective deliberation; ledger E16: minted claims are anti-hidden-profile, *nobody has said X* is anti-cascade; E17: **polarization has no structural answer**).\n- **Bail et al. 2018** and **Kalla & Broockman** ([cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md) §§ 8–9; ledger I2–I5): a month of opposing *conclusions* increased polarization; arguments alone null, arguments plus non-judgmental narrative durable at four months. The corpus's reading: **flags versus reasons** is the variable, and nobody has run Bail's design with reasons.\n- **The DeepMind facilitation study** (FAccT '26; threat-model entry 2): LLM facilitation did not improve consensus, participants preferred it anyway, and *summarising* steered allocations.\n- **Our own two measurements** ([what-human-judgment-is-for.md](what-human-judgment-is-for.md) § the population; ledger L2, L3): **co-presence is not engagement** (run 6: 402 candidate cross-source pairs, zero at threshold), and **N users is not N frames** — a thousand readers from one community are one frame sampled a thousand times, the correlated-error finding one level up, so frame diversity must be selected for, never awaited.\n\n## 2. New evidence FOR the premise\n\n- **Groups solve what individuals cannot.** Moshman & Geil (1998): 70–80% of participants solved the Wason selection task when discussing it in small groups, against 10–20% of the same people individually. Trouche, Sander & Mercier (2014, *JEP: General*): the gain comes from the **exchange of arguments**, not from confidence or social facilitation — a *truth wins* dynamic in which the member with the right answer convinces the rest by argument ([Sperber & Mercier, Reasoning as a social competence](https://www.dan.sperber.fr/wp-content/uploads/2012_mercier_reasoning-as-a-social-competence.pdf); [Trouche et al.](https://www.apa.org/pubs/journals/features/xge-a0037099.pdf)).\n- **Small deliberating groups beat large silent crowds.** Navajas, Niella, Garbulsky, Bahrami & Sigman (2018, *Nature Human Behaviour*; N = 5,180): structuring a crowd into small independent groups that deliberate briefly and then aggregating the groups' consensus estimates beat every method of combining the same people's individual answers ([Nature](https://www.nature.com/articles/s41562-017-0273-4)).\n- **Dissent is the active ingredient.** Schulz-Hardt, Brodbeck, Mojzisch, Kerschreiter & Frey (2006, *JPSP*; 135 three-person groups on a hidden-profile task): groups with pre-discussion dissent solved the problem more often, the effect carried by greater discussion intensity and less biased discussion; highest when the dissenter held the correct answer ([Aston](https://research.aston.ac.uk/en/publications/group-decision-making-in-hidden-profile-situations-dissent-as-a-f/)). Nemeth (1986; Nemeth & Kwan 1987): exposure to a *minority* view produces divergent thinking — more strategies considered, more correct solutions found ([Nemeth & Kwan](https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1559-1816.1987.tb00339.x)).\n- **Even partisan crowds get wiser by exchanging information.** Becker, Porter & Centola (2019, *PNAS* 116:10717; two web experiments on factual questions known to elicit partisan bias): *\"In contrast to polarization theories, we found that social information exchange in homogeneous networks not only increased accuracy but also reduced polarization\"* — abstract read in full ([PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC6561169/)). The exchanged objects were **numeric estimates on factual questions**, which is the condition to carry: it says nothing yet about value claims.\n- **Deliberative polling.** Fishkin's design — random samples, balanced briefing, moderated small groups, before-and-after surveys — reports across more than a hundred polls in some thirty countries that deliberation raises knowledge and moderates extreme positions, with depolarization on complex policy issues rather than identity-saturated symbolic ones ([Participedia](https://participedia.net/method/deliberative-polling)). The corpus's citizens'-assembly section holds the same finding.\n- **Machine-side mechanism evidence (2026).** Multi-agent LLM debate improves truth-seeking when the debaters are *epistemically diverse*, even from weak individual performers; the process analysis finds **majority pressure suppresses independent correction while effective teams overturn a wrong consensus** ([arXiv 2605.30391](https://arxiv.org/abs/2605.30391)) — the argumentative theory reproduced in silicon, with the same conditions.\n\n## 3. New evidence AGAINST, or conditioning it\n\n- **Like-minded groups polarize, and arguments do it more than comparison.** Isenberg's 1986 meta-analysis (21 articles, 33 effects): polarization arises from both social comparison and persuasive argumentation, **with the argument effects larger** ([APA](https://psycnet.apa.org/record/1986-24477-001)). So *reasons rather than flags* is not a free pass: **reasons exchanged inside one frame move the group further out**; it is reasons across frames that correct. Schkade, Sunstein & Hastie (2007; 63 Colorado citizens): liberal Boulder groups moved left and conservative Colorado Springs groups moved right on all three issues, and within-group diversity fell ([Chicago Unbound](https://chicagounbound.uchicago.edu/cgi/viewcontent.cgi?httpsredir=1&article=12230&context=journal_articles)).\n- **Exposure to the other side's conclusions can backfire.** Bail et al. (2018, *PNAS* 115:9216): a month of a bot retweeting the opposing party's officials and opinion leaders left Republicans substantially more conservative; Democrats slightly more liberal, not significantly ([PNAS](https://www.pnas.org/doi/10.1073/pnas.1804840115)).\n- **Groups sit on what only one member knows.** Lu, Yuan & McLeod's 2012 meta-analysis (65 studies, 101 effects, 3,189 groups): groups mention two standard deviations more *shared* than *unique* information, and hidden-profile groups are **eight times less likely** to find the solution than fully informed ones; the share of unique information that gets mentioned predicts decision quality ([Sage](https://journals.sagepub.com/doi/abs/10.1177/1088868311417243)).\n- **Mild social influence kills the diversity that makes a crowd wise.** Lorenz, Rauhut, Schweitzer & Helbing (2011, *PNAS*): showing people others' estimates narrowed the spread of answers without improving accuracy ([PNAS](https://www.pnas.org/doi/10.1073/pnas.1008636108)). Independence before pooling is a condition, not a nicety.\n- **\"Groups have an IQ\" replicates weakly.** Woolley et al.'s 2010 collective-intelligence factor has one reported replication by its own group; Bates & Gupta (2017) find group performance largely explained by member IQ; a 2024 study found two factors with neither social sensitivity nor turn-taking related to them; a meta-analysis found ~80% of studies underpowered ([Bates & Gupta](https://www.sciencedirect.com/science/article/abs/pii/S0160289616303282); [2024 review](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11318883/); [g versus c](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8019454/)). **The corpus's collective-intelligence doc § 1 overstates this and now carries a caveat.**\n- **\"Diversity trumps ability\" is not a theorem you can lean on.** Thompson (2014) and successors show the Hong–Page result holds only under conditions its own experiments weaken ([AMS Notices](https://www.ams.org/notices/201409/rnoti-p1024.pdf); [Thompson's challenge, a note](https://www.tandfonline.com/doi/full/10.1080/08913811.2017.1288455)). The corpus already records what survives: diversity helps under conditions, as an empirical matter, not by mathematics.\n- **Kahan** (already in the corpus): every reasoning-proficiency measure worsens polarization on contested questions; only curiosity reverses it. More reasoning in a group of the uncurious is a better weapons factory.\n\n## 4. Synthesis — the premise is true under five conditions, and the graph supplies three of them\n\n| Condition | Evidence | What the graph does about it |\n|---|---|---|\n| 1. **Genuine dissent is present** | Mercier's own rider; Schulz-Hardt; Nemeth | Attack edges and *nobody has said X* only exist if someone brings them; run 6 showed co-presence without engagement (L2). **Partial.** |\n| 2. **Reasons, not conclusions, are what travels — and across frames** | Trouche; Isenberg (arguments within a frame polarize *more*); Bail | Structure, never summaries; the map shows premises. **Supplied**, but the across-frames half depends on condition 5. |\n| 3. **Private information gets surfaced** | Lu, Yuan & McLeod; Sunstein's hidden profiles | Every reason minted as a claim; always-mint; the completeness oracle. **Supplied.** |\n| 4. **Independence before pooling** | Lorenz; Navajas (deliberate in small independent groups, then aggregate) | Strength is computed from arguments, never from votes (tonight's two-axis ruling); the map is read at a break, not during the talk (kartpaus). **Supplied.** |\n| 5. **The minds actually differ** | Colorado; Becker 2019's network design; L3 | **Not supplied.** Frame diversity must be selected for: register diversity as precondition, ingestion by debate cluster, the matching rule for sessions. |\n\nPolarization keeps its verdict from E17: no structural answer. What the graph offers against it is conditions 2 and 5 together — reasons across frames — which is the bet nobody has run (I3), stated as a bet.\n\n## 5. What this means for the tile\n\n*\"Thinking works best in groups. The graph is one.\"* fails the over-broad test on both halves. The first sentence is false in four documented ways for an unconditioned group, and Isenberg makes it worse than false for like-minded ones. The second is aspirational: a graph is a group only when frames engage in it, and the one measurement we have shows they need not.\n\nCandidates, the founder's call:\n\n- **(a) *\"Different minds beat one, when every reason is on the table.\"*** — 11 words. *Different minds* carries condition 5, *every reason on the table* carries 2 and 3; the graph's job (supplying the table) is named and the recruitment job (supplying the difference) is implied. **Recommended.**\n- **(b) *\"Many minds beat one, when every reason is on the table.\"*** — drops the diversity condition, which is the one the graph cannot supply and the one most of the counter-evidence turns on.\n- **(c) keep as is**, hover carries the caveats — leaves a stranger-facing sentence the evidence contradicts.\n\nThe hover, whichever wins: the five conditions and the one the graph cannot supply.\n\n**RULED 2026-09-09.** The founder took the direction of (a) and sharpened the word: *\"Minds that disagree beat one mind, when every reason is on the table.\"* — his path to it, verbatim: *\"Thanks, great. \\\"A divergence of minds beat one\\\"? Perhaps? Or something like it?\"* → *\"Agree on the second. I was about to try to use the word \\\"diversity\\\" but \\\"disagree\\\" is more on-point here, right?\"* *Disagree* over *diversity* because diversity is who is in the room and disagreement is what they do with it, which is what the evidence rewards.\n\n## 6. Corrections owed elsewhere, made 2026-09-09\n\n- [collective-intelligence-and-epistemic-democracy.md](collective-intelligence-and-epistemic-democracy.md) § 1 — the collective-intelligence factor is now caveated as weakly replicated.\n- Ledger row N7 added (the conditional premise); E16, I2, I3, L2, L3 point here.\n\n## Cross-references\n\n[what-human-judgment-is-for.md](what-human-judgment-is-for.md) (N users is not N frames) · [cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md) §§ 8–9 (flags versus reasons) · [collective-intelligence-and-epistemic-democracy.md](collective-intelligence-and-epistemic-democracy.md) · [islands-of-coherence.md](islands-of-coherence.md) (the matching rule) · [convergence-wager-red-team.md](convergence-wager-red-team.md) (the DeepMind result) · `specs/landing-redesign/constellation-map.md` (the tile).\n"}