When twelve voices become one error

The usual story about collective intelligence begins with arithmetic. Ask many people, average their answers, and idiosyncratic errors cancel. That story is useful, but incomplete. Cancellation requires the errors to point in sufficiently different directions. Ten measurements generated by ten independent disturbances can be informative. Ten measurements copied from one miscalibrated instrument are still one measurement wearing ten name badges.

This is the first hidden assumption in the phrase “wisdom of crowds”. It is not crowd size. It is error diversity under a common question. Participants must be estimating the same target, possess at least some relevant information, and retain enough independence for one mistake not to force the others. Aggregation then converts distributed evidence into a group estimate.

The count that matters is not the headcount

Suppose each of n judges has an error variance of σ², and every pair of errors has the same correlation ρ. Under these simplifying assumptions, the variance of their equal-weight mean is:

When ρ is zero, averaging reduces variance by a factor of n. When ρ is positive, the bracketed term returns part of that error. We can rewrite the expression as if the crowd contained a smaller number of independent voices:

Influence is an intervention, not background noise

Independence need not mean silence. A group can improve after interaction if people with worse information move towards people with better information. The crucial issue is when and how influence enters. Lorenz and colleagues showed that even mild social influence could narrow answer diversity and raise confidence without reliably improving accuracy.[1] Becker, Brackbill and Centola later found a more conditional result: in decentralised networks, social learning improved estimates because less accurate participants revised more, while centralised networks made outcomes depend heavily on the central person.[2]

The lesson is narrower than “deliberation is bad”. Navajas and colleagues had 5,180 people answer independently, deliberate in groups of five, and produce group consensuses. Aggregating a small number of those consensuses could outperform a very large pool of isolated answers.[3] The gain came from a sequence: private judgement, small-group exchange, then aggregation across groups. Deliberation operated inside cells while diversity survived across cells.

Why the revision record matters

A final answer is a lossy compression of a collective process. It hides who supplied the decisive observation, which estimates were copied, and whether uncertainty fell because evidence improved or because dissent became costly. This matters even when the final answer is correct. A system that reached the truth by following one lucky speaker is fragile under the next case. A system that reached the same answer by exposing and testing distinct clues has learned something about its own reliability.

Revision behaviour can itself become evidence, but only with care. Madirolas and de Polavieja showed that resistance to social influence could help identify more reliable estimators in collective estimation.[18] The mechanism is plausible: a participant who holds a strong private signal may move less than one who is guessing. Yet stubbornness, status and overconfidence can produce the same behavioural trace. Resistance should therefore be combined with calibration history and source quality, not treated as an intrinsic mark of expertise.

flowchart TD
  A["When twelve voices become one<br/>error"] --> B["The topology decides what survives"]
  B --> C["People, models and the<br/>illusion of diversity"]
  C --> D["Build for truth before consensus"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#D95545,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

The topology decides what survives

A group does more than contain people. It contains paths along which evidence, confidence and authority travel. Two groups with identical members can produce opposite results when the communication graph changes. A hub allows fast coordination but concentrates error. A mesh lets evidence travel around a bad node but may create echoing loops. Independent cells preserve diversity but can miss a clue that only becomes visible when fragments meet.

Topology matters because social influence is recursive. When I revise towards you, and a third person revises towards me, your original signal has been counted twice unless the system remembers its lineage. This is why a visible tally can be deceptive. It reports the current state of belief, not the number of independent observations that produced it.

Four communication geometries

Controlled evidence supports this topology-sensitive view. In forty-person estimation networks, decentralised communication reduced dispersion and improved central estimates, while a centralised network could pull the whole group towards the central node whether that node was right or wrong.[2] In adaptive networks, connections could shift towards better performers, allowing the group to respond to biased or changing information; accurate individuals also tended to resist poor influence.[4] Structured information-sharing networks reduced diagnostic error across 84 group trials involving 2,941 practising clinicians, with the weakest performers improving and the strongest not being dragged down.[5]

These results do not mean that a mesh automatically creates truth. A network can be decentralised yet still homogeneous in data, profession or model family. It can also polarise if people preferentially connect to similar beliefs and assimilate congenial evidence more strongly than contrary evidence. Models of biased assimilation and homophily show how two camps can move apart despite continuing to exchange information.[6] Social ratings can also create asymmetric herding, where an early positive signal accumulates further positive evaluations even when the initial signal was random.[7]

Truth, coordination and polarisation are different outputs

This distinction matters because a system can become better coordinated while becoming less accurate. A one-word communication channel increased cooperation from zero to 48.3 per cent in a four-agent Stag Hunt experiment with language-model agents.[8] That result is evidence about coordination under a game structure, not about discovering truth. The agents may converge on a mutually beneficial convention even if every factual belief they use is wrong.

The human layer is part of the topology

Edges on a diagram are not equivalent communication channels. One member may interrupt freely while another waits for permission. A specialist may translate uncertainty into a crisp recommendation because the meeting format rewards brevity. A junior participant may possess the only disconfirming observation but interpret silence from senior members as evidence against it. Formal connectivity can therefore overstate practical connectivity.

Research on human groups offers one useful clue. Woolley and colleagues found a group-level performance factor across varied tasks that was associated with social sensitivity and more equal conversational turn-taking, rather than simply the average or maximum individual intelligence in the group.[17] The result does not establish a unitary group mind, nor should the factor be treated as a universal trait. It does show why participation structure can be causally relevant. When one voice monopolises the channel, the group loses access to information held elsewhere.

For a human–agent interface, equal speaking time is not always the right target. A database query may deserve more weight than an unsupported opinion. The design objective is equal opportunity for decision-relevant evidence to enter, not equal authority for every utterance. Useful controls include anonymous private entry, structured rounds for disconfirming evidence, a protected dissent channel, and a chair who summarises source lineages rather than personalities.

People, models and the illusion of diversity

Human–AI systems make the independence problem harder because visible participants are not reliable units of evidence. Eight agents may use different personas yet share a base model, training corpus, retrieval index and system prompt. Three human reviewers may have read the same briefing note produced by one model. A vote among them can look diverse while tracing back to one upstream generator.

The practical mistake is to count outputs rather than lineages. In statistical language, shared dependencies induce correlated error. In engineering language, they create a common-mode failure. In epistemic language, they make several testimonies less independent than they appear. Diversity is useful only when it changes the probability of being wrong together.

Human plus AI is not automatically better than the better partner

A large meta-analysis of 106 experiments and 370 effect sizes found two results that are easy to conflate. Human–AI combinations performed better than humans alone on average, with a standardised effect of g = 0.64. Yet they performed worse than the better of the human or AI alone, with a synergy effect of g = −0.23.[9] The first result demonstrates augmentation. The second shows that complementarity is not automatic.

Why can combination fail? A human may defer to a stronger model on cases where the human has the rare corrective clue. The model may provide an explanation that increases trust without increasing error detectability. Confidence may be uncalibrated across partners. Or the interface may ask a human to approve a recommendation rather than make an independent judgement. In the meta-analysis, explanations and confidence displays were not reliable moderators of synergy on average, and decision tasks showed poorer complementarity than creation tasks.[9]

Multi-agent language-model systems face the same trap. ReConcile reported gains from rounds of discussion among diverse model agents and found model diversity important.[10] More recent work challenges broad interpretations of “debate”: majority voting can account for much of the apparent gain, while intrinsic model quality and genuine group diversity dominate outcomes; majority pressure can also suppress a correct minority.[11][12] The latter study remains a preprint, and the literature is young, but both identify the right causal question. Does exchanging arguments add corrective information beyond sampling several answers and voting?

Topology benchmarks also resist a universal answer. MultiAgentBench compares star, chain, tree and graph organisations across collaborative and competitive tasks; graph coordination performed best in one research scenario, but the benchmark’s broader lesson is task dependence rather than a winning shape.[19] A coding task with decomposable modules, an investigation with cross-cutting clues and a negotiation with strategic incentives require different communication paths.

A disciplined evaluation therefore needs matched baselines. Compare one strong model with tools, several independent samples plus voting, agents that exchange only conclusions, and agents that exchange evidence records. Hold model calls and token budgets constant where possible. Then intervene on topology, speaking order and shared sources. Without these controls, a multi-agent gain may be a sampling gain, a compute gain or an interface gain described as “collective reasoning”.

Agreement is not yet joint agency

A second confusion appears when a set of aligned outputs is described as a collective agent. Coincident intention is too weak. Four systems can independently choose the same action because they share a reward function. A crowd can converge on a forecast. Neither fact creates a subject or team that owns a plan.

Accounts of shared agency emphasise relations such as mutually responsive intentions, meshing subplans and commitments that constrain how participants treat one another.[13][14] For engineered systems, a useful operational translation is narrower: explicit roles, compatible plans, authority boundaries, veto conditions, and obligations to repair or hand over after interruption. These conditions establish governed joint action, not metaphysical group consciousness.

stateDiagram-v2
  [*] --> Claim
  state "The count that matters is not<br/>the headcount" as Claim
  state "Influence is an intervention,<br/>not background noise" as Test
  state "Why the revision record matters" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build for truth before consensus

The design problem is not to eliminate influence. It is to stage influence so that it can correct rather than overwrite. A practical architecture needs four gates. Each gate preserves something that later stages would otherwise destroy: private evidence, causal reasons, dependence structure, and accountable authority.

Gate one: register private evidence

Before showing peer answers, record each participant’s estimate, confidence, decisive evidence and provenance. For agents, record model and version, prompt or policy, retrieval sources, tool outputs and stochastic seed where available. For humans, record which briefing, data view or prior discussion they saw. The goal is not surveillance. It is to preserve the pre-influence state.

Gate two: exchange reasons, not only positions

Participants should expose the observation or assumption that would change another member’s answer. Showing only a tally encourages conformity. Showing full prose can create rhetorical overload. A better exchange uses compact claim records: conclusion, confidence, decisive evidence, strongest counterevidence and a requested test. Speaking order can be randomised or rotated so status does not become a hidden prior.

Gate three: synthesise with dependence in view

Aggregation should cluster shared evidence lineages before weighting outputs. Confidence earns weight only after calibration. A source used by five agents should normally enter once, with the agents’ interpretations represented separately. When dependence is uncertain, report a range rather than one precise probability. The system should also preserve a minority report when a dissenting lineage contains evidence absent from the majority.

Gate four: contract authority and recovery

If the result will cause action, specify who may execute, who can veto, which state must be re-read, what receipt proves the effect, and how partial execution is reversed or resumed. This gate keeps epistemic synthesis separate from operational authority. A model proposal, a group belief and a permitted action are different objects.

A decision instrument for choosing the communication mode

A preregistered test of influence latency

The central thesis can be tested without building a large agent society. Use six-member groups on tasks with objectively scorable answers and decomposable evidence. Give each member one private clue. On half the tasks, secretly give three members a common briefing containing the same plausible error. This creates a known dependence structure while preserving individual uncertainty.

Randomise groups into a two-by-two design. The first factor is influence timing: members either register a private claim before seeing peers, or see an evolving tally immediately. The second factor is provenance visibility: the synthesis interface either shows shared-source clusters or presents all six voices as separate. Keep total deliberation time, incentives, task order and available evidence constant. Human, model-only and mixed teams can be run as separate cohorts rather than pooled.

Primary outcomes should be final accuracy and calibration. Secondary outcomes should include correction of an initially wrong majority, creation of a new error from an initially correct majority, survival of a correct minority, concentration of influence, time to decision and the effective number of evidence lineages represented in the final answer. Store every revision with its preceding evidence view so causal order can be reconstructed.

The predicted interaction is specific. Private-first registration should help most when the common briefing is wrong. Provenance visibility should add further benefit by preventing three repetitions from receiving three independent weights. When the common briefing is correct, immediate social influence may be faster and can sometimes be equally accurate. That result would define the cost of protection rather than refute the mechanism.

This experiment also separates three rival explanations. A reflection-time control tests whether private registration merely makes people think longer. A hidden-random-label control tests whether provenance labels act as authority cues. An independent-sampling baseline tests whether discussion adds anything beyond drawing more answers. The resulting evidence would not establish a general law of collective intelligence, but it would identify a concrete break condition for one communication design.

Executable lab: when does communication help?

The lab below simulates a binary judgement. Each agent receives a private signal and may also rely on one shared source. You can compare four protocols: a private majority, a public cascade, a central hub and the four-gate synthesis. The fourth protocol knows which part of each opinion came from the shared source and counts that source once.

What the lab tests

The simulation asks one narrow question: when several agents partially copy a shared source, what changes if the aggregator can separate common evidence from private evidence? Increase shared-source overlap. Private majority accuracy may initially rise if the common source is useful, then saturate because the same source dominates many voices. Increase social influence. The public cascade becomes more sensitive to order, while the central hub becomes more sensitive to one agent’s luck. The gated protocol counts the common source once and pools the remaining private components.

function gatedVote(privateParts, sharedPart) {
  // Invariant: each private part belongs to one agent.
  // Invariant: the common source enters exactly once.
  const privateEvidence = privateParts.reduce((sum, x) => sum + x, 0);
  return Math.sign(privateEvidence + sharedPart) || 1;
}

function effectiveVoices(n, rho) {
  if (n < 1 || rho < 0 || rho > 1) throw new RangeError("Invalid inputs");
  return n / (1 + (n - 1) * rho);
}

The full simulation runs in the page script. It uses a seeded random generator, agent-specific private skill, one common source, and protocol-specific influence rules. The compact function above exposes the article’s central assumption: dependence-aware synthesis requires a typed distinction between private and shared evidence. Without that distinction, an aggregator cannot know whether five matching statements are five observations or five copies.

The strongest failure case

A second failure occurs when confidence is treated as competence. In perceptual dyads, communication improved performance when partners had similar sensitivity, but a large skill mismatch could make the pair worse than its better member.[15] People can also display an “equality bias”, weighting partners more equally than their demonstrated competence justifies.[16] The opposite failure is blind deference to a prestigious expert or high-confidence model. A dependence-aware system therefore needs calibration evidence, not just self-reported certainty.

The smallest useful implementation is modest. Add a private-estimate field to the review interface. Store the evidence source and confidence before exposing peer answers. Display source overlap alongside vote counts. Preserve revisions as revisions, not replacement votes. For consequential action, require a named owner, veto and readback. No multi-agent framework is required.

sequenceDiagram
  participant Q as Question
  participant M as Mechanism
  participant E as Evidence
  participant D as Decision
  Q->>M: When twelve voices become one error
  M->>E: The count that matters is not the headcount
  E-->>M: compare rival explanations
  E->>D: Protect independent evidence before seeking agreement
  D-->>Q: revise scope or act

Protect independent evidence before seeking agreement

When assembling a committee, multi-agent workflow or human–AI review, stop asking only how many members it contains. Ask four different questions: How many independent evidence lineages exist? When does social influence enter? Which topology can carry correction without concentrating error? Who owns action and recovery after the judgement?

This changes several design decisions immediately. Do not display a live tally before private registration when the question is consequential. Do not treat role prompts as evidence diversity. Do not add agents merely to create debate if independent sampling and voting explain the gain. Do not let a central expert or model become both the dominant information source and the final authority. Do not call a unanimous output “joint agency” unless roles, commitments, vetoes and recovery survive interruption.

Communication is valuable because it can move information to where it is missing. It is dangerous because it can also move error, status and confidence along the same paths. The aim is not maximum agreement. It is a collective in which better evidence can still defeat a popular mistake.