Consensus

Abstract

Consensus is a type of decision making: the general agreement or collective judgment a group arrives at when its members' individual positions converge. Cognitive psychology studies how that convergence is produced, distinguishing agreement that reflects pooled evidence from agreement manufactured by social pressure. Two broad routes operate: the aggregation of independent judgments, which can make a group more accurate than most of its members, and social influence during interaction, which can either sharpen or corrupt the result. Formal models specify when iterated opinion exchange converges to a shared value and when majority processes track the truth. This article surveys the mechanisms of social influence, the conditions under which crowds are wise, and the formal theories of how groups reach, and sometimes mistakenly reach, a common position.

Keywords: consensus, group decision making, wisdom of crowds, social influence, opinion aggregation

Consensus is the point at which the distribution of opinion in a group collapses toward a single value, but the same observable agreement can arise from opposite processes: a group may converge because its members have pooled independent evidence, or because they have yielded to one another regardless of the evidence (Galton, 1907). Cognitive psychology is therefore less interested in whether a group agrees than in how it came to agree, because only the first route reliably improves on the individual members (Davis, 1973).

What Consensus Is

Consensus denotes general agreement or the collective judgment a group settles on, together with the process that produces it. As a form of decision making, it differs from an individual choice in that the inputs are distributed across several minds that must be combined, whether by explicit voting, by discussion, or by the implicit averaging of opinions over repeated contact (DeGroot, 1974). The defining feature is convergence: a set of initially dispersed positions moves toward a common value, and the central scientific questions concern the rule by which the combination happens and whether that rule leaves the group better or worse than its members.

Figure 1

Two properties make consensus a distinct object of study rather than a scaled-up individual decision. First, the quality of the collective judgment depends on statistical structure that has no individual analogue, chiefly the independence and the error distribution of the contributing judgments. Second, the process is reflexive: members observe one another and adjust, so the act of reaching agreement changes the very opinions being aggregated (Kerr & Tindale, 2004). A theory of consensus must therefore track both the aggregation rule and the social dynamics that feed it.

Mechanisms of Social Influence

The classic analysis of why people move toward a group position distinguishes two influences that usually co-occur (Deutsch & Gerard, 1955). Normative social influence is the pressure to conform in order to be accepted, which can produce public agreement without private acceptance; informational social influence is the acceptance of others' judgments as evidence about reality, which changes private belief. The distinction matters for consensus because only informational influence transmits information: a consensus built on normative pressure records how much members wanted to fit in, not what they collectively knew.

Asch's conformity experiments made the normative route vivid, showing that a unanimous majority giving an obviously wrong answer on a line-judgment task could induce a substantial minority of participants to endorse the error against the plain evidence of their senses (Asch, 1956). The result establishes that visible unanimity exerts a force of its own, independent of the content of the judgment, so an observed consensus cannot be taken at face value as a signal of shared evidence.

Influence is not only a majority phenomenon. A consistent, confident minority can shift a majority over time, and group discussion typically polarizes opinion, driving the average position further in the direction it already leaned rather than toward a bland midpoint (Moscovici & Zavalloni, 1969). These findings reframe consensus as the product of an active influence process in which positions are argued and converted, not a static tally of pre-existing views. One account holds that reasoning itself is built for this arena, evolved to produce and evaluate arguments in group discussion rather than to perfect solitary inference, which predicts that groups can outperform individuals when argument quality, not social pressure, drives the outcome (Mercier & Sperber, 2011).

Aggregating Independent Judgments

The oldest quantitative result on consensus is also the most counterintuitive: a collection of fallible, independent estimates can combine into a highly accurate one. Galton observed that the median of hundreds of independent guesses of an ox's weight fell within a percent of the truth, a demonstration now called the wisdom of crowds (Galton, 1907). The mechanism is statistical. Each estimate carries a signal plus an error; when errors are large but uncorrelated, averaging cancels them while the signal survives, so the aggregate beats most individuals and often all of them.

The principle is systematically underappreciated. People asked to combine their own estimate with an adviser's tend to choose one or anchor on their own rather than average, forgoing the error reduction that simple averaging guarantees whenever the two judgments bracket the truth (Larrick & Soll, 2006). Research on advice taking confirms that the weight people give to others' opinions is usually too low relative to the accuracy gains available, so the wisdom of crowds is left partly unclaimed even when it is within reach (Bonaccio & Dalal, 2006).

The precondition is independence, and it is fragile. When members see one another's estimates before responding, their judgments correlate, the diversity of errors collapses, and the crowd's accuracy can degrade even as the group grows more confident and more tightly agreed (Lorenz et al., 2011). Yet influence is not uniformly harmful: structured exchange can help. Aggregating the revised estimates produced by a small number of deliberating dyads can outperform the mean of a large independent crowd, because brief discussion lets members correct gross errors without fully homogenizing (Navajas et al., 2018). Pooling independent judgments also raises accuracy in applied diagnosis, where combining several readers' independent assessments improves detection over the average individual reader (Kurvers et al., 2016).

Formal Models of Reaching Consensus

The DeGroot model gives the canonical formal account of how iterated social influence yields agreement (DeGroot, 1974). Each member holds an opinion and a set of non-negative trust weights on the others that sum to one; at every step each replaces its opinion with the weighted average of the group's current opinions. When the resulting influence matrix is connected and aperiodic, the opinions converge to a single consensus value equal to a fixed weighted average of the initial opinions, where each member's weight is its long-run social influence rather than its own confidence. The model makes precise how a group can reach exact agreement through local averaging alone, and why the member others trust most, not the one most sure of itself, dominates the result.

Where DeGroot describes continuous opinions, social decision scheme theory models discrete group choices as a probabilistic mapping from the distribution of members' initial preferences to the group's decision (Davis, 1973). The scheme is the rule, estimated from data, that says how often a group with a given split of opinions adopts each option, letting majority rule, proportionality, and truth-wins processes be compared directly. Analyses in this tradition show that majority or plurality schemes describe most unstructured group decisions and that, for judgment tasks with a demonstrably correct answer, majority processes are a robust and efficient route to accuracy, usually matching more complex schemes (Hastie & Kameda, 2005). A unifying theme across these models is social sharedness: shared task representations, shared preferences, and shared information all gain disproportionate weight in the group's process, which explains both the power and the predictable failures of consensus (Tindale & Kameda, 2000).

Combination also operates at the smallest scale, the dyad. Two observers who share their confidence can form a joint judgment more accurate than either alone, but only when they are of comparable reliability and communicate how sure they are; when one is much worse, naive confidence-sharing drags the pair below the better member (Bahrami et al., 2010). The benefit depends on the validity of subjective confidence as a cue, which holds for easy items but reverses for deceptive ones, so two heads are better than one chiefly when confidence tracks accuracy (Koriat, 2012).

Worked Example

Consider three analysts estimating a quantity, with initial estimates of 70, 40, and 90, as rendered in the interactive DeGroot model above. Their trust weights form the rows of an influence matrix: the first analyst places weights 0.5, 0.3, 0.2 on the three members; the second places 0.2, 0.6, 0.2; the third places 0.1, 0.3, 0.6. Each analyst's updated opinion is the weighted average of the current opinions.

After one round the first analyst holds 0.5 times 70 plus 0.3 times 40 plus 0.2 times 90, which equals 35 plus 12 plus 18, or 65; the second holds 0.2 times 70 plus 0.6 times 40 plus 0.2 times 90, which equals 14 plus 24 plus 18, or 56; the third holds 0.1 times 70 plus 0.3 times 40 plus 0.6 times 90, which equals 7 plus 12 plus 54, or 73. The spread has already fallen from 50 points to 17. Iterating the same averaging drives all three opinions together, and they converge to a single value of about 63.81.

That limit is not the plain average of the starting estimates, which would be 66.67. It is the weighted average using each member's long-run influence, the stationary weights 5/21, 9/21, and 7/21 implied by the trust matrix: 5 times 70 plus 9 times 40 plus 7 times 90, all divided by 21, equals 1340 divided by 21, or 63.81. The second analyst, whose estimate was the lowest, pulls the consensus furthest because the others trust that analyst most, illustrating the model's central lesson that influence, not confidence or correctness, determines whose opinion prevails.

Discussion

The two routes to consensus impose a single design principle: aggregate independent judgments before letting members influence one another. Independence is what makes averaging cancel error, so any process that correlates opinions early, open round-robin disclosure, visible tallies, a dominant speaker, spends the statistical advantage that made the crowd worth consulting (Lorenz et al., 2011). The same logic explains why structured protocols that collect private estimates first and only then permit discussion tend to outperform free interaction.

A recurring failure mode arises when members share most of their information and hold little that is unique. Groups discuss what everyone already knows and neglect the privately held facts that would change the decision, so the consensus reproduces the common prior rather than pooling the distributed evidence, the hidden-profile problem (Stasser & Titus, 1985). The effect is worsened, not helped, by knowing others' preferences in advance, which biases discussion toward preference-consistent information and degrades the decision further (Mojzisch & Schulz-Hardt, 2010). Structured elicitation and accuracy incentives mitigate these failures, as in forecasting tournaments where training, weighting, and recombining judgments produced consensus estimates that beat prediction markets and intelligence analysts (Mellers et al., 2014).

When a cohesive group values agreement for its own sake, the drive toward consensus becomes its own failure mode. Janis called this groupthink: members suppress doubts, discount outside information, and converge prematurely on a course of action, mistaking the resulting unanimity for evidence that the decision is sound. The mechanism is normative influence turned inward, cohesion and the wish to preserve it override the independent appraisal that would otherwise correct the group, so the very agreement that signals a healthy consensus in an accuracy task signals its corruption here. Reviews of the evidence find the classic antecedents only loosely predictive and locate the reliable damage in defective decision process, poor information search and the suppression of dissent, rather than cohesion as such (Esser, 1998). The practical remedy is the same design principle in reverse: protect dissent and independent appraisal so that agreement, when it comes, reflects pooled evidence rather than the pressure to concur.

Current Directions

The most active empirical front studies consensus on realistic social networks rather than in isolated crowds or fully connected groups. The network's structure turns out to govern the outcome: decentralized communication, in which no member occupies a privileged broadcasting position, can let social influence improve a group's collective estimate instead of corrupting it, reversing the pessimistic reading of earlier herding results (Becker et al., 2017). Adaptive rewiring, in which members preferentially attend to better-performing peers, further amplifies the wisdom of the crowd by routing influence toward accuracy (Almaatouq et al., 2020).

A parallel line examines how consensus on conventions tips. A committed minority holding a fixed position can overturn an established group convention once it exceeds a critical fraction of the population, giving an experimental handle on when and how collective agreement reorganizes rather than merely drifts (Centola et al., 2018). At the mechanistic end, work on how people share and communicate confidence is clarifying the cognitive ingredients that let interacting minds combine well, and when metacognitive miscalibration makes them combine badly (Bang & Frith, 2017). Across these strands the field is moving from asking whether crowds are wise to specifying the communication structure and metacognitive conditions under which influence builds an accurate consensus rather than a confident error.

Key Researchers

Solomon Asch

(1907-1996). Demonstrated through the line-judgment experiments that a unanimous majority can induce individuals to endorse a plainly wrong answer, isolating normative conformity pressure. Wikipedia

Dan Bang

(contemporary). Investigates the cognitive and neural basis of how people share confidence and combine opinions to reach joint decisions better than either member alone. ORCID

Damon Centola

(contemporary). Shows experimentally how social-network structure governs whether a population reaches consensus and how committed minorities can tip collective conventions. ORCID

Morris H. DeGroot

(1931-1989). Gave the canonical formal model of consensus, in which agents repeatedly replace their opinion with a trust-weighted average and converge to a shared value. Wikidata

Morton Deutsch

(1920-2017). Drew the foundational distinction between normative and informational social influence, the two routes by which members move toward a group position. Wikipedia

Francis Galton

(1822-1911). Reported the first quantitative demonstration of the wisdom of crowds, showing that the median of hundreds of independent guesses can nearly equal the true value. Wikipedia

Reid Hastie

(contemporary). A leading authority on group and jury decision making who formalized how groups combine judgments through social decision schemes and documented the robustness of majority rules. ORCID

Ralph Hertwig

(contemporary). Extended the wisdom-of-crowds principle to repeated and pooled judgments and studies how groups aggregate estimates under uncertainty. ORCID

Tatsuya Kameda

(contemporary). Studies the adaptive and ecological rationality of group decision making, including when majority processes track the truth and the evolutionary logic of consensus. ORCID

Serge Moscovici

(1925-2014). Established group polarization and minority influence, showing that consensus is produced by an active influence process rather than a simple averaging of prior views. Wikipedia

Glossary

Condorcet Jury Theorem.

The result that if each member decides correctly with probability above one-half and members are independent, the probability that a majority is correct rises toward certainty as the group grows.

DeGroot Model.

A formal model in which each agent repeatedly replaces its opinion with a trust-weighted average of the group's opinions, converging to a consensus equal to a fixed weighted average of the initial values.

Diversity Prediction Theorem.

The identity that a crowd's squared error equals the average individual squared error minus the diversity of the individual estimates, so that more varied errors yield a more accurate aggregate.

Group Polarization.

The tendency of group discussion to shift the average opinion further in the direction members already favored rather than toward a moderate midpoint.

Hidden Profile.

A situation in which the information needed for the best decision is distributed so that no member holds it alone, and groups fail to surface it because discussion favors shared over unique information.

Informational Social Influence.

Acceptance of others' judgments as evidence about reality, which changes private belief and transmits information into the consensus.

Majority Rule.

A social decision scheme that adopts the option preferred by more than half the members, the most common aggregation rule and a strong tracker of the truth when members are individually better than chance.

Minority Influence.

The process by which a consistent, confident minority gradually converts a majority, shifting the group's consensus over time.

Normative Social Influence.

Conformity driven by the desire to be accepted, which can produce public agreement without private acceptance and so manufacture a consensus that carries no evidence.

Opinion Aggregation.

The combination of several individuals' judgments into one collective judgment, by averaging, voting, or weighting, that can be more accurate than its inputs.

Social Decision Scheme.

A probabilistic rule mapping the distribution of members' initial preferences onto the group's decision, used to compare majority, proportional, and truth-wins processes.

Social Sharedness.

The tendency of group discussion and judgment to be dominated by information, preferences, and beliefs the members already hold in common, at the expense of unique information held by only one member.

Truth-Wins Scheme.

A social decision scheme in which a group adopts the correct answer whenever at least one member finds and can demonstrate it, characteristic of problems with verifiable solutions.

Wisdom of Crowds.

The phenomenon in which the aggregate of many independent, fallible judgments is more accurate than most or all of the individual judgments, provided their errors are large and uncorrelated.

Frequently Asked Questions

What is consensus in group decision making?

Consensus is the general agreement a group reaches when its members' positions converge on a common judgment, studied not as an outcome but as a process whose quality depends on how the individual judgments were combined (Kerr & Tindale, 2004).

How can a group be more accurate than its members?

When independent estimates carry uncorrelated errors, averaging cancels the errors while preserving the shared signal, so the aggregate beats most individuals, the wisdom-of-crowds effect first quantified by Galton (Galton, 1907).

What is the difference between normative and informational social influence?

Normative influence is conforming to be accepted, producing public agreement without belief change; informational influence is treating others' views as evidence, which changes private belief and transmits information (Deutsch & Gerard, 1955).

When does social influence undermine the wisdom of crowds?

When members see one another's judgments before responding, their estimates correlate and the diversity of errors collapses, so accuracy can fall even as the group becomes more confident and more tightly agreed (Lorenz et al., 2011).

What is the DeGroot model of consensus?

It models each agent repeatedly adopting a trust-weighted average of the group's opinions; under mild connectivity the opinions converge to one value weighted by each member's long-run social influence rather than its confidence (DeGroot, 1974).

Does majority rule produce good group decisions?

For judgment tasks with a demonstrably correct answer, majority and plurality schemes describe most real group decisions and are a robust, efficient route to accuracy, usually matching more elaborate aggregation rules (Hastie & Kameda, 2005).

What is a hidden profile?

It is a decision in which the information favoring the best option is split across members so no one holds it alone; groups often miss it because discussion dwells on shared rather than unique information (Stasser & Titus, 1985).

Can a minority change a group's consensus?

Yes; a consistent and confident minority can convert a majority over time, and committed minorities can tip an established convention once they pass a critical fraction of the group (Moscovici & Zavalloni, 1969).

References