When an institution makes a serious mistake despite the involvement of many intelligent people, one explanation often appears quickly: groupthink. The term evokes a familiar picture in which individuals suppress doubts, follow dominant personalities or avoid challenging a consensus because disagreement feels professionally or socially costly. These behaviours certainly occur, but treating groupthink primarily as a weakness of individual judgement can obscure a deeper problem. Sometimes people converge because the institution has been organised in ways that make independent judgement progressively more difficult. In those cases, groupthink is not only something that happens among people. It is something an architecture can produce.
The distinction matters because organisations frequently assume that the presence of many decision-makers provides protection against error. If ten capable people examine a problem, their agreement may seem more reassuring than the judgement of one. Yet this confidence depends on an assumption that is rarely examined: that their judgements are sufficiently independent for agreement to contain additional information. If everyone has been exposed to the same evidence, filtered through the same categories, interpreted through the same hierarchy and shaped by the same incentives, ten agreeing judgements may provide far less confirmation than they appear to.
This is the central architectural problem. CONSENSUS ≠ INDEPENDENT CONFIRMATION. Agreement becomes strong evidence only when the routes through which people reached that agreement retain enough independence to make convergence informative. When those routes are highly correlated, agreement can simply reproduce the same blind spot several times.
Consider a senior policy meeting in which participants receive a briefing prepared through a single reporting chain. The briefing already reflects decisions about which evidence matters, which uncertainties deserve attention and which interpretations appear credible. Junior analysts whose observations challenge the prevailing account may have seen their concerns softened as information moved upwards. By the time senior participants discuss the issue, they may sincerely and independently conclude that the evidence supports the same position. Nobody in the room needs to be consciously conforming. Their agreement can still be structurally produced because the architecture has already narrowed what each participant is able to see.
This is why GROUPTHINK ≠ SIMPLY BAD INDIVIDUAL JUDGEMENT. An organisation can contain thoughtful, critical people and nevertheless generate highly correlated conclusions. The relevant question is not only whether individuals are willing to think independently, but whether the institutional environment gives them sufficiently independent material with which to think.
Information architecture is one mechanism through which this happens. If every decision-maker receives evidence through the same channel, omissions in that channel become shared omissions. If all performance information is generated by the same measurement system, limitations in that system become common limitations. If uncertainty is systematically removed from reports as they travel upwards, senior decision-makers may independently become overconfident because they are reasoning from evidence that has been made uniformly confident before it reaches them.
Hierarchy can produce a similar effect, although HIERARCHY ≠ GROUPTHINK AUTOMATICALLY. Hierarchy can clarify responsibility, resolve conflicts and make action possible. The cognitive risk appears when organisational position determines which interpretations are safe to express or which questions can remain open. If a senior interpretation becomes an implicit reference point early in a process, later participants may adjust not only what they say but what they investigate. Evidence compatible with the dominant interpretation receives attention; alternatives appear increasingly peripheral. Eventually the institution may experience convergence that looks epistemic even though it was partly produced by authority.
Incentives can synchronise judgement in subtler ways. People learn which forms of uncertainty are rewarded, which objections delay decisions, which kinds of analysis advance careers and which problems are considered outside their remit. No instruction to conform is necessary. If the same incentive environment acts on many people, it can make their cognitive behaviour more similar. The organisation begins to select not merely among actions but among ways of seeing.
This is the mechanism preserved by the idea of correlated cognitive error. If several people independently make unrelated mistakes, the errors may cancel, contradict or expose one another. If their mistakes share a common source, however, adding more people does not necessarily improve the judgement. An organisation can increase the number of participants while leaving the underlying error architecture untouched.
CORRELATED ERROR ≠ MANY PEOPLE MAKING THE SAME INDEPENDENT MISTAKE. The important feature is common causation. Participants may overlook the same risk because they rely on the same dataset, accept the same conceptual model, receive information through the same hierarchy or respond to the same organisational incentives. Their agreement then reflects shared exposure to a cognitive structure rather than multiple independent confirmations of reality.
This helps explain why adding consultation to a decision process does not automatically solve groupthink. If new participants encounter the same information, answer questions framed in the same way and operate within the same assumptions, participation can increase without cognitive independence increasing. The institution has added voices but not necessarily added genuinely different routes to judgement.
Nor is dissent sufficient on its own. DISSENT ≠ CORRECTNESS. A contrarian view can be poorly informed, strategically motivated or simply wrong. The objective of cognitive architecture is not to reward disagreement regardless of quality. It is to preserve enough independence that alternative interpretations can arise, encounter evidence and be evaluated before convergence becomes irreversible. What matters is not disagreement as a virtue but the institution’s ability to distinguish convergence produced by evidence from convergence produced by its own architecture.
Meeting design offers a simple illustration. If participants state their views only after hearing the most senior person, later judgements are no longer formed under the same conditions as the first. If initial assessments are recorded independently before discussion, the institution preserves information about the distribution of judgement before social and hierarchical interaction begins. The meeting may ultimately reach exactly the same conclusion, but it can now know more about how that conclusion was reached. The architecture has preserved a degree of cognitive independence without eliminating coordination.
The same principle applies beyond meetings. Parallel analytical teams can examine a problem using different methods. Forecasts can be generated before analysts see one another’s estimates. Red teams can test assumptions from outside the dominant project structure. Frontline observations can reach decision-makers through channels that do not depend entirely on the hierarchy responsible for implementing the policy being evaluated. None of these arrangements guarantees better judgement, but each can reduce the probability that one cognitive failure becomes universal before it is visible.
This also reveals an important distinction between COGNITIVE INDEPENDENCE and ORGANISATIONAL INDEPENDENCE. People do not need to work in separate institutions to think through sufficiently different routes, and separate organisations can still reproduce identical assumptions. Independence is cognitive when evidence, models, perspectives or inferential paths retain enough variation for one judgement to provide information that another does not already contain.
Artificial intelligence can either weaken or strengthen this independence. If every team relies on the same model to summarise evidence, generate scenarios or recommend interpretations, apparently independent human judgements may inherit correlated assumptions from a common computational source. Ten officials may each review a problem separately while unknowingly beginning from near-identical AI-generated representations. The number of human reviewers has increased, but the diversity of cognitive pathways may have decreased.
Conversely, AI can sometimes be used deliberately to expose alternative hypotheses, compare interpretations or identify evidence inconsistent with a dominant account. The relevant design question is therefore not whether AI participates in institutional judgement, but whether its role increases or reduces the independence of the cognitive pathways on which the institution relies.
Groupthink also has a temporal dimension. Once an interpretation becomes institutionalised, it can shape which data are subsequently collected and which questions continue to be asked. Early convergence can therefore manufacture later evidence for itself. A policy classified as successful may generate reporting designed around indicators of successful implementation; contradictory effects that fall outside those indicators become less visible; later reviews then encounter an increasingly coherent body of evidence supporting the original judgement. The architecture has not merely coordinated belief. It has begun to coordinate observation.
At that point, shared blind spots can become remarkably stable. New participants enter an environment in which the dominant interpretation is already embedded in reports, categories, routines and expectations. They may reproduce it without ever experiencing direct pressure to conform. What began as a particular judgement becomes part of the institutional conditions under which future judgement occurs.
This is why interventions aimed only at individual behaviour can be insufficient. Telling employees to “speak up” does little if the relevant evidence never reaches them, if dissent appears only after commitments have become costly to reverse or if alternative interpretations have no recognised route into decision processes. Psychological safety matters, but the architecture of information, timing, incentives and authority matters as well.
A cognitively resilient institution therefore asks not merely whether people agree, but what produced the agreement. Did participants begin from genuinely independent observations? Were alternative hypotheses possible before the dominant account became established? Could inconvenient evidence travel through the organisation? Did hierarchy determine the conclusion before deliberation began? Are multiple judgements actually multiple sources of information, or are they repetitions of one cognitive pathway?
These questions do not imply that consensus is suspicious by definition. Institutions need convergence in order to act, and genuine evidence should often cause different observers to reach similar conclusions. The purpose is not to preserve disagreement indefinitely. It is to ensure that convergence remains informative.
Groupthink is an architectural problem when institutional structures make cognitive errors correlated. Hierarchies, incentives, information flows and decision processes can cause intelligent people to see through increasingly similar lenses until collective agreement becomes difficult to distinguish from a shared blind spot. The solution is not permanent disagreement, but architectures that preserve enough cognitive independence for errors to encounter alternatives before the institution mistakes repetition for confirmation.