Wisdom: Collective Judgment & AI Governance

Updated 14 July 2026

Wisdom is a multifaceted phenomenon that embodies successful judgment under uncertainty through aggregation, diversity, and context-sensitive reasoning.
It is applied in social decision-making and AI governance, where collective estimates and metacognitive regulation refine outcomes.
Key aspects include leveraging independent judgments, managing social influence, and implementing control mechanisms to regulate optimization processes.

Searching arXiv for recent and foundational papers on "wisdom" and "wisdom of crowds" to ground the article. Searching arXiv for exact paper (Engelhardt et al., 2020) and adjacent work on social influence, collective wisdom, and AI wisdom. Wisdom, in contemporary technical literature, is not a unitary construct but a family of related phenomena concerning successful judgment under uncertainty, complexity, and interdependence. In social decision-making, wisdom commonly denotes the “wisdom of crowds”: the possibility that aggregation of many estimates yields a collective judgment closer to truth than most constituent estimates. In networked opinion dynamics, the same term marks conditions under which interaction improves or degrades collective accuracy. In artificial intelligence and systems engineering, wisdom is used more narrowly for higher-order regulation: metacognitive control over reasoning strategies, or architectural governance over whether optimization should proceed at all. Across these domains, recurring themes are aggregation, diversity, independence, social influence, humility, context sensitivity, and the distinction between optimizing a goal and interrogating the goal itself (Engelhardt et al., 2020, Challet, 2016, Johnson et al., 2024, Chang, 15 Jun 2026).

1. Collective wisdom as aggregation

A central research meaning of wisdom is the functional superiority of aggregated judgments over individual ones. In the econophysics framing, this is “functional wisdom of the crowds”: collective performance can emerge from heterogeneous, noisy, and even individually suboptimal agents through aggregation, learning, and selection (Challet, 2016). The canonical mathematical expression is the aggregate estimate

$\overline{x} = \frac{1}{N}\sum_{i=1}^N x_i,$

which serves as the simplest formalization of a crowd judgment when individuals supply estimates $x_i$ (Challet, 2016).

A second line of work reframes crowd wisdom as a one-dimensional unsupervised dimension-reduction problem. In that formulation, the response matrix is approximately rank-one because multiple observers are noisy measurements of a shared latent signal. Principal component analysis and Isomap can then recover a consensus latent dimension, accommodating binary and continuous responses, and the first principal component simultaneously acts as a consensus score and a weighting of individual reliability (Wang et al., 2017). This extends classic crowd-aggregation models beyond binary voting and suggests that collective wisdom can be represented as latent-structure recovery rather than mere arithmetic averaging.

The empirical status of crowd wisdom is mixed. One experimental study of online threads found that, for difficult tasks, seeing preceding estimates aids the wisdom of crowds, but that if participants only see extreme estimates, wisdom quickly turns into folly; the same work assigns each participant a persuadability score using a Gaussian Mixture Model and reports that persuadability increases with task difficulty and with the amount of social information provided (Engelhardt et al., 2020). By contrast, an analysis of the Survey of Professional Forecasters reported that the crowd beats all individuals in less than 2% of the forecasts and beats most individuals in less than 70% of the forecasts, while also finding a positive correlation between diversity and crowd error and no informative relation between skew and crowd error (Reia et al., 2020). This suggests that “wisdom of crowds” is not a universal law but a contingent property of task structure, error distribution, and information flow.

A persistent question is whether interaction among estimators strengthens or destroys collective wisdom. Experimental evidence on threads indicates an explicitly conditional answer: social information can improve accuracy on difficult estimation tasks, yet filtering that exposes participants only to extreme prior estimates quickly converts wisdom into folly (Engelhardt et al., 2020). This conditionality is echoed in analytic and agent-based models of opinion dynamics.

In a stochastic opinion-dynamics model, each agent updates according to

$\frac{dx_i(t)}{dt} = \alpha \left[\bar{x}(t)-x_i(t)\right] + \beta \left[x_i(0)-x_i(t)\right] + A\xi_i(t),$

where $\alpha$ is social influence and $\beta$ is individual conviction (Mavrodiev et al., 2020). Two systemic observables are central:

$\mathcal{E}(t) = [\ln \mathcal{T} - \langle \ln x(t)\rangle]^2,$

the collective error, and

$\mathcal{D}(t)=\mathrm{Var}[\ln x(t)],$

the group diversity (Mavrodiev et al., 2020). The analytical result is unambiguous on one point: increasing $\alpha$ reduces diversity. Its effect on accuracy, however, is ambiguous. Social influence improves the wisdom of crowds only if the initial collective error is large; in most other cases it deteriorates the outcome, while individual conviction can mitigate harmful drift when the group begins near the truth (Mavrodiev et al., 2020).

Agent-based simulations sharpen the same conclusion. Under both aggregated-information and full-information regimes, social influence only in rare cases enhances the wisdom of crowds; more often agents converge to a collective opinion that is farther from the true answer, with stronger distortion under full information than under aggregated information (Mavrodiev et al., 2020). The theoretical influence-systems literature formalizes this in terms of social power allocation: wisdom is improved when relatively more social power is allocated to relatively more accurate individuals, and undermined when more accurate individuals receive less social power (Tian et al., 2022). In democratic influence networks, if relatively more accurate individuals are relatively less susceptible, the wisdom is improved; if more accurate individuals are more susceptible, the wisdom is undermined (Tian et al., 2022).

Finite-time analysis further shows that wisdom preservation is not a single property. In French-DeGroot systems,

$x_i(k+1)=\sum_{j=1}^n P_{ij}x_j(k),$

one-time wisdom, finite-time wisdom, infinite-time wisdom, uniform wisdom, and pre-uniform wisdom are distinct notions (Bullo et al., 2019). A key obstruction is the presence of “prominent families”: groups of size $o(n)$ whose cumulative influence on the population average remains order one. Their absence implies finite-time wisdom in general, and in equal-neighbor models the absence of prominent individuals is equivalent to one-time wisdom and pre-uniform wisdom (Bullo et al., 2019). The theoretical implication is that independence is not merely a psychological desideratum; it is a structural property of influence matrices and social power distributions.

3. Diversity, deliberation, and the “wisdom of crowds” versus the “wisdom of a few”

Diversity is often treated as a precondition for crowd wisdom, but the literature shows that diversity alone is insufficient, and that its operational meaning depends on how groups are organized. One influential experiment structured a live crowd of 5,180 participants into independent groups of five. Participants first answered general-knowledge questions individually, then deliberated to produce group consensus answers, and then revised their individual estimates. Averaging the consensus decisions of as few as four independent groups outperformed the aggregation of the initial independent opinions of thousands of individuals (Navajas et al., 2017). Deliberation reduced within-group variance, increased between-group variance, and decreased bias; participants reported that “sharing arguments and reasoning together” rather than merely following the most confident member was the principal mechanism (Navajas et al., 2017). The result challenges the blanket claim that social influence necessarily undermines crowd accuracy.

Yet other evidence suggests that the relevant unit may often not be a broad crowd at all. An analysis of Facebook and Twitter found that a small percentage of active users represent 50% of user-generated content, that this set is stable over time, and that these active users are highly connected among themselves (Baeza-Yates et al., 2016). The same study introduced the “digital desert,” referring to content that is never seen, thereby challenging the assumption that every contribution enters collective judgment (Baeza-Yates et al., 2016). This motivates the phrase “wisdom of a few” rather than “wisdom of crowds” in some online settings.

Negative ties make the problem even harder. In a DeGroot-like model with in-group attraction and out-group opposition,

$x_i$ 0

agents may follow friends and oppose enemies (Eger, 2013). Under sufficiently strong negative relations, wisdom is difficult to attain; persistent disagreement, bi-polarization, or multi-polarization can arise even in connected societies, and only “neutral” consensus profiles are admissible unless fixed-point conditions on deviation functions are met (Eger, 2013). A plausible implication is that diversity contributes to wisdom only when embedded in interaction structures that preserve informational complementarity rather than convert heterogeneity into factional opposition.

4. Humility, persuadability, and metacognition

Several papers shift the focus from aggregation rules to the cognitive and metacognitive traits that permit wise judgment. In the thread experiment, persuadability is treated as an individual-level parameter inferred from observed updating behavior. It increases with task difficulty and with the amount of social information provided, and filtered threads produce an increasing gap between highly persuadable participants and skeptics (Engelhardt et al., 2020). This identifies a micro-level mechanism through which social information can be either corrective or distorting.

A network-level analogue appears in work on intellectual humility. There, a DeGroot updating rule

$x_i$ 1

is extended so that self-weight is modulated by evidence quality and homophily (Mahjabin et al., 4 Feb 2025). The reported effect is that intellectual humility improves the accuracy of collective estimations, reduces polarization, and increases the Revision Coefficient, meaning that less accurate individuals revise more after social feedback while more accurate individuals appropriately retain confidence (Mahjabin et al., 4 Feb 2025). The mechanism is explicitly metacognitive: greater sensitivity to comparative evidence quality and reduced susceptibility to partisan homophily.

The same metacognitive emphasis appears in AI-centered work. One program defines wisdom as the ability to successfully navigate intractable problems through both task-level strategies and metacognitive strategies, with particular emphasis on intellectual humility, epistemic deference, scenario flexibility, context adaptability, perspective seeking, and viewpoint balancing (Johnson et al., 2024). On this account, current AI systems are optimized primarily for task outputs and exhibit “metacognitive myopia”: overconfidence, poor calibration, failure to recognize limitations, and difficulty handling context shifts (Johnson et al., 2024). Another empirical line studies LLMs through “decision trajectories” across transformer depth. It distinguishes stable-correct, stable-wrong, unstable-correct, and unstable-wrong trajectories, and reports that the largest group is unstable-correct rather than stable-correct (Rana, 31 May 2026). Correctness and stability are thus dissociated, and “wisdom” in this setting refers less to final accuracy alone than to settled, robust, and auditable decisiveness.

5. Wisdom in artificial intelligence and system development

In AI and software engineering, wisdom is often distinguished from intelligence. “Wise computing” proposes a development environment endowed not with general wisdom, and not AI in the standard sense, but wisdom geared toward classical system-building, manifested in creativity, proactivity, and deep insights into the system’s own structure, behavior, goals, and rationale (Harel et al., 2015). The proposed Wise Development Suite is conceived as an active participant in development, maintenance, evolution, and runtime operation rather than a passive tool. Its prototype architecture includes Athena, Regina, and Livia, which respectively support offline formal analysis, offline empirical exploration, and online observation and interaction (Harel et al., 2015).

A more recent architectural program makes the distinction sharper. “Architectural wisdom” is defined as a corrigible objective-governance layer above the optimization substrate: intelligence accepts a goal and optimizes within it, whereas wisdom interrogates whether the goal should be optimized at all (Chang, 15 Jun 2026). The framework requires three structural commitments before any action—temporal horizon, relational boundary, and irreversibility—and realizes them through four components: Structural Utility Transform, Moral Admissibility Interface, Arbitration and Escalation Controller, and Value Revision Channel (Chang, 15 Jun 2026). These components compute a six-coordinate wisdom tuple over horizon, relational coverage, irreversibility, admissibility, value revision, and auditability:

$x_i$ 2

(Chang, 15 Jun 2026).

This architectural conception responds to failures that capability scaling alone does not repair: engagement optimization that amplifies harmful pathways, tool-using agents that commit irreversible actions, and preference-trained LLMs that become sycophantic (Chang, 15 Jun 2026). It aligns with the metacognitive view that wise AI should be robust to novel environments, explainable to users, cooperative with others, and safer because it risks fewer misaligned goals with human users (Johnson et al., 2024). The common element is higher-order self-governance: the system must monitor, revise, or refuse objectives, not merely optimize them.

6. Operationalizations, controversies, and research directions

The technical literature offers multiple operationalizations of wisdom, each tied to a different object of study. In crowd estimation and forecasting, wisdom is quantified through collective error, diversity, social power, or outperformance relative to individuals (Mavrodiev et al., 2020, Tian et al., 2022, Reia et al., 2020). In influence systems, it is a property of stochastic matrices, convergence regimes, and the absence of prominent agents or families (Bullo et al., 2019). In online discussion threads, it is affected by task difficulty, the amount and extremity of visible social information, and individual persuadability (Engelhardt et al., 2020). In LLMs, it can be studied through answer margins, next-layer margin changes, boundary distances, and the stability of decision trajectories through depth (Rana, 31 May 2026). In AI governance, it is represented through architectural commitments and non-scalar audit tuples rather than a single objective score (Chang, 15 Jun 2026).

Several controversies follow directly from these operational differences. First, social influence is neither uniformly beneficial nor uniformly harmful. Deliberation inside small, independent groups can improve accuracy (Navajas et al., 2017), but full exposure to others’ estimates or extreme opinions can degrade it (Mavrodiev et al., 2020, Engelhardt et al., 2020). Second, diversity is not automatically a resource. It may support collective accuracy, yet empirical forecasting results show positive correlation between diversity and crowd error (Reia et al., 2020), and bias mitigation work on hybrid human-LLM crowds shows that simple averaging of homogeneous LLM crowds can exacerbate existing biases because diversity is too limited (Abels et al., 18 May 2025). That work further reports that locally weighted aggregation better leverages the wisdom of the LLM crowd, and that hybrid crowds combining human diversity with LLM accuracy further reduce biases across ethnic and gender-related contexts (Abels et al., 18 May 2025). Third, wisdom is not reducible to accuracy. The AI metacognition literature treats robustness, perspective integration, calibrated uncertainty, and context adaptability as constitutive rather than ancillary (Johnson et al., 2024).

A plausible synthesis is that wisdom, across these literatures, names a constrained success condition for judgment systems operating under uncertainty and interdependence. In human groups, it requires aggregation schemes and communication structures that preserve useful diversity without producing correlated error, extremal herding, or elite capture. In networks, it depends on how social power, susceptibility, and topology interact with the distribution of accuracy. In artificial systems, it increasingly denotes metacognitive or architectural machinery for regulating objectives, explanations, and irreversible consequences before optimization proceeds. The term therefore spans collective intelligence, social epistemology, interpretability, and AI governance, but in each setting it retains a common core: successful judgment requires not only problem-solving capacity, but disciplined control over how information, influence, and goals are admitted into the decision process (Challet, 2016, Engelhardt et al., 2020, Johnson et al., 2024, Chang, 15 Jun 2026).