High-performing teams: Is collective intelligence the answer?

The study investigates whether a single, general collective intelligence factor (c-factor) exists in human groups, analogous to Spearman’s g-factor for individual intelligence, and whether this factor uniquely predicts group performance. Building on Woolley et al. (2010), who reported a positive manifold across group tasks and a single c-factor with strong predictive validity, the authors contextualize the importance of intelligence in academic, occupational, and health domains and ask whether an analogous group-level construct exists. Given mixed evidence from subsequent work and meta-analyses, the paper aims to test competing models (single c-factor, individual IQ-driven models, and multidimensional factor structures) and to evaluate antecedents (turn-taking, social perceptiveness, proportion female) and predictive validity for complex tasks and academic outcomes. The work is important for theories of group learning and team effectiveness, as understanding the factor(s) underpinning group performance could inform how teams are formed, assessed, and developed.

Prior research by Woolley et al. reported a single collective intelligence factor accounting for substantial variance across group tasks and predicting complex criterion tasks, with proposed antecedents including equal conversational turn-taking, higher social perceptiveness (Reading the Mind in the Eyes test), and greater proportion of women. Follow-up studies and a meta-analysis (Riedl et al.) supported positive manifolds and substantial variance explained but were limited to specific task batteries. Other reviews (Graf-Drasch et al.) found task structure effects, suggesting the c-factor applies mainly to well-structured tasks, while ill-structured tasks show multidimensionality. Rowe et al. conducted meta-analyses indicating modest predictive correlations for c (r ≈ .26) and limited evidence for individual IQ predicting group tasks, but emphasized methodological limitations. Bates and Gupta argued that average individual IQ accounts for most group-IQ differences (~80%), implying group performance reflects individual intelligence writ large. These mixed findings motivate testing alternative factor structures and clarifying the role of individual versus collective intelligence.

Design: Within-subjects correlational study with three phases (individual, group, prediction). Participants: 85 university students (Mean age = 25.21; 71.76% female; 96.47% full-time; 83.53% born overseas) allocated to 29 groups (mean group size = 2.93; range 2–5). Recruitment occurred over 6 weeks; inclusion required English fluency; consent obtained; compensation provided; ethics approved. Power analyses targeted detection of effects comparable to Woolley et al. (c-factor correlation with criterion task r = .52), estimating minimum 26 groups (two-sided) with ~85% observed power. Measures: Individual phase included demographics, ICAR 16-item cognitive ability test (operationalizing individual IQ), Reading the Mind in the Eyes (36-item social perceptiveness), and secondary measures (Big Five personality IPIP-50, BEIS-10 trait emotional intelligence, SIMS motivation, PCS cohesion, task satisfaction). Group phase included group IQ subtests: advanced multiple-choice vocabulary (ETS Kit, 18 items), open-form vocabulary (Mill Hill, 5 definitions), Raven’s Standard Progressive Matrices Plus (15 items), group brainstorming (5 minutes for novel ideas), and group memory/attention (10 multiple-choice questions from a 5-minute film clip). Group process measures included friendship status, communication recording with lapel microphones to quantify speaking turns and word counts, perceived cohesion, motivation, and satisfaction. Scoring: Multiple-choice items were auto-scored; open-form responses were rated independently by two researchers (97% agreement; disagreements resolved by discussion). Predictive phase: Criterion task Moon Landing Exercise (MLE), a judgmental, multi-domain group task requiring ranking 15 salvaged items based on survival utility; scoring by deviation from expert rankings; administered ~5 minutes after the group IQ battery. External validation: Group assignment grades collected from a subset of participants (n = 30) via verified documentation; for those with multiple grades, means were used. Analytic strategy: Exploratory factor analysis (principal axis factoring with Promax oblique rotation) and Horn’s parallel analysis (paran package in R) to determine factor retention. Comparative structural equation modeling (SPSS AMOS v27) tested four models: (A) single c-factor (Woolley), (B) individual IQ indirectly accounting for group-IQ via latent c (Bates & Gupta), (C) individual IQ directly accounting for group-IQ (correlational paths across subtests), and (D) a two-factor cFluid–cCrystal model based on Cattell’s Gf–Gc theory. Fit indices (χ², p, GFI, CFI, RMSEA) guided model selection. Correlational analyses (Pearson’s and Spearman’s where applicable) examined relationships between antecedents (speaking turn variance, social perceptiveness, proportion female), individual IQ metrics (average, maximum, g loadings), the group-IQ composite (raw scores across 49 items), factor loadings (cFluid, cCrystal), and outcomes (MLE, group assignment scores). One-way ANOVA tested proportion female effects on group-IQ. Linear regressions assessed predictive validity for MLE.

Factor structure: Horn’s parallel analysis suggested retaining three factors (accounting for 41%, 22%, and 20% of common variance), and comparative SEM favored a two-factor cFluid–cCrystal model over a single c-factor or individual IQ models. Model fits: Single c-factor (χ²(5)=10.251, p=.068, GFI=.883, CFI=.694, RMSEA=.194); individual IQ via latent c (χ²(9)=21.649, p=.010, GFI=.824, CFI=.652, RMSEA=.224); individual IQ direct (χ²(10)=27.731, p=.002, GFI=.746, CFI=.512, RMSEA=.252); cFluid–cCrystal (χ²(4)=6.926, p=.140, GFI=.914, CFI=.829, RMSEA=.162), selected for subsequent analysis. The cFluid and cCrystal factors correlated strongly with the group-IQ composite (r > .54) but were weakly correlated with each other (r = .25), consistent with a positive manifold (mean inter-subtest r ≈ .26). Antecedents: Variance in speaking turns (r = .19, p = .33), average social perceptiveness (r = .27, p = .16), and proportion female (r = −.17, p = .38) were not significantly related to the group-IQ composite; none showed significant associations with cFluid or cCrystal. One-way ANOVA found no effect of proportion female on group-IQ (F(6,22)=1.2, p=.34). Individual intelligence: Highest individual IQ, average individual IQ, and average individual g-loadings correlated strongly with group matrices (r = .69, .69, .60; all p < .001), suggesting alignment between individual and group performance on matched constructs. Criterion task prediction (MLE): No significant correlations or regressions for group-IQ, cFluid, cCrystal, or Woolley’s c-factor (e.g., group-IQ r = .11, p = .56; cFluid r = .10, p = .59; cCrystal r = .04, p = .84). External academic validation: Moderate, statistically significant associations for individual IQ with group assignment scores (Spearman’s rho r_s = .40, p = .014) and for group-IQ with group assignment scores (r_s = .47, p = .005), based on n = 30 verified participants. Secondary findings: Conscientiousness showed positive relationships with collective intelligence indicators (e.g., with cCrystal and group-IQ), aligning with prior literature on personality composition and group performance.

The findings challenge the single-factor conception of collective intelligence by supporting a multidimensional structure analogous to Cattell’s fluid and crystallized intelligence at the group level. This addresses RQ1 by demonstrating superior fit for a cFluid–cCrystal model compared to a unitary c-factor or models driven purely by individual IQ. Regarding RQ2, the canonical antecedents proposed by Woolley et al. (turn-taking equality, social perceptiveness, proportion female) did not relate significantly to collective intelligence measures, suggesting that previously reported effects may be context- or method-dependent. Evidence that individual IQ strongly relates to group performance on matched constructs (e.g., matrices) implies that some group-level variance may reflect aggregated individual-level cognitive abilities. For RQ3, collective intelligence measures did not predict performance on a near-term, complex judgmental task (MLE), contrasting with earlier reports. For RQ4, both individual and group intelligence measures moderately predicted university group assignment grades, suggesting practical relevance of cognitive abilities for real-world academic collaboration. Overall, the results advocate for nuanced models of group cognition, careful task sampling aligned with psychometric theory, and consideration of both individual and collective contributions to team outcomes.

This study proposes and supports a two-factor cFluid–cCrystal model of collective intelligence, paralleling established individual-level intelligence theory. It questions the dominance of a single collective intelligence factor and highlights the potential nesting of group-level cognition within individual-level abilities. Practically, both individual IQ and group-IQ measures showed utility in predicting academic group assignment outcomes, while collective intelligence measures did not predict immediate complex judgmental performance. The findings suggest caution in using c-factor assessments for organizational or educational decision-making until further validation clarifies their distinctiveness from individual intelligence. Future research should employ psychometrically validated, parallel task batteries across individual and group levels, collect richer process data in digital environments, conduct longitudinal studies to capture team dynamics, and develop methods to model within- and between-group variability (e.g., weighting individual contributions) to disentangle c from g.

The study’s face-to-face, paper-based context limited capture of rich process data (effort, time on task, conflict, nonverbal cues). External validation of academic outcomes relied on voluntary reporting, introducing potential social desirability bias. The subset providing grades was small (n = 30). The observed cFluid–cCrystal pattern may reflect individual-level intelligence structures rather than uniquely group-level constructs. The design was cross-sectional with immediate criterion task assessment, lacking longitudinal observation of team dynamics. Broader generalizability may be constrained by the sample’s demographics and the academic context.