Re-ranking recommendations led teams to pick more different-race and different-gender collaborators, and nudged teams outscored self-assembled ones on creativity

Re-ranking recommendations by heterogeneity criteria significantly shifted collaborator choices: in the Nudged condition participants were more likely to invite different-race (β=0.69, p<0.01) and different-gender (β=0.68, p<0.01) collaborators; in the top-5 recommendations, different-race individuals rose from 2.47 to 2.85 and different-gender from 1.85 to 2.53, with different-race invitations rising from 56.03% to 73.47% and different-gender from 40.52% to 64.29%. Prior CSCW work showed that explicitly displaying diversity information (scores or badges) can paradoxically strengthen users' attachment to similar others; this study shows that merely reordering recommendations, without making the criteria salient, can subtly counteract homophily while preserving choice. Generalized linear mixed-effects model on recommender trace data, with 4,351 recommendations from 185 participants as observations, random effects for senders, standardized variables, and VIF below 10.

Granting user agency reduced teams' surface-level differences: teams in agency conditions showed significantly lower surface-level heterogeneity than assignment conditions (F(1,79)=11.53, p<0.01, η²=0.13), with Assigned-Diverse teams significantly more heterogeneous than Self-Assembled (Δ=−0.62, p_adj<0.01) and Nudged (Δ=−0.50, p_adj<0.05); gender showed the largest disparity, with Assigned-Diverse highest (M=0.83) and Self-Assembled lowest (M=0.43). The study moves the agency–homophily relationship from correlational discussion to experimental causal evidence and separates surface- from deep-level differences, finding agency affected only surface-level differences, not deep-level skill differences (F(1,79)=1.84, p>0.10), because participants actively sought skill-complementary collaborators across conditions (project skills accounted for 49% of search queries). Team-level two-way ANOVA on N=83 teams, using Blau index and coefficient of variation for heterogeneity, with Bonferroni correction and Tukey HSD post-hoc comparisons.

Including heterogeneity criteria alone did not significantly increase overall surface- or deep-level differences (F(1,79)=2.08 and 0.62, both p>0.10), but assignment-based heterogeneity carried process costs: Assigned-Diverse members reported the highest perceived surface-level differences (M=3.26 vs. Self-Assembled M=3.0, p_adj<0.05) and significantly lower communication indegree than Random (Δ=−0.52, p_adj<0.01) and Self-Assembled (Δ=−0.47, p_adj<0.05). The study advances the “double-edged sword” discussion from whether composition is diverse to how diversity is achieved, showing that assignment-based optimization triggers social categorization and communication barriers, whereas nudged choice did not show comparable communication declines. Individual-level two-way ANOVA (N=332) combined with descriptive analysis of open-ended responses and Zoom transcripts generated by Otter.AI and reviewed by trained research assistants.

The interaction of agency and heterogeneity criteria significantly affected team creativity (F(1,79)=8.23, p<0.01, η²=0.09): Nudged teams scored significantly higher than Self-Assembled teams (Δ=0.45, p_adj<0.05), while agency alone or heterogeneity criteria alone showed no significant improvement over Random teams (Δ=−0.24 and −0.16, both p_adj>0.10). The result indicates that the effect of organizational values (heterogeneity) depends on how they are embedded in the algorithm, not merely whether they are included; nudged choice introduced enough variation while preserving psychological ownership. Team-level two-way ANOVA on N=83 teams; creativity rated by ten undergraduate marketing and communication majors blind to condition on 5-point originality and uniqueness scales, Cronbach's α=0.96.

The work speaks to researchers and organizational practitioners designing team-formation recommender systems, in settings where collaboration platforms must balance individual autonomy with collective goals. Its conclusions rest on a laboratory 30-minute creativity task, four-member teams, and an online Zoom environment, with participants drawn mainly from one university's students, staff, and local residents and not stratified by discipline or expertise. The study suggests that embedding heterogeneity criteria through re-ranking, while preserving choice, can improve team creativity without significantly raising members' awareness of differences or reducing communication, whereas direct assignment maximizes surface-level heterogeneity but comes with reduced communication. This positions nudged choice as a design option between free choice and forced assignment, and signals that organizations delegating value judgments to algorithms need accompanying transparency, explanation, and contestation mechanisms.

Readers should still watch: heterogeneity criteria alone did not significantly raise overall surface- or deep-level differences, suggesting nudging cannot fully offset homophily when agency is retained; deep-level differences were largely non-significant, possibly because the participant pool was relatively homogeneous and not stratified by discipline; the nudge worked without users' awareness, leaving transparency and contestability open; the 30-minute fictitious creativity task means teams had no past or future, and heterogeneous teams may need more time and support to translate composition into performance in longer-term settings; and the heterogeneity metric, the sequential team-assembly procedure, and the limited information participants saw (names and photos) may all shape results, so alternative metrics and more realistic platforms remain to be tested.