PaperlayerView exact revision on arXiv

Source-linked AI summary

The Effects of Generative AI on Collaboration and Social Networks Among Knowledge Workers

Giuseppe Soda, Alessandro Iorio, Leonardo Rizzo

arXiv:2610.32461v1physics.soc-phcs.CY

TL;DR

The paper asks whether generative AI shrinks or expands organizational collaboration networks. Using staggered difference-in-differences and fixed-effects instrumental-variable models, it finds that users collaborate more broadly while relying more on trusted colleagues.

  • Problem

    The paper asks whether generative AI adoption shrinks collaboration networks through substitution or expands them through new coordination needs.

    Source: §1 · p. 4

  • Method

    The study combines staggered difference-in-differences estimates of adoption with fixed-effects instrumental-variable estimates of Copilot-use intensity, instrumented by feature releases interacted with baseline exposure.

    Source: §2 · pp. 9–10

  • Results

    Generative AI users collaborate more frequently, draw on broader groups, move beyond their original small groups, and rely more on a few trusted colleagues.

    Source: §4 · pp. 18–19

  • Takeaways & Limitations

    The patterns suggest that interpersonal interactions help workers check and validate AI-assisted outputs and decide whether to accept, change, or reject them.

    Source: §4 · p. 19

    The study covers one business unit at one multinational company, so its findings may not generalize to other industries, workplaces, job types, or institutional settings.

    Source: §4 · p. 21

Abstract

from arXiv · show

The rapid diffusion of generative artificial intelligence has raised concerns that knowledge work may become less dependent on human-to-human interaction. Yet because AI-generated outputs require interpretation, verification, and social legitimation, generative AI may instead increase collaboration among workers. We test these competing predictions using passively collected behavioral data from more than 12,000 knowledge workers observed over 13 months during the staged organizational rollout of an AI-powered assistant (Microsoft Copilot). Using two complementary quasi-experimental designs -- a staggered difference-in-differences model and a fixed-effects instrumental-variable approach -- we estimate the effects of generative AI on workers' collaboration behavior and social network structure. We estimate that generative AI increased collaboration time across communication channels and reshaped individual networks. Those using generative AI interacted with more colleagues and their networks became less clustered, indicating broadening in collaboration structures. At the same time, they also concentrated more of their collaboration on ties with stronger prior histories, deepening reliance on already trusted colleagues. Together, these findings suggest that, while generative AI changes how people work, it may not displace human-to-human interactions.

1 Main

The paper asks whether GenAI substitutes for colleague interaction or creates new demands for validation, and examines how adoption changes collaboration networks.

1 Main

  • Prior research largely examines human-AI dyads, leaving limited knowledge about how GenAI changes interaction, dependence, and coordination among workers.The paper frames the unresolved choice as whether adoption shrinks or expands workplace collaboration networks.Source: §1 · pp. 2, 4
  • Because AI outputs can be difficult to assess, workers may depend on colleagues to understand, question, and legitimate them.The authors describe this gap between rapid output production and the time needed to understand and defend an output as epistemic debt.Source: §1 · pp. 3–4
  • The study uses behavioral data from more than 12,000 workers during a staged Copilot rollout to estimate effects on collaboration time and network structure.It combines staggered difference-in-differences with fixed-effects instrumental-variable models to address adoption timing and usage endogeneity.Source: §1 · p. 4
  • GenAI users collaborate more, reach more colleagues, show lower network clustering, and concentrate more collaboration on colleagues with strong prior histories.The pattern is broadening without uniform expansion: workers rely more heavily on a subset of already-trusted colleagues.Source: §1 · p. 4

2 Methods

The study measures digital collaboration and Copilot use in a monthly worker panel, then estimates adoption and usage effects with complementary quasi-experimental designs.

2 Methods

  • The analysis follows 12,040 employees in one multinational company's focal business unit from September 2024 to September 2025 during Copilot's staged rollout.After excluding 172 always-adopters, the estimation sample contains 11,868 workers: 1,992 adopters and 9,876 never-adopters.Source: §2 · pp. 5–7
  • Generated load measures monthly meeting, call, email, and chat hours initiated by a worker, with overlapping activities deduplicated.It captures active outreach rather than communication received from colleagues; total collaboration hours and modality-specific hours are also examined.Source: §2 · p. 7
  • Collaboration outcomes and Copilot usage come from separate Viva Insights systems, while collaboration measures draw on Microsoft 365 activity logs.Weekly usage and workload measures are aggregated to the monthly panel; network observations are also monthly.Source: §2 · pp. 5–7
  • Monthly directed, weighted network ties require at least one collaboration hour; out-degree, clustering, and preferential reinforcement capture distinct network dimensions.Preferential reinforcement is the correlation between current tie weights and cumulative past weights across a worker's outgoing ties.Source: §2 · p. 8
  • The staggered difference-in-differences design compares adopters' outcome changes with never-adopters' changes, conditional on worker characteristics and common time shocks.The Callaway and Sant’Anna estimator uses doubly robust weighting; its key assumption is conditional parallel trends, alongside no anticipation.Source: §2 · pp. 8–9
  • Fixed-effects instrumental-variable models estimate the intensive margin using monthly Copilot feature releases interacted with workers' baseline exposure to Copilot users.Never-adopter placebo and exposure-by-month tests provide support for the instrument, but cannot prove its exclusion restriction.Source: §2 · pp. 9–11

3 Results

Copilot adoption was followed by increased collaboration time and networks that broadened while deepening reliance on strong prior ties. Fixed-effects IV estimates report similar collaboration-time and network patterns for more intensive use.

3.1 Effects of Generative AI on Collaboration Time

  • Generated collaboration averaged 1.430 hours over the first three post-adoption months, while the overall ATT of 0.948 was not statistically distinguishable from zero.Event-study estimates show increases in both synchronous and asynchronous collaboration; estimates remained consistent when directly connected never-adopters were excluded.Source: §3.1 · pp. 12–14

3.2 Effects of Generative AI on Social Networks

  • Copilot adoption increased outdegree and reduced clustering, while strong-tie reliance rose in event studies but its overall ATT was not statistically significant.The reported overall ATT estimates are 2.179 for outdegree and −0.009 for clustering; the strong-tie ATT is 0.008 (p = 0.088).Source: §3.2 · p. 15

3.3 Additional Analysis: TWFE-IV

  • In fixed-effects IV estimates, higher Copilot intensity predicted more collaboration time and broader, less clustered networks with stronger ties.A tenfold increase in monthly actions corresponded to about 3.4 additional total collaboration hours; asynchronous effects were significant, while user-generated synchronous estimates were not.Source: §3.3 · pp. 17–18

4 Discussion

GenAI appears to reorganize rather than displace collaboration: workers collaborate more broadly while also relying more on trusted colleagues.

4 Discussion

  • Because AI outputs are probabilistic and opaque, workers may consult colleagues for context, independent judgment, and social validation.These interactions can help determine whether AI-assisted outputs are accepted, changed, or rejected.Source: §4 · p. 19
  • GenAI use is associated with broader, less-clustered collaboration and heavier interaction with a subset of already-trusted colleagues.These patterns suggest workers seek expertise and independent judgment while relying more heavily on trusted ties.Source: §4 · pp. 18–19
  • The findings add a relational dimension to GenAI research centered on productivity and cognition, showing that adoption reshapes workers’ social structure.The authors note that this reshaping differs from other major technology-enabled transitions.Source: §4 · p. 19
  • Broader, less-clustered networks may bring in less-overlapping evaluators, while trusted colleagues offer informed judgment but may also receive less scrutiny.The data cannot distinguish validation from endorsement because they capture interaction partners and duration, not discussion content.Source: §4 · p. 20
  • Organizations seeking to offset GenAI’s opacity through collaboration need norms that reward questioning AI-assisted work and disclosing its use.The authors also call for research on when collaboration improves judgment versus creates friction or cognitive overload.Source: §4 · pp. 20–21
  • Network-level analysis finds path lengths and diameter decline, global efficiency rise, and clustering fail to keep pace with degree, yielding a more integrated than locally cohesive network.The authors leave open how individual-level collaboration shifts scale up to the network’s overall structure.Source: §4 · pp. 20–21
  • The study covers one business unit and only the first three post-adoption months, does not observe discussion content, and relies on causal assumptions.The authors note that transferability and whether collaboration patterns persist over time remain unknown.Source: §4 · pp. 21–22

Tables and Figures

The figures track collaboration and network outcomes around Copilot adoption and summarize staggered difference-in-differences estimates. The tables report first- and second-stage instrumental-variable regressions.

Figures

  • Figure 1 plots adopters’ monthly outcome trajectories around first Copilot use, normalized to the pre-adoption baseline and shown with 95% confidence intervals.Values are detrended by calendar-month means and expressed in pre-period standard-deviation units; the shaded region marks the adoption month.Source: Figure 1 · p. 25
  • Figure 2 shows collaboration-outcome event-study estimates from t = −3 to 2, with separate outcome scales and pre-treatment estimates used to assess parallel trends.Markers distinguish pre- and post-treatment estimates; bars show 95% confidence intervals.Source: Figure 2 · p. 26
  • Figure 3 compares each outcome’s average effect over the first three post-adoption months with its overall ATT across event times and adoption cohorts.Estimates and confidence intervals are standardized by their standard errors, with raw coefficients also reported.Source: Figure 3 · p. 27
  • Figure 4 presents network-outcome event-study estimates from t = −3 to 2, with separate outcome scales and pre-treatment estimates used to assess parallel trends.Markers distinguish pre- and post-treatment estimates; bars show 95% confidence intervals.Source: Figure 4 · p. 28

Tables

  • Tables 1–3 report first-stage and second-stage IV regressions, with separate regressions for collaboration and network outcomes and user and month fixed effects.The first-stage outcome is total Copilot actions; the second-stage specifications instrument Copilot actions using a shift-share instrument. Network measures include out-degree, clustering, and preferential reinforcement.Source: Tables and Figures · pp. 29–30, Table 1 · p. 29, Table 2 · p. 30
Loading 2610.32461v1…