Generative AI and the Restructuring of Workplace Collaboration

This is a 2×2 design that crosses team structure (Individual vs. Team) with technology access (No AI vs. AI Access), giving four conditions: Individual with no AI, Individual with AI, Team with no AI, and Team with AI. Our main interest is the effect of AI access within the Team conditions. Within those Team conditions we also randomize the basis of the performance bonus, rewarding workers either on expert ratings of their output or on their collaborator's evaluation. Workers are recruited through Upwork's normal hiring process against set skill and eligibility criteria, complete a real deliverable on a platform we built, and can consult a supervisor who holds firm-specific information. We log all communication and AI use, and independent expert raters evaluate the output.
Beyond the main 2×2 contrast, we will run three secondary analyses. First, supervisor consultation across the AI and No AI conditions, which identifies whether AI substitutes for the firm-specific knowledge the supervisor holds or changes the cost of consulting them (H1, H2). Second, the bonus basis within the Team conditions, which examines whether the return to collaboration drives AI's effect on collaborative behavior, comparing pairs rewarded on their collaborator's evaluation to those rewarded on expert ratings (H1). Third, worker status, which investigates whether AI's effect on collaboration differs across the status distribution (H3), using the baseline characteristics on which we stratify. Because Upwork makes worker ratings and experience public, we have a platform-validated measure of status to draw on here.

2026-07-22

2026-11-30

Our main outcomes fall into two groups. The first is how workers collaborate: how often a worker consults peers and the supervisor; when in their process those consultations happen, and the content of the collaboration, both its routine versus non-routine cognitive content and its communicative function. The second is the work itself: the quality, novelty, and firm-specific integration of the output, judged by independent expert raters at the component and the full-deliverable level.

The collaboration measures come from the platform's logs of communication between teammates and with the supervisor. Frequency is a count of the consultations a worker initiates. Timing is where those consultations fall across the stages of the task, summarized as the share that happen before versus after the worker drafts their own component. Content is built by classifying each exchange two ways: whether it is routine or non-routine cognitive work (Autor, Levy & Murnane 2003), and what function it serves (explanatory talk, complex problem-solving, integrative critique, help-seeking, coaching, or task-information sharing). We automate this classification and have trained research assistants cross-validate it. Output quality, novelty, and firm-specific integration come from independent expert evaluations collected on Prolific, where each deliverable is rated by several raters who do not know which condition produced it.

How workers use AI, measured the same way as peer collaboration: the frequency, timing, and content of their AI queries. How often workers consult the supervisor across the AI and No AI conditions. Workers' ratings of their teammate after the task. And post-task self-reports on how they used firm-specific information, whether they found the collaboration and the AI useful, whether they would take on a similar project again, and which working arrangement they would prefer.

The AI and supervisor measures come from the same platform logs as the peer-collaboration measures. The teammate ratings and self-reports come from a short post-task survey. Together these support the secondary analyses: whether AI substitutes for the firm-specific knowledge the supervisor holds, whether the return to collaboration drives AI's effect on how workers collaborate, and how AI's effect on collaboration differs across the worker status distribution.

This is a 2×2 design that crosses team structure (Individual vs. Team) with technology access (No AI vs. AI Access), giving four conditions: Individual with no AI, Individual with AI, Team with no AI, and Team with AI. Our main interest is the effect of AI access within the Team conditions. Within those Team conditions we also randomize the basis of the performance bonus, rewarding workers either on expert ratings of their output or on their collaborator's evaluation. Workers are recruited through Upwork's normal hiring process against set skill and eligibility criteria, complete a real deliverable on a platform we built, and can consult a supervisor who holds firm-specific information. We log all communication and AI use, and independent expert raters evaluate the output.

Not available

. We sort workers into blocks defined by baseline characteristics we observe before assignment — platform tenure, prior work experience, and demographics — and randomize them to conditions within each block.

The unit assigned to a condition is the individual worker in the Individual conditions and the worker pair in the Team conditions. The bonus-basis treatment is randomized within the Team conditions at the pair level.

Yes

About 110 individual workers in the Individual conditions plus 110 worker pairs in the Team conditions — roughly 220 randomized units.

330 workers: 110 in the Individual conditions and 220 in the Team conditions.

Individual conditions: about 55 workers with no AI and 55 with AI. Team conditions: about 110 workers (55 pairs) with no AI and 110 workers (55 pairs) with AI, with the bonus-basis treatment randomized within each Team arm.

The design detects a minimum effect of about 0.40 standard deviations on the main collaboration outcome at 80% power and two-sided α = 0.05, following Athey & Imbens (2017). This is in line with effects from recent field experiments on generative AI and teamwork, which fall around 0.2 to 0.45 standard deviations on output and quality measures (Noy & Zhang 2023; Dell'Acqua et al. 2025), and is comparable to the teamwork effects found in online labor markets (Lyons 2017).