Effort Opacity: Generative AI and Reciprocal Effort in Collaborative Work
Job Market Paper
Generative AI is often described as a technology for increasing individual productivity. But organizations do not convert individual output into collective value automatically. Contributions must be reviewed, improved, trusted, and integrated by others. This paper argues that generative AI creates an organizational problem: it can make human effort harder to infer from the work people produce. When actors use generative AI to produce work, the work output can obscure the role the human played producing it. Therefore, from the contribution alone, collaborators may struggle to infer how much effort the actor exerted compared to how much the AI produced. I call this inference problem effort opacity: the difficulty collaborators face in inferring, from the work product itself, how much effort fellow collaborators put into producing it. I test this argument in open-source software development, where individual contributions become collective value only when developers invest scarce attention in reviewing and integrating code through a dyadic collaborative process. Across one observational study and two pre-registered experiments, I look at what happens when GAI is used in a dyadic collaboration. In the context of dyadic collaboration, I find that work appearing or disclosed as GAI-assisted is associated with lower perceived human effort and receives less reciprocal effort compared to work that does not appear or disclose GAI assistance. I operationalize reciprocal effort as the costly attention collaborators give to others’ work, measured through context-specific outcomes: rejection and constructive feedback in the observational study, and prioritization decisions in an online experiment. In Study 1, I use a custom AI prediction algorithm to classify code contributions as being AI-generated or human-generated. Contributions predicted to be AI-generated are 31 percentage points more likely to be rejected and receive less constructive feedback, even after controlling for observable code quality. Study 2 validates the assumption behind this GAI classifier measure using the code from the corpus used to build the classifier. Because the true production process of GitHub pull requests is not observed, the experiment does not establish ground truth AI use in Study 1. Instead, it asks whether developers perceive GAI-generated and human-written code in the known-provenance corpus in ways that align with the classifier’s signal, and whether perceived GAI assistance reduces human effort attributions. Study 3 then tests whether reducing effort opacity can restore reciprocal effort. In a within-subjects 2×2 design where I vary AI disclosure and effort signals, developers become more willing to prioritize AI-assisted work when contributors display high effort cues. Together, the findings reveal a collaboration paradox: AI can make individual work easier to produce while making human effort harder to infer, weakening the reciprocity through which individual contributions become collective value.
Which Ideas Attract Workers? Gender and Labor Supply in Startups
with Solène Delecourt (UC Berkeley Haas), Katelyn Cranney (Stanford), and Rembrand Koning (Harvard Business School)
Revise & Resubmit at Management Science
Nominated for Best Conference Paper Award at the Strategic Management Society Annual Conference 2023
Nominated for Research Methods Paper Prize at the Strategic Management Society Annual Conference 2023
Which startups can attract the talent they need to grow? We argue that worker preferences over which problems to solve shape labor supply to early-stage firms. In male-dominated labor markets, startups focused on women may struggle to recruit, especially
for critical, male-dominated roles like engineering and senior leadership. We combine observational data on over 52,000 U.S. high-potential startups with a pre-registered field experiment. Linking startups to over 2.5 million worker histories, we find that
female-focused startups employ fewer workers: a one standard deviation increase in a startup’s female-focus predicts 11% fewer male employees. To isolate worker preferences, we run a field experiment where we randomize the startup idea itself, thereby
developing a new experimental paradigm to causally identify how the nature of an idea matters. We built a job-search platform recruiting real tech job seekers in a within-subject design, where each participant evaluates 15 startup ideas, each randomized
to be more or less female-focused. Crucially, our experiment is incentive-compatible, because participants’ evaluations affect which real job postings our platform subsequently recommends to them. Male job seekers are 6.8 percentage points less likely to
apply to female-focused startups than other startups, a 10% decrease from baseline. Offsetting the male penalty would require approximately $47,000 in additional yearly compensation. Our findings demonstrate that who benefits from an idea is an important nonpecuniary job attribute shaping talent acquisition, potentially helping explain why ventures serving women grow more slowly.
The Fabricated Front: Generative AI and the Opacity of Workplace Performance
with Tom Van Nuenen (D-Lab, UC Berkeley) and Pratik Sachdeva (D-Lab, UC Berkeley)
Under Review at the PCAIDE Conference Journal – The Paris Journal on AI & Digital Ethics
Generative AI (GenAI) has become a fixture of workplace life. Current research asks chiefly what this implies for jobs and outputs, measured in productivity, displacement, or bias. What remains underexamined are the interactional reconfigurations that GenAI produces at work. Chopra’s (2026) theory of effort opacity has begun to fill this gap by noting the systematic decoupling of observable output from human engagement. When GenAI makes interactional cues less diagnostic, it weakens the reciprocal exchange that sustains collaborative trust. Extending this account of effort opacity, we examine the interactional mechanics that produce opacity in everyday workplace encounters. Drawing on Erving Goffman’s dramaturgical framework and 1,250 interview transcripts from Anthropic’s AI Interviewer dataset, we identify five opacity mechanisms through which workplace fronts are reorganized: voice (whose stance the words index), provenance (who can stand behind the artifact), vulnerability (whether the worker is uncertain), attention (whether the worker is engaged), and investment (how much labor the output reflects). We show that professionals defend the identity mechanisms while freely producing opacity around the labor mechanisms, and trace this asymmetry to the output-centered organization of contemporary work, where deliverables already stand in for the labor process
that produced them. The governance task, accordingly, is one of involvement management: specifying which forms of human involvement (attention, effort, judgment) must remain inspectable, and to whom. Workplace AI policies built on universal disclosure will systematically misrecognize a social field in which inspectability is already audience-relative.
Whose Transparency Is It Anyway?
with Genevieve Smith (BAIR, UC Berkeley), Natalia Luka (BAIR, UC Berkeley), and Min Kyung Lee (UT Austin)
Transparency has become a central principle in generative AI governance, yet there is little agreement about what transparency should actually mean, for whom, and in what contexts. Drawing on the Social Construction of Technology and Value Sensitive Design, this paper examines how different stakeholders understand and experience transparency in generative AI. We use a two-phase mixed-methods study: a survey of 400 end users, AI builders, and policymakers, followed by participatory design sessions with end users. We find that stakeholders broadly agree that transparency matters, but differ in what kinds of transparency they need and whether those needs are currently being met. In particular, existing transparency tools appear to serve builders more effectively than end users, revealing how power shapes which interpretations of transparency become embedded in practice. We argue that meaningful AI transparency requires moving beyond one-size-fits-all disclosures toward participatory, pluralistic transparency infrastructures that reflect the needs of diverse stakeholders.
The Technical Debt Tradeoff: Generative AI, Shipping Speed, and the Future Cost of Work
with Dhruv Agarwal (Middleware HQ)
Generative AI promises to help organizations build and ship products faster, but it may also change the amount and kind of technical debt they accumulate. Technical debt is often treated as a software problem, but it is also an organizational tradeoff: teams sometimes take shortcuts to move quickly, release products, and respond to market pressure, while creating future costs for maintenance, coordination, and reliability. This project examines whether GAI changes that tradeoff. In collaboration with a B2B SaaS company that works with organizations on software delivery and productivity data, we use de-identified workplace data to study whether GAI use is associated with changes in technical debt and whether those changes help or hinder the shipping of products over time. The project asks whether GAI enables organizations to ship more by reducing development bottlenecks, or whether it accelerates the creation of work that is harder to review, maintain, and build on. By treating technical debt as an organizational outcome rather than only a code-quality issue, this study examines how GAI may reshape the relationship between speed, productivity, and long-term organizational capacity.
From AI Tools to AI Systems: How Organizations Redesign Work Around Generative AI
with Calista Cooper (Vendasta) and Kim Coutts (Vendasta)
Generative AI is often introduced into organizations as a tool for individual productivity, but its organizational effects depend on whether firms can turn scattered experimentation into reliable systems of work. This project examines how organizations move from employees using AI on their own to building structured, governed, GAI-enabled workflows. In collaboration with a technology company undergoing an AI-first transformation, we study how GAI adoption changes the way organizational knowledge is documented, shared, governed, and used. The project focuses on a shift from individual GAI use toward systems in which humans increasingly design, manage, and oversee GAI agents, workflows, and knowledge infrastructures. We ask how organizations decide what knowledge becomes usable by GAI, how roles change when employees become builders and orchestrators of GAI-enabled work, and what forms of governance are needed when GAI systems begin carrying out organizational tasks. By examining GAI adoption as a process of organizational redesign rather than simple tool use, this study contributes to research on technology and work by showing how firms build the infrastructure needed to use GAI at scale.
Sahiba Chopra 