"Generative AI and the Nature of Work" is a 2025 paper by Hoffmann, Boysel, Nagle, Peng, and Xu, affiliated with Harvard Business School, Microsoft, and GitHub. It examines how access to generative AI changes not only the productivity of software developers but the composition of their work, drawing on a natural experiment from the deployment of GitHub Copilot that covers millions of work activities over two years.
It circulates as Harvard Business School Working Paper No. 25-021, first issued October 2024 and revised April 2025, and as CESifo Working Paper Series No. 11479; the version dated April 18, 2025 runs 75 pages (Source: papers.ssrn.com). The paper assigns JEL classifications H4, O3, and J0 and lists the keywords generative artificial intelligence, digital work, open source software, and knowledge economy (Source: papers.ssrn.com). Author affiliations are listed differently across outlets: the SSRN record places Manuel Hoffmann and Sam Boysel at Harvard Business School, Frank Nagle at MIT and the Linux Foundation, Sida Peng at Microsoft, and Kevin Xu at GitHub (Source: papers.ssrn.com), while a later UC Irvine account lists Hoffmann at UC Irvine's Paul Merage School of Business, Boysel and Nagle at Harvard Business School, Peng at Microsoft, and Xu at GitHub (Source: merage.uci.edu).
Summary of findings
The paper studies how software developers reallocate their tasks when given access to GitHub Copilot, distinguishing the effect on the nature of work from the effect on the volume of output. Developers with Copilot access shift toward core coding activities and away from non-core project management.
The authors state that little prior attention had been paid to how AI changes the nature of work itself, as distinct from the productivity effects documented in an emerging literature, and that the study addresses how individuals in the knowledge economy adjust how they work when they begin using AI (Source: papers.ssrn.com). They attribute the shift to two mechanisms: an increase in autonomous rather than collaborative work, and an increase in exploration rather than exploitation. The effects are larger for lower-ability individuals, which the authors describe as AI flattening the ability distribution. They argue that AI therefore has the potential to flatten organizational hierarchies in the knowledge economy.
Setting and method
The study uses open source software development on GitHub as its setting. It exploits a natural experiment created by the public launch of GitHub Copilot in June 2022, when GitHub granted free access to a subset of its most active developers based on an internal eligibility ranking whose precise criteria the developers themselves did not know (Source: merage.uci.edu). Because eligibility near the cutoff was effectively random, the authors use the program eligibility threshold within a quasi-experimental regression discontinuity design to estimate causal effects on task allocation, drawing on millions of panel observations of work activities over a two-year period (Source: papers.ssrn.com). The UC Irvine account describes the dataset as covering nearly 190,000 developers observed weekly from June 2022 through June 2024 (Source: merage.uci.edu). The authors report that their results are robust to alternate identification strategies, bandwidth and kernel selections, and variable definitions (Source: papers.ssrn.com).
Reported magnitudes
The UC Irvine summary reports the following estimated effects of Copilot access. Coding activities as a share of total work rose by more than 12 percent, while project management activities — coordinating with others, triaging issues, reviewing contributions, and related community overhead — fell by nearly 25 percent (Source: merage.uci.edu). The number of distinct collaborators that Copilot-eligible developers interacted with dropped by roughly 79 percent relative to baseline, which the authors read as AI substituting for the role a colleague would otherwise fill (Source: merage.uci.edu).
On exploration, developers with access engaged with roughly 15 more new repositories on average and expanded their exposure to programming languages by more than 21 percent, gravitating toward languages associated with higher market salaries (Source: merage.uci.edu). The effect on coding for lower-ability developers — those with shorter tenure, fewer platform achievements, and less established reputations — is reported as more than four times the effect observed in higher-ability peers under some measures (Source: merage.uci.edu). The authors note that the effects held across both volunteer contributors and developers working on behalf of employers (Source: merage.uci.edu).
Authors' interpretation
The authors interpret the reduced reliance on other people as evidence that AI can lower the need for the management layers that exist to coordinate interactions, and that hierarchies premised on coordinating traffic become less necessary when developers can resolve problems independently (Source: merage.uci.edu). Hoffmann frames the innovation finding as conditional on how firms deploy the technology, arguing that AI "can be an equalizer" when used for innovation rather than cost-cutting, but only for organizations "on the innovation side and less on the efficiency side" (Source: merage.uci.edu). He also argues that lower-ability individuals can "learn more quickly how to engage in certain types of work and potentially move up the hierarchy more quickly when they have access to the tools," and that relieving overhead on overburdened "linchpin" contributors lets them focus on work only they can do (Source: merage.uci.edu).
Status
As of April 2026 the working paper was under review at the journal Management Science (Source: merage.uci.edu). The SSRN record reports more than 10,500 downloads and over 34,000 abstract views (Source: papers.ssrn.com).
Relation to other work
The study connects to research on how AI changes the type of work performed rather than only its quantity, with the largest changes accruing to lower-skill workers (AI Labor Disruption). Its task-level granularity complements Stanford HAI's findings on the organizational-structure determinants of corporate AI projects (Source: hai.stanford.edu). It serves as a companion to Brynjolfsson et al.'s Generative AI at Work, which measures productivity gains where this study measures task reallocation. The documented shift from collaborative to autonomous work parallels the broader transition toward Agentic AI.
Provenance
PDF converted to markdown with images on 2026-04-13. Additional detail on the paper's status, method, and reported magnitudes added 2026-06-09 from the SSRN working-paper record (Source: papers.ssrn.com), the Harvard Business School faculty publication record (Source: hbs.edu), and a UC Irvine Paul Merage School of Business feature on the research (Source: merage.uci.edu).