« Back to Results

Who Wins and Loses with Generative AI?

Paper Session

Saturday, Jan. 4, 2025 2:30 PM - 4:30 PM (PST)

Hilton San Francisco Union Square, Union Square 23 and 24
Hosted By: American Economic Association
  • Chair: Sonia Jaffe, Microsoft

The Impact of Generative AI on Collaboration at Work

Eleanor Dillon
,
Microsoft
Nicole Immorlica
,
Microsoft
Sonia Jaffe
,
Microsoft
Danielle Li
,
Massachusetts Institute of Technology

Abstract

Generative AI tools are increasingly capable of performing complex knowledge-based tasks, fueling excitement about their potential to substantially increase worker productivity. At the same time, a key question remains: whose work will be augmented and whose will be replaced? Forecasting the impact of generative AI on labor demand (and understanding its attendant implications for economic inequality) requires understanding not just what tasks new AI systems are good at, but how their use shapes how responsibilities and opportunities are allocated to workers.
We conduct a novel large-scale field experiment that randomizes access to a general purpose LLM-based tool to assist information workers with their regular work. Our study includes data from 60 major multinational firms across a range of industries. Each participating firm identified 100 employees, who were randomized (50/50) into receiving early access to Copilot for Microsoft 365 (Copilot). Firms agreed to maintain this randomized access to Copilot for six months so we can follow the evolution of work patterns as workers adapt to using this new AI tool.
Copilot integrates GPT-like capabilities (e.g. creating and summarizing) into Microsoft Word, Excel, PowerPoint, Outlook and Teams, which these workers already use regularly. Using data from Microsoft applications, we observe email and meeting behavior and collaborative document creation for all workers at these firms. We contrast the behavior of four distinct groups: treated workers, their managers and teammates, control workers, and their managers and teammates.
Treated workers edit more documents and use more advanced (non-Copilot) features available in Word/Excel/PowerPoint. They also view fewer emails (presumably because they use Copilot’s summary function), reply to emails faster, and participate in shorter email threads. We are also analyzing how collaboration networks evolve and how access to the tool affects the division of tasks within teams.

The Uneven Impact of Generative AI on Entrepreneurial Performance

Nicholas Otis
,
University of California-Berkeley
Rowan Clarke
,
Harvard Business School
Solene Delecourt
,
University of California-Berkeley
David Holtz
,
University of California-Berkeley
Rembrand Koning
,
Harvard Business School

Abstract

There is a growing belief that scalable and low-cost AI assistance can improve firm decision-making and economic performance. However, running a business involves a myriad of open-ended problems, making it hard to generalize from recent studies showing that generative AI improves performance on well-defined writing tasks. In our field experiment with 640 Kenyan entrepreneurs, we as- sessed the impact of AI-generated advice on small business revenues and profits. Participants were randomly assigned to a control group that received a standard business guide or to a treatment group that received a GPT-4-powered AI busi- ness mentor via WhatsApp. We are unable to reject the null hypothesis that generative AI access has no impact, but are able to rule out the large effect sizes reported by other studies of generative AI’s economic impact. Our over- all null result masks treatment effect heterogeneity with respect to the baseline business performance of the entrepreneur: our point estimates suggest that high performers benefited by just over 15% from AI advice, whereas low performers did about 8% worse with AI assistance. Exploratory analysis of WhatsApp in- teraction logs shows that both groups sought the AI mentor’s advice, but that low performers did worse because they sought help on more challenging business tasks. Our findings highlight the potential and limitations of generative AI to enable entrepreneurs across the globe.

Generative AI and the Nature of Work

Manuel Hoffmann
,
Harvard Business School
Sam Boysel
,
Harvard Business School
Frank Nagle
,
Harvard Business School
Sida Peng
,
Microsoft
Kevin Xu
,
Github

Abstract

Artificial intelligence (AI) is at the beginning of a technological evolution that is slowly and steadily reaching all sectors of the economy. However, due to the novelty and breadth of AI, we still know relatively little about the labor impacts of AI, including both the effects on strategic human capital decisions as well as decentralized, often voluntary, contributions to public goods. Therefore, it imperative to understand how different work processes can be affected by this transformational technology. Using the context of open source software, we study the individual consequences of using a generative AI tool designed to help developers work more efficiently and effectively. We exploit a natural experiment arising from the deployment of GitHub Copilot, a generative programming AI released by the social coding platform in June 2022. We study the impact of generative AI for a whole year on millions of work activities by exploiting threshold eligibility requirements through a quasi-experimental regression discontinuity approach. We find that after having access to Copilot, top maintainers are able to code more and spend less time on project management tasks. Similar to other modern technologies, generative AI also induces maintainers to substitute away from high cost interactions (i.e. comments) towards low cost interactions (i.e. emoticons). Overall, our estimates point towards a large potential for AI to transform work processes and social interactions.

Impact of Github Copilot on Software Engineer Skill Acquisition and Hiring and Retention Patterns

Matthew Baird
,
LinkedIn
Mar Carpanelli
,
LinkedIn
Brian Xu
,
LinkedIn

Abstract

Generative AI is poised as a revolutionary technology to transform the workforce. One area of particular potential is for software engineers (SWE), who can use the technology to assist with programming and other related tasks. To evaluate the impact of GAI, we merge data from LinkedIn and Github at the company-by-month level in the US. We use this merged data and estimate Callaway-Sant’Anna’s difference-in-difference with staggered treatment model to investigate the impact of Github copilot adoption on several outcomes. Specifically, we examine the extent to which copilot adoption leads to faster programming and non-programming skill development, as well as the impact on labor market outcomes (job posts, job applications, and new hires), in terms of frequency and type (seniority of position and worker, types of skills held by workers).

Discussant(s)
Rembrand Koning
,
Harvard Business School
Sam Boysel
,
Harvard Business School
James Brand
,
Microsoft
Teng "Ted" Liu
,
Upwork Inc.
JEL Classifications
  • J4 - Particular Labor Markets
  • O3 - Innovation; Research and Development; Technological Change; Intellectual Property Rights