This website uses cookies.

By clicking the "Accept" button or continuing to browse our site, you agree to first-party and session-only cookies being stored on your device to enhance site navigation and analyze site performance and traffic. For more information on our use of cookies, please see our Privacy Policy.

  1. Home
  2. ASSA Annual Meeting
  3. 2025
  4. Program
  5. Productivity and Competition Effects of AI
« Back to Results

Productivity and Competition Effects of AI

Paper Session

Saturday, Jan. 4, 2025 8:00 AM - 10:00 AM (PST)

Hilton San Francisco Union Square, Union Square 5
Hosted By: Industrial Organization Society
  • Chair: Imke Reimers, Cornell University

The Productivity Effects of Generative AI: Evidence from a Field Experiment with GitHub Copilot

Kevin Zheyuan Cui
,
Princeton University
Mert Demirer
,
Massachusetts Institute of Technology
Sonia Jaffe
,
Microsoft
Leon Musolff
,
University of Pennsylvania
Sida Peng
,
Microsoft

Abstract

We are providing a preview of a project that analyzes two field experiments with 1,974 software
developers at Microsoft and Accenture to evaluate the productivity impact of Generative AI. As
part of our study, a random subset of developers was given access to GitHub Copilot, an
AI-based coding assistant that intelligently suggests “completions” for code. Our preliminary
results provide suggestive evidence that these developers became more productive, completing
12.92% to 21.83% more pull requests per week at Microsoft and 7.51% to 8.69% at Accenture
(depending on specification). Due to low compliance in the Microsoft experiment and internal
organizational changes at Accenture, our estimates are not very precise and only reach
statistical significance if we weight more heavily periods in which Copilot uptake differs across
control and treatment.

Decision-Making with Machine Prediction: Evidence from Predictive Maintenance in Trucking

Adam Harris
,
Massachusetts Institute of Technology
Maggie Yellen
,
Massachusetts Institute of Technology

Abstract

In this paper, we study how professional human decision-makers use predictive artificial
intelligence (AI). Using a rich decision-level data set from the maintenance of heavy-duty
trucks, we document how the repair decision-making of expert technicians changes with
the introduction of an AI tool designed to predict truck breakdowns. To quantify the effects
of this AI tool on decision-making quality and fleet outcomes, we develop and estimate a dynamic
discrete choice model of technician decision-making. The resulting estimates show
that technicians with the AI tool exhibit a substantially better understanding of breakdown
risk than those without the tool. This improvement in the ability to predict breakdowns
translates into better decision-making and better outcomes: The AI tool reduces the total
costs that technicians incur by $240-$480 per truck per year. We show that this represents
85% of all of the cost savings that could feasibly be achieved through improvements
in decision-making quality; that is, with the AI tool, technicians are close to the efficient
frontier. The AI tool enables these cost savings by helping technicians avoid doing costly,
unnecessary repairs.

Information and the welfare benefits from differentiated products

Imke Reimers
,
Cornell University
Christoph Riedl
,
Northeastern University
Joel Waldfogel
,
University of Minnesota

Abstract

Differentiated product consumption choices made without full information can lead to regret and missed opportunities, but a lack of post-purchase usage data has prevented its exploration. Using novel data on individual ownership and post-purchase usage of video games, we explore both the potential welfare benefits of full pre-purchase information and the ability of contemporary prediction technology to allow the exploitation of the gains. We find large potential gains: Fully informed consumers could achieve 90 percent of their baseline playtime with 40 percent of current expenditure of owned games; and current expenditure reallocated among all available games could double baseline playtime. We develop a tractable model of consumer choice among bundles based on hours of playtime relative to overall spending, which we implement using both a Cobb Douglas calibration and a logit model of bundle choice. Full information would raise CS by more than the value of initial spending while reducing spending by one half. Consumers heeding sophisticated predictions would obtain roughly 40 percent of these welfare benefits with a fifth less spending.

The Market Effects of Algorithms

Lindsey Raymond
,
Massachusetts Institute of Technology

Abstract

While there is excitement about the potential of algorithms to optimize individual decision-making, changes in individual behavior will, almost inevitably, impact markets. Yet little is known about these effects. In this paper, I study how the availability of algorithmic prediction changes entry, allocation, and prices in the US residential real estate market, a key driver of household wealth. I identify a market-level natural experiment that generates variation in the cost of using algorithms to value houses: digitization, the transition from physical to digital housing records. I show that digitization leads to entry by investors using algorithms, but does not push out investors using human judgment. Instead, human investors shift towards houses that are difficult to predict algorithmically. Algorithmic investors predominantly purchase minority-owned homes, a segment of the market where humans may be biased. Digitization increases the average sale price of minority-owned homes by 5% or $5,000 and nearly eliminates racial disparities in home prices. Algorithmic investors, via competition, affect the prices paid by humans for minority homes, which drives most of the reduction in racial disparities. This decrease in racial inequality underscores the potential of algorithms to mitigate human biases at the market level.̊

Discussant(s)
J. Frank Li
,
University of British Columbia
Daniel Martin
,
University of California-Santa Barbara
John Horton
,
Massachusetts Institute of Technology
Robert Seamans
,
New York University-Stern
JEL Classifications
  • L1 - Market Structure, Firm Strategy, and Market Performance
  • L8 - Industry Studies: Services