Artificial Intelligence Efficiency and Results in Software Development
Research from MIT Sloan and other institutions shows that while AI increases the coding efficiency of software developers, final product outputs do not increase at the same rate due to human bottlenecks.
New academic research has revealed that AI tools significantly increase software developers' efficiency during the code-writing stage, but this increase is not fully reflected in final outcomes due to human-induced bottlenecks later in the production cycle.
Artificial Intelligence in Software Development
While organizations wanted to see whether AI could boost worker productivity, studies on accountants and manufacturing firms had shown that this was possible. New research reveals that AI does the same for software developers.
An associate professor of applied economics at the MIT Sloan School of Management noted that tasks that used to take a whole day can now be done in minutes. However, this productivity gain does not always translate into final results.
Coding Activity and Project Increase
A new paper by Demirer, University of Pennsylvania assistant professor Leon Musolff, and Liyuan Yang concluded that AI tools increase efficiency while writing code, but this effect diminishes in shipping the code.
The researchers found that developers using AI tools could increase their coding activity with a cumulative impact of up to 180 percent, but these gains diminished throughout the production cycle.
App Reviews and Downloads
The AI-enhanced coding study led to only 50 percent more projects and 30 percent more actual releases compared to developers not using AI.
Additionally, while the adoption of AI tools led to an increase in releases to popular app stores, it did not lead to a corresponding increase in app downloads or user reviews.
The Role of Human Bottlenecks
The researchers stated that human processes remaining unchanged are responsible for the disconnect between AI productivity gains and actual outcomes.
Demirer stated that some bottlenecks are actual technical and mechanical tasks that AI is not good at doing and that humans must step in to complete.
GitHub Data and Tool Categories
Examining data from over 100,000 developers on GitHub, the researchers evaluated how three categories of AI tools affected productivity at each stage of software development.
Using different AI tools increased developers' efficiency in writing code; autocomplete activity increased by 40 percent, synchronous tool activity by 140 percent, and asynchronous tool activity by 180 percent.
Recommendations for the Future
The productivity gain provided by AI in the early stages diminished due to human bottlenecks in later stages, such as reviewing changes and launching applications.
To fix these bottlenecks, researchers recommend smaller teams, reallocating resources to subsequent tasks, upskilling workers, and using AI to streamline human workflows.