Generative AI changes how employees spend their time

As artificial intelligence models become more sophisticated, researchers and employers are exploring its potential to have a positive impact on workplace productivity. When relatively routine activities take less time, AI users can ostensibly devote more time to more interesting tasks.

A recent paper co-authored by Frank Nagle, a research scientist at the MIT Initiative on the Digital Economy, shifts the focus from productivity gains to how AI changes the nature of the work itself. When they analyzed the behavior of software developers with access to generative AI tools, the researchers found that those workers did more core coding work and fewer non-coding tasks.

“Generative AI gave people the ability to do more of what they want to do and less of what they have to,” said Nagle.

Observing how developers work

To study generative AI’s impact on work tasks, the researchers looked at the open-source development platform GitHub. Software developers who use the platform have “observable work patterns,” as the researchers put it. These include generating code, merging code snippets and, for those who manage projects on the platform, reviewing code contributions and looking over support requests.

In June 2022, the platform launched GitHub Copilot, a generative AI code completion tool for software developers. Instead of completing written text like a more traditional large language model, the tool generates the next snippet of code. Copilot aims to help programmers “code faster, solve problems more quickly, and learn code that they previously did not know,” the researchers write.

As part of GitHub’s efforts to support open-source developers, it gave free access to Copilot to those working on highly ranked projects. Nagle and his co-authors looked at the behavior of 187,000 developers on GitHub (some with free access, some without) in the year before Copilot’s launch and then again in the two years afterward. In doing so, they sought to test several hypotheses about how access to a generative AI tool would affect the day-to-day work of developers who need to balance programming and project management activities.

AI access leads to long-term changes in work patterns 

At the onset of the study, the developers spent roughly 44% of their time on coding activities and 37% on project management. After Copilot was made available, developers increased coding activities by 12.4% and decreased project management activities by 24.9%. At the same time, developers using Copilot dramatically reduced their peer collaborations, by nearly 80%.

Over time, Nagle said, the proportion of time spent coding has leveled off, but it hasn’t returned to the baseline. “This gives some indication that people will change the way they work, and [this change] will stick around in the long term,” Nagle said. 

The findings suggest that Copilot helped developers produce more accurate code. That reduced the need for them to interact with peers to review code and resolve issues, which are considered auxiliary project management tasks. This helped the platform’s “power users” — the top contributors who maintain projects and who often spend considerable time resolving issues — avoid becoming overwhelmed and burned out. 

Less-experienced developers in the sample saw the biggest increase in time devoted to coding activities. This outcome aligned with the results of a 2023 paper that found that call-center workers with less experience gained the most from generative AI compared with their more experienced colleagues. 

It also reinforced a point Nagle made in a recent Fortune article: Enterprises make a “profound strategic error” when they replace junior employees with AI models as a cost-cutting move. In fact, allowing AI to complement the work of junior workers helps them build skills and take on additional responsibilities faster.

“When companies stop hiring entry-level people, it’s short-term thinking at the expense of investing in the future,” Nagle said. 

Leading Technical Professionals and Teams

In person at MIT Sloan

Get information

Consider AI’s impacts on teamwork and learning 

The results of the study brought up two issues worth additional exploration, according to Nagle.

The first is a “retreat away from teamwork” that occurred as developers started using Copilot. While they spent less time conferring (or potentially arguing) with their peers, they also moved away from the collaborative approach typically associated with open-source software.

“That’s the isolating effect of technology,” Nagle said. “If no one has to ask their colleagues for advice, is that good or bad? Yes, you avoid wasting time, but it reduces the human interaction that’s so valuable to what companies do and people do.”

The other point of interest is Copilot’s impact on developers learning new programming languages. Researchers found those with access to the AI tool increased their cumulative exposure to new programming languages by nearly 22% relative to the baseline. 

This shows that Copilot can enable what Nagle called “low-cost experimentation” and shorten the learning curve. At the same time, developers — and their employers — need to be sure that generative AI is being used to help them learn, not to replace their thoughts.

“We still teach young children basic arithmetic even though they’ll eventually have access to a calculator,” he said. “To do more advanced math, you need to learn the basics. If we offload everything to LLMs, then we may forget how to do things.”

Read the paper: “Generative AI and the Nature of Work”

This article is based on a research paper by Manuel Hoffmann, Sam Boysel, Frank Nagle, Sida Peng, and Kevin Xu.  

Frank Nagle is a research scientist at the MIT Initiative on the Digital Economy and the chief economist at the Linux Foundation. He studies how competitors can collaborate on the creation of core technologies while still competing on products and services built on top of them — especially in the context of artificial intelligence. His work frequently explores the domains of open-source software, crowdsourcing, free digital goods, cybersecurity, and the generation of strategic predictions from unstructured big data.

Manuel Hoffmann is an assistant professor at the University of California, Irvine. Sam Boysel is a data scientist at the Linux Foundation. Sida Peng is a senior director of economics in the Office of the Chief Economist at Microsoft. Kevin Xu is a staff software engineer at GitHub.