Software engineering was supposed to be artificial intelligence’s easiest win. Today companies such as OpenAI, Anthropic, Microsoft and Google have all released AI products geared specifically to coding. And a survey of nearly 5,000 technology professionals released in a report last year by Google’s DevOps Research and Assessment (DORA) team found that 90 percent of respondents said they were using AI at work—with more than 80 percent saying the technology had boosted their productivity.
“We see a large majority of folks that are relying on AI to get their job done, at least a moderate amount, which is really fascinating,” says Nathen Harvey, who leads the DORA team.
AI can generate code for everything from Web and mobile apps to data management tools. It often automates some of the tedious elements of the job, such as building testing infrastructure and updating software to run on new devices and systems. In some cases, even inexperienced developers can create working prototypes simply by describing their intentions to AI systems in a process often called “vibe coding,” a term coined by OpenAI co-founder and researcher Andrej Karpathy. But writing code is only part of the job; developers still have to verify that it does what it’s supposed to and fix it if it fails.
On supporting science journalism
If you’re enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.
Another finding from the DORA report was that while individual coder effectiveness appeared to rise with the use of AI, so, too, did “software delivery instability”—an assessment of how frequently code needed to be rolled back or patched after release to address unexpected issues.
“As you use more AI, you’re more likely to roll back changes that you’ve pushed into production,” Harvey says. “And this, obviously, is something that you want to avoid.”
Even as it becomes increasingly adept at writing code, AI doesn’t eliminate the need for human software engineering. Developers often still need to craft bespoke code—or at least tweak an AI tool’s output—to handle unusual cases or specific business needs that might not be reflected in AI training data. They also still need to carefully confirm that machine-generated programs behave exactly as intended and meet company standards.
AI tools don’t automatically shorten the workday. In some workplaces, studies suggest, AI has intensified pressure to move faster than ever.
If employers don’t manage its effects, AI may even exacerbate stress and burnout among software engineers. In a report published in the Harvard Business Review in February, researchers at the University of California, Berkeley’s Haas School of Business found that employees at one U.S. tech company took on more tasks, worked at a faster pace and worked more hours after adopting AI. Even without the company mandating use of the technology, employees began prompting AI during lunch, breaks and meetings, with some finding former downtimes less refreshing. There’s a risk that initial excitement and productivity boosts could give way to fatigue, lower-quality output and greater employee turnover, the researchers warned.
This pressure isn’t happening in a vacuum. Following years of industry-wide layoffs and corporate mandates for efficiency, AI is often deployed alongside the expectation that those left behind will do more with less.
Additionally, a report assessing more than 500 developers, released late last year by Multitudes, a New Zealand–based company that helps businesses track and optimize software engineering practices, found indications that AI can expand worker productivity but also working hours. On average, engineers merged 27.2 percent more “pull requests”—packages of code that were approved for insertion into existing software projects. But they also experienced a 19.6 percent rise in “out-of-hour commits”—submissions of coding work outside of their ordinary schedules. That could be a sign of problems to come.
“If that out-of-hours work is going up, it’s not good for the person,” says Multitudes founder and CEO Lauren Peate. “It can lead to burnout.”
The Multitudes report doesn’t definitively prove that AI directly caused the measured changes, but Peate says interviews suggest that the observed changes in hours among engineers are likely a sign that businesses expect greater productivity from employees in the AI era.
“People were feeling additional pressure to get more work done, and it looks like that was contributing to them putting in more hours,” she says.
While some research has suggested that less experienced developers might be among those who benefit the most from AI’s assistance, and vibe coding can let people with a minimal programming background build programs that run, a recent assessment from Anthropic suggests that overreliance on AI may affect the development of coding skills.
In a report released in January, Anthropic researchers found that software engineers working with a new software library saw a small, statistically insignificant boost in speed when they solved a task with the aid of AI compared with a control group working without AI assistance. When the coders were quizzed about the software library after the task, however, the group given AI assistance scored 17 percent lower than the AI-free group. Those who asked questions of the AI rather than just relying on it to generate code generally performed better, but the researchers raised concerns that using AI to simply complete tasks as quickly as possible under workplace pressure could be harmful to engineers’ professional development.
Additionally, they noted, the biggest gap in quiz performance was in questions related to debugging code—the process of finding and fixing the flaws that make code malfunction. In other words, junior developers who rely too much on AI might have a harder time not only writing code on their own but also understanding and putting the finishing touches on the code they generated in the first place. In a statement to Scientific American, Anthropic researcher Judy Hanwen Shen said the goal “shouldn’t be to use AI to avoid cognitive effort—it should be to use AI to deepen it.”
Already, the U.C. Berkeley researchers noted, engineers can find themselves helping co-workers who’ve created incomplete software solutions through vibe coding. And some open-source projects have reported a rise in low-quality, AI-driven submissions that sap core developers’ time.
That comes after a 2025 Harvard Business School working paper indicated that AI can lead to open-source developers shifting their time from handling project management tasks, such as reviewing code contributions and maintaining lists of issues for contributors to fix, to generating code themselves.
“You can do it by yourself now, so there’s not a lot of need to interact much with others,” says Manuel Hoffmann, a co-author of the paper and an assistant professor of information systems at the University of California, Irvine’s Paul Merage School of Business. “And that’s not necessarily a bad thing.”
Still, such use of AI may limit another channel for less experienced programmers to hone their skills, develop professional networks and expand their résumés.
And as AI redefines what productivity means, workplace structures that prevent burnout, keep workloads manageable, and provide avenues for advancement and training may be more important than ever.
“When you’ve got great things happening, and you add some AI to the mix, they’re probably going to get better,” Harvey says. “And when you have painful things that are happening, [and] you add some AI to the mix, [you’re] probably going to feel that pain a little bit more acutely.”
