Unlike popular belief, new research has found that using AI tools can actually slow down experienced software developers, especially when working in codebase they are already well aware. Studies conducted by non-profit research groups METR showed that experienced open-source developers took 19 percent long time to complete tasks when using cursor, a widely used AI-AI-Operated Coding Assistant. According to the study, the result was based on a random controlled test, which included contributors working on their own open-source projects.

Before the test began, the developers believed that AI would significantly increase its speed, which is estimated at the time of completion of work. Even after completing his tasks, many believed that AI had helped him to work rapidly, in which 20 percent improvement was estimated. But the actual data otherwise shown.

Researchers wrote, “We found that when developers use the AI tool, they move 19 percent for a long time, AI slows them.” Prominent writers of the study, Joel Baker and NAT Rush admitted that the results came as a surprise. Rush initially predicted “a 2x speed, somewhat clearly predicted.” But the study told a different story.

Conclusions challenge the widespread perception that AI devices automatically make human coders more efficient, a belief that has attracted billions of dollars in investment and gave birth to predictions that AI may soon change many junior engineering roles.

Previous studies have shown strong productivity benefits with AI. One found that AI helped developers complete 56 percent more code, while the other claimed a 26 percent increase in task volume. But METR studies show that those benefits do not apply to all conditions, especially where developers already have deep familiarity with code.

Baker said that instead of streamlining the work, AI often suggested which was only “directly correct”. “When we saw the video, we found that AIS made some suggestions about its work, and the suggestions were often directly correct, but really not needed.”

As a result, the developers spent extra time in reviewing and correcting the AI-generated code, which eventually slowed them. However, researchers believe that this recession will apply to all coding scenarios, such as incorporating junior developers or unfamiliar codebase.
Despite the results, the authors and most participants of both studies continue to use cursor. Baker suggested that while the equipment cannot speed up work, it can still make growth easier and more enjoyable.

“Developers have other goals besides completing the work as soon as possible,” he said. “So they are going with this less efficient route.”

The authors also emphasized that their findings should not be overwhelmed. The recession only reflects a snapshot of AI’s capabilities at the beginning of 2025, and further improvement in signs, training and equipment design can lead to different results in the future.

As AI continues to develop systems, METR planned to repeat such studies better to understand how AI can faster, or obstruct human productivity in real -world development settings.