Contrary to popular belief, AI coding assistants may not always boost productivity for experienced software developers. A new study by the nonprofit METR (Model Evaluation and Testing for Responsible AI) found that using AI tools like Cursor slowed down seasoned developers when working in familiar codebases.
AI Coding Tools Increase Task Times by 19% for Experienced Developers
The METR study, conducted earlier this year, observed a group of experienced open-source developers using Cursor to complete tasks in projects they already knew well.
Before starting, these developers expected AI to help them work faster, estimating it would reduce task times by 24%. Even after completing the tasks with AI, they still believed it had reduced their workload by 20%.
However, the study’s actual results showed the opposite effect: AI increased task completion time by 19%.
Why AI Slowed Down Seasoned Software Engineers
According to the study’s lead authors, Joel Becker and Nate Rush, the findings were surprising. Rush initially expected “a 2x speed up.”
The slowdown occurred because developers had to spend extra time reviewing and correcting AI-generated code suggestions.
“The AIs made some suggestions about their work, and the suggestions were often directionally correct, but not exactly what’s needed,” Becker explained.
While AI assistants provide quick suggestions, those suggestions often don’t fully align with the complex requirements of large, established codebases, especially when used by developers who already know those systems well.
How This Differs From Other AI Productivity Studies
Previous research on AI coding assistants found productivity gains such as:
- 56% faster coding speeds in some controlled studies
- 26% more tasks completed within a given period
However, the METR study highlights that these benefits don’t apply universally, particularly in real-world development scenarios with experienced engineers and complex projects.
Many prior studies rely on simplified benchmarks, which the METR team believes don’t reflect real-world software development challenges.
Will AI Replace Junior Developers Instead?
While AI might slow down experienced coders, it’s still widely expected to affect entry-level coding jobs.
Dario Amodei, CEO of Anthropic, recently predicted AI could eliminate half of all entry-level white-collar jobs within one to five years.
The METR authors noted that the slowdown effect might not apply to junior engineers or those working in unfamiliar codebases, where AI tools could still offer meaningful speed boosts.
Why Developers Still Use AI Coding Assistants Like Cursor
Interestingly, despite the slowdown, most study participants and the authors themselves continue using Cursor.
Becker suggests that while AI may not always speed up task completion, it makes the development experience easier and more pleasant, similar to editing an essay draft rather than starting from scratch.
“Developers have goals other than completing the task as soon as possible,” Becker said. “So they’re going with this less effortful route.”
AI in Software Development Isn’t One-Size-Fits-All
The METR study challenges the growing assumption that AI automatically makes software developers more productive. While junior coders or those tackling unfamiliar projects may benefit, experienced engineers working in well-known codebases might slow down due to AI suggestion corrections.
