This post is also available in:
עברית (Hebrew)
A new study challenges the common assumption that AI coding tools consistently boost productivity. In controlled tests, seasoned developers working within codebases they already knew well were actually slower when using an AI assistant.
The research, conducted by nonprofit group METR, observed experienced open-source contributors performing real tasks with the help of Cursor, a popular AI tool for software development. Developers were asked to complete tasks in open-source projects they were already familiar with—code they had worked on before, not unfamiliar code or toy problems.
Despite anticipating a significant performance boost—developers expected AI to reduce task time by nearly a quarter—the results showed the opposite. On average, the AI tool led to a 19% increase in completion time.
According to Reuters, the core issue wasn’t the accuracy of the AI’s suggestions, but the added overhead in reviewing, interpreting, and correcting its output. While many suggestions were technically relevant, they often didn’t align with the specific needs or structure of the project at hand. This meant that developers had to spend time adjusting the AI-generated code, adding friction rather than saving time.
This contrasts with earlier studies that reported large productivity gains using AI tools. However, those studies often relied on artificial tasks or focused on junior developers, rather than real-world coding scenarios with experienced professionals.
Interestingly, even though the tool made them slower, most participants continued to use it after the study. The reason, researchers suggest, may be psychological rather than performance-driven: AI can reduce the cognitive load of getting started or thinking through structure.
The findings underline a growing understanding in the developer community: generative AI tools can be helpful, but their effectiveness depends heavily on the context. For familiar environments and seasoned engineers, speed might not be the biggest benefit—ease of workflow and reduced mental effort could be just as valuable.
