Study Finds AI Tools May Slow Down Experienced Developers’ Task Completion Time

A recent study conducted by AI research nonprofit METR challenges the widely held belief that artificial intelligence tools always boost productivity among experienced software developers. Instead of accelerating their workflow, the study found that developers familiar with their codebases actually experienced a 19% slowdown when using Cursor, a popular AI-powered coding assistant. This result contradicts the developers' initial expectations — before the experiment, they had predicted a 24% reduction in task time, and even after completing their tasks, they still believed they had worked 20% faster with AI.

The study focused on seasoned developers working on open-source projects they knew well. While they anticipated AI to enhance their productivity, they were instead slowed down by the need to review, edit, or correct the AI-generated suggestions. The model’s suggestions were often directionally correct but lacked precision, requiring developers to spend extra time refining the results. According to lead authors Joel Becker and Nate Rush, these results were surprising. Rush had even predicted a twofold speed increase before the study began.

These findings carry broader implications for the tech industry, especially as major investments pour into AI tools with the promise of making human developers significantly more efficient. Many companies are betting on AI’s potential to revolutionize software development workflows and are even suggesting that AI could replace a large portion of entry-level programming jobs. For instance, Dario Amodei, CEO of Anthropic, recently projected that AI could eliminate half of all entry-level white-collar jobs in the next one to five years.

While earlier research has shown dramatic improvements — one study suggested AI could boost coding speed by 56%, and another reported a 26% increase in completed tasks — METR’s findings indicate these gains are not universally applicable. Specifically, experienced developers who deeply understand the quirks, dependencies, and architecture of large, mature codebases may not benefit from AI in the same way. In such scenarios, AI may offer limited assistance or even become a hindrance by introducing noise that the developers need to clean up.

The authors also pointed out a significant limitation in most existing AI productivity studies: many rely on simplified or synthetic software development benchmarks that don’t accurately reflect real-world challenges and complexities. METR’s study aimed to address that gap by analyzing genuine developer behavior in authentic development environments.

Interestingly, despite the observed slowdown, the majority of the study participants continue to use Cursor, as do the authors themselves. This apparent contradiction is explained by the subjective benefits AI offers. Developers reported that the AI-assisted experience felt smoother and more enjoyable. The comparison was made to editing an essay rather than writing from scratch — even if it takes a bit longer, the process feels less mentally taxing.

As Becker summarized, developers often pursue goals beyond sheer speed. They value comfort, ease of use, and cognitive support, even if it doesn’t translate into faster task completion. In conclusion, the study offers a nuanced view of AI’s impact on software development — it’s not a universal accelerator but a tool whose effectiveness depends heavily on the context and the user’s familiarity with the codebase.