Do you know what is slowing down senior coders? As it turns out, it's AI

Picture this: the very thing that was created to help you is now your performance bottleneck. Quite the irony.

There's a new study that wreaks havoc to the idea that artificial intelligence supercharges the work of seasoned software developers. Instead of speeding them up, using AI tools actually slowed down experienced coders when they tackled projects they already knew well.



The study is carried out by AI research nonprofit METR and is focused on veteran developers working with Cursor, a popular AI coding assistant, on open-source projects they were familiar with. Before diving into the study, the developers expected the AI would save them time, guessing it could cut task completion by nearly a quarter. Even after using the AI, many still felt it had made them roughly 20% faster. But the hard data told a different story: AI extended the time needed to finish tasks by 19%.

Joel Becker and Nate Rush, who led the research, admitted they were caught off guard by the outcome. Rush, before the study began, had predicted the AI would double productivity – and that's what many of us would think. The findings, however, cast doubt on the widespread assumption that AI tools reliably boost the productivity of highly skilled, high-cost software engineers. This idea is something that many companies have chosen to invest heavily into and big investments have been made recently.

This comes as some tech leaders, including Dario Amodei, CEO of AI company Anthropic, have suggested that AI could eliminate as much as half of all entry-level white-collar jobs within the next five years. Previous studies have added to the hype, reporting significant productivity gains: one claimed AI sped up coding by 56%, while another found developers using AI completed 26% more tasks within the same timeframe.

But METR's study tells a very different story. The boost in productivity doesn't hold up when developers are working on large, complex codebases they know well. In these cases, AI not only failed to help but actively slowed developers down. The problem stemmed from the need to double-check and often correct the AI's suggestions – suggestions that were frequently close to correct, but not precise enough to be trusted without careful review.

Becker explained that video recordings of the participants showed how AI often nudged developers in the right direction, but rarely delivered exactly what was needed:



That led to additional time spent reviewing, editing, and sometimes discarding the AI's output.

The researchers were careful to point out that these results likely don't apply across the board. Less experienced engineers or those working on unfamiliar codebases could still benefit from AI assistance.

What about smartphone AI?

The study's findings may also offer a glimpse into how everyday smartphone users (like you and me) interact with AI-powered features on their devices. Just like developers, many smartphone users expect AI tools – from predictive text and voice assistants to photo editing suggestions – to streamline the day-to-day tasks.

But these features often require us to pause, review, and correct AI missteps, sometimes making simple actions feel more complicated than they should be. Whether it's an autocorrect blunder, a poorly framed photo enhancement, or a confusing AI-generated message reply, we as smartphone users are also discovering that AI is far from flawless.

Convenience can sometimes come at the cost of time!

Grab Surfshark VPN now at more than 50% off and with 3 extra months for free!

Secure your connection now at a bargain price!

We may earn a commission if you make a purchase

Check Out The Offer