Artificial Intelligence

Study: AI Tools Slow Down Experienced Developers by 19%

Published JUL 11 2025

Story

Above: Ian Collins, CEO of software developer Wysdom.AI, in his home office in Markham, Canada on June 13, 2023. Image credit: Lance McMillan/Toronto Star via Getty Images

The Facts

AI research nonprofit METR conducted a randomized controlled trial with 16 experienced developers from major open-source projects, finding that using early-2025 AI tools like Cursor increased task completion time by 19% rather than speeding up work.
METR
Reuters
Finimize
Before the study, developers predicted that AI would decrease their task completion time by 24%. Even after experiencing the slowdown, they still believed AI had sped them up by 20%, revealing a significant gap in perception.
Cryptopolitan
Reuters
The study involved 246 coding tasks, randomly assigned to either allow or disallow AI use, with developers working on familiar open-source codebases that averaged over 1 million lines of code and multiple years of development history.
METR
The Economic Times
Developers spent substantial time reviewing and correcting AI-generated code, with only 39% of code generations from Cursor being accepted, often requiring multiple rounds of prompting before manual coding was needed.
METR
Dev
Secondthoughts
The slowdown primarily affected experienced developers familiar with large, established codebases that have strict quality standards. The authors noted that the results may not apply to junior engineers or unfamiliar projects.
StartupNews.fyi
Reuters
Despite the productivity decrease, most study participants and authors reported that they continue to use AI tools because they make the development experience more pleasant and less effortful, similar to editing an essay versus starting from scratch on a blank page.
Reuters
The Economic Times

Sources Split

LEFT

CENTER

RIGHT

ESTAB
CRIT

CENTER

PRO
ESTAB

The Spin

Techno-skeptic narrative

This study exposes a critical flaw in the AI hype surrounding software development. Experienced developers who know their codebases intimately are being slowed down by AI tools that generate mediocre code requiring constant review and correction. The 19% productivity loss suggests that AI isn't the silver bullet promised by tech companies selling these expensive tools to unsuspecting businesses.

CodingIT

Ycombinator

Techno-optimist narrative

The study reveals important limitations but doesn't invalidate AI's potential in software development. These tools are still evolving rapidly, and the learning curve for effective AI-assisted coding is steeper than expected. The future belongs to developers who can effectively combine AI assistance with human expertise, not those who reject these powerful tools entirely.

Secondthoughts