Software program engineer workflows have been remodeled in recent times by an inflow of AI coding instruments like Cursor and GitHub Copilot, which promise to boost productiveness by routinely writing strains of code, fixing bugs, and testing modifications. The instruments are powered by AI fashions from OpenAI, Google DeepMind, Anthropic, and xAI which have quickly elevated their efficiency on a variety of software program engineering assessments in recent times.
Nevertheless, a brand new examine printed Thursday by the non-profit AI analysis group METR calls into query the extent to which as we speak’s AI coding instruments improve productiveness for knowledgeable builders.
METR performed a randomized managed trial for this examine by recruiting 16 skilled open-source builders and having them full 246 actual duties on giant code repositories they frequently contribute to. The researchers randomly assigned roughly half of these duties as “AI-allowed,” giving builders permission to make use of state-of-the-art AI coding instruments corresponding to Cursor Professional, whereas the opposite half of duties forbade using AI instruments.
Earlier than finishing their assigned duties, the builders forecasted that utilizing AI coding instruments would scale back their completion time by 24%. That wasn’t the case.
“Surprisingly, we discover that permitting AI truly will increase completion time by 19%— builders are slower when utilizing AI tooling,” the researchers stated.
Notably, solely 56% of the builders within the examine had expertise utilizing Cursor, the principle AI device provided within the examine. Whereas practically all of the builders (94%) had expertise utilizing some web-based LLMs of their coding workflows, this examine was the primary time some used Cursor particularly. The researchers notice that builders had been skilled on utilizing Cursor in preparation for the examine.
However, METR’s findings elevate questions concerning the supposed common productiveness good points promised by AI coding instruments in 2025. Primarily based on the examine, builders shouldn’t assume that AI coding instruments — particularly what’s come to be referred to as “vibe coders” — will instantly pace up their workflows.
METR researchers level to some potential the reason why AI slowed down builders fairly than dashing them up.
First, builders spend way more time prompting AI and ready for it to reply when utilizing vibe coders fairly than truly coding. AI additionally tends to wrestle in giant, advanced code bases, which this take a look at used.
The examine’s authors are cautious not to attract any robust conclusions from these findings, explicitly noting they don’t consider AI techniques presently fail to hurry up many or most software program builders. Different giant scale research have proven that AI coding instruments do pace up software program engineer workflows.
The authors additionally notice that AI progress has been substantial in recent times, and that they wouldn’t anticipate the identical outcomes even three months from now. METR has additionally discovered that AI coding instruments have considerably improved their capacity to finish advanced, long-horizon duties in recent times.
Nevertheless, the analysis gives but another excuse to be skeptical of the promised good points of AI coding instruments. Different research have proven that as we speak’s AI coding instruments can introduce errors, and in some circumstances, safety vulnerabilities.
