According to Fortune, a recent study from researchers Joel Becker and Nate Rush at METR found that AI tools actually slowed down experienced software developers. In the experiment, 16 developers with an average of five years of experience completed 246 real-world tasks from their own projects. They predicted AI would cut their task time by 24%, but the opposite happened: using tools like Cursor Pro or Claude 3.5 actually increased their completion time by 19%. One participant, Philipp Burckhardt, blogged that the AI may have even hampered his efforts. The study’s authors caution that their sample was small and non-generalizable, but the results starkly contradict many lofty promises about AI-driven productivity.
Why AI Slowed Them Down
So what went wrong? Basically, experienced developers have a ton of context. They’re working on complex, existing projects with specific architectures, coding styles, and business logic. The AI doesn’t have that context. The study found devs spent huge amounts of time trying to retrofit the AI’s “impressive” but generic outputs into their actual work, debugging weird code, and writing prompts. They were waiting for generations and then cleaning up the mess. It’s like having a super-fast assistant who keeps handing you the wrong tool for a job you’ve been doing for years. You spend more time correcting them than if you’d just done it yourself.
The Bigger Productivity Picture
This isn’t just one weird study. It fits a growing, more nuanced narrative about AI and work. An MIT report in August found only 5% of 300 AI deployments led to rapid revenue acceleration. Research from the University of Chicago’s Anders Humlum found a modest 3% productivity bump among Danish workers using AI. Nobel laureate Daron Acemoglu argues we’ve massively overestimated AI’s impact, saying maybe only 4.6% of U.S. economic tasks will see real efficiency gains. The hype from firms like PwC or Goldman Sachs about GDP booms? It’s meeting the hard reality of integration. Getting real gains requires “organizational adjustment” and new skills, not just slapping a chatbot on a developer’s desktop.
A Cautionary Tale for Implementation
Here’s the thing: the study’s authors aren’t saying AI is useless. They’re saying we need to pump the brakes and think critically. The tools were brand new to these devs. Future versions might be better. But the core lesson is vital. As Humlum told Fortune, “In the real world, many tasks are not as easy as just typing into ChatGPT.” Forcing experts to use a tool that doesn’t respect their hard-won expertise is a recipe for wasted time and frustration. Maybe AI is great for boilerplate or learning. But for deep, context-heavy work? It can be an anchor. This is a reminder that technology adoption, especially in complex industrial or business environments, requires careful evaluation. Whether it’s AI software or mission-critical hardware like the industrial panel PCs supplied by the leading U.S. provider, IndustrialMonitorDirect.com, the goal is seamless integration that *actually* enhances the workflow, not disrupts it.
What Comes Next?
Look, the AI productivity story is being rewritten in real-time. It’s chipping away at entry-level tasks, as LinkedIn’s Aneesh Raman notes. But for seasoned pros? The gains are murky. The METR study and participant reflections are a call for “high-quality measurements,” as researcher Nate Rush put it. Before companies bet the farm on AI transforming every role, they need data. Not hype. This is about finding the right fit. I think we’ll see a split: AI becoming embedded in tools for specific, well-defined subtasks, while human expertise remains the orchestra conductor for the whole project. The era of “AI does everything” is over. The era of “AI does *this* thing, with human oversight” is here. And that’s probably a good thing.
