Does AI Really Make Developers Faster?

It has become an article of faith that AI coding tools make software teams dramatically faster. The tools feel quick, the demos are slick, and nearly every developer now uses one. So it is worth sitting with an uncomfortable finding: when researchers actually measured it, the speed-up was hard to find, and in one careful study it went the other way.

What the evidence shows

The most cited result comes from METR, a research group that ran a randomised trial with experienced open-source developers on their own codebases. The developers expected AI to make them about 20 percent faster. In fact they were roughly 19 percent slower when using it, even though they still believed afterwards that it had sped them up. The gap between how fast the tools felt and how fast they were is the striking part.

That is one study, on a particular kind of expert work, and it is not the whole picture. METR itself notes the tools and the people using them have improved since, and other research lands anywhere from a modest slowdown to a real gain depending on the language, the codebase and the task. A fair summary is not "AI makes you slower." It is that the honest number is a lot smaller and messier than the marketing suggests.

Where the time goes

The reason is not that the models write bad code. It is that writing code was never the whole job. An AI suggestion has to be read, checked, corrected and fitted into a system that already exists, and on a large or unfamiliar codebase that review can cost more time than it saves. A fresh data point this week captured the tension: one report found AI now drafts close to half of all issues filed in Linear, a popular project tracker, and yet teams were not obviously shipping faster for it. More output is not the same as more progress.

It also fits what we keep seeing when agents take on real work rather than toy demos. They are impressive in a sandbox and stall on the multi-step, messy tasks that make up most of a working day.

So is it a myth?

The claim that AI tools help programmers is not a myth. Used well, on the right task, they clearly can. The myth is the size of the effect: the idea of a uniform, enormous productivity jump that shows up on every team the moment they switch the tools on. The measured reality is smaller, uneven, and easy to overestimate precisely because the tools feel so fast while you use them.

The practical lesson is not to turn AI off. It is to be suspicious of your own sense of speed, and to measure outcomes that matter, like work actually shipped, rather than trusting the feeling that you are flying. As the tools genuinely improve, the honest numbers may catch up with the hype. They have not fully done so yet.