The Number That Held Still

The number everyone watches is whether people feel more productive. That number is easy to collect, easy to put on a slide, and, according to a new longitudinal study of professional engineers, easy to hold steady while something underneath it gets worse. Over six months with AI coding assistants, productivity perception did not budge. The share of engineers reporting that the developer experience itself had degraded nearly doubled, from 14 percent to 27 percent, in the same cohort, over the same window.

The study

Annie Vella and Kelly Blincoe ran a mixed-methods longitudinal study, which is a plain way of saying they asked the same people the same things twice and watched what moved. Two questionnaires went out six months apart. The first drew 158 eligible professional software engineers, the second 101, with a matched cohort of 95 who answered both and can therefore be tracked as individuals rather than as two separate crowds. That matching is what makes the finding more than a vibe: it is the same people, changing.

What moved first was where the time goes. Participants reported spending less time on most development tasks, and 82 percent reported spending less time specifically on writing code. The authors describe a broader shift in focus from creation to verification. You are not typing the thing into existence as much anymore. You are reading what the assistant produced, judging whether it is right, and correcting it when it is not.

A new category of work

Vella and Blincoe give that shift a name: supervisory engineering work, the direction, evaluation, and correction of AI output. It is a useful piece of vocabulary because it separates two things that get lumped together. Writing less code is not the same as doing less work. The work is still there; it moved. It moved from the part of the job that produces flow, the heads-down build, to the part that produces vigilance, the constant read-and-check of output you did not write and cannot fully trust.

That relocation is the mechanism behind the headline. When your day becomes supervision, the texture of the day changes even if the throughput does not. You spend it evaluating instead of making. And evaluation, it turns out, is where the experience quietly erodes.

The paradox

Here is the split the study calls a productivity-experience paradox. On one side, perception held: 84 percent reported a productivity improvement at both time points, a number stable enough that any dashboard tracking it would show a flat, healthy line for half a year. On the other side, among the matched participants, the proportion reporting a worsened developer experience in at least one dimension nearly doubled, from 14 percent to 27 percent. The two dimensions that eroded were flow state and cognitive load. The one that improved was feedback loops, which makes sense: the assistant answers fast. Fast feedback and broken flow can coexist, and here they did.

So the same engineer, on the same survey, says yes, I am more productive, and also says the work feels worse than it did. Both are true reports. They just measure different things. Productivity perception is an answer about output. Flow and cognitive load are answers about what it costs you to produce it. Ask only the first question and you will hear that everything is fine, for as long as you care to keep asking.

The wrong meter

This is the companion finding to the token studies we keep covering, arriving from the other direction. Those measured a cost that used to be invisible because nobody metered it, and found the meter changes the argument. This one measures a cost that stays invisible precisely because the popular meter, self-reported productivity, holds still while it happens. One is a number nobody watched. The other is the number everybody watches, doing its job as reassurance while the thing worth worrying about moves somewhere the number cannot see.

Keep the scale honest and the point survives. This is one cohort of 95 matched engineers reporting on themselves over six months, not a law of nature, and a jump from 14 to 27 percent still leaves most people unmoved. But it is the shape that should give a manager pause. If you adopted these tools and your only instrument is a productivity-sentiment number, the study is telling you your instrument will read normal in exactly the situation you would most want it to flag. The work got quieter to describe and heavier to do, and the flat line never twitched.