AI makes experienced devs 19% slower. They just don't feel it.

Senior devs used AI coding tools and got 19% slower. But they felt 20% faster. Read that again.

No, you read that correctly. A 39-percentage-point difference between what people think about something and what the truth is. And it might be the most important finding about AI-assisted development this year.

The Study

Experienced open-source developers who worked on real tasks were measured by researchers with and without AI coding assistants. They weren't juniors going through tutorials. These were seasoned contributors familiar with their codebases.

The outcome? These AI tools actually made them 19% slower. Those same developers self-reported feeling roughly 20% faster.

The Productivity Placebo

I believe I understand the situation here. And if you've used Copilot or Cursor for more than a week, you probably recognize it too.

The typing feels faster. Autocomplete fills in boilerplate. Tab-tab-tab and you've got a function skeleton. That dopamine hit of "wow, I didn't have to type that" is real. Your fingers are doing less work. Your brain registers that as speed. 🧠

But here's what your brain doesn't track well:

→ The 4 minutes you spent reading generated code to check if it actually does what you wanted
→ The 12 minutes debugging a subtle hallucination buried in otherwise-correct output
→ The context switches when AI suggestions pull you down a different path than you intended
→ The false confidence that lets a bug slip into code review

While you save time on generation, you end up spending that time on verifying and debugging the code. So, in the end, there is no real advantage. But it doesn't feel negative because the slow parts are spread out and invisible.

Why This Should Worry Us

This does not oppose the use of AI tools. It's an argument against blind trust in them.

If you feel faster but you're actually slower, you won't course-correct. You'll keep doing the thing that feels productive. This is what we call a productivity placebo.

The situation is even more alarming at the organizational level where managers observe that developers are creating more code, more commits, more PRs, and all the vibes-based metrics are trending positively, yet if actual throughput is down 19% nobody realizes it until deadlines start slipping and the bug count keeps growing. 😬

What I Think This Actually Means

I don't think AI coding tools are useless. I think we're using them wrong.

The experienced developers involved in this research were working on codebases that they were already familiar with - i.e., they already had the mental models and knew what to type; for them, AI assistance was an overhead, not a leverage.

I think the calculus changes in the unknown unknowns. Trying out a new API. Writing in a language you touch once per year. Filling in a project structure you've never made before. That's where generation speed might actually outweigh verification cost.

The lesson isn't "stop using AI tools." The lesson is stop assuming they're helping just because they feel good.

The Uncomfortable Takeaway

Track your time and effort spent on debugging. Time yourself while debugging code before and after testing. Not "do I feel productive" but "did this task take less time." Use a timer if you have to. Be honest with yourself about the debugging tax.

The top developers in my circle consider AI assistants as junior pair programmers. They are good for helping with ideas and providing boilerplate code, but you still need to carefully check and review every line of code they produce as if the assistant is a new hire. The bottom developers treat them like magical oracles. 🎯

If the feeling of speed and the reality of speed are pointing in opposite directions, the feeling is lying to you.

What's your experience — have you ever actually timed yourself with and without AI tools on the same kind of task?