The AI coding speedup is smaller than it feels

The most uncomfortable finding in AI coding isn't that the tools don't work. It's that they make you feel faster than you are.

In a now-famous METR study, experienced open-source developers worked on tasks in their own repositories, with and without AI. They predicted AI would speed them up by 24%. Afterward they believed it had sped them up by about 20%. The stopwatch said the opposite: on those tasks they were measurably slower with AI. Slower — while certain they were faster. That gap between the feeling and the measurement is the whole subject of this post.

Why the feeling and the fact come apart

Watch yourself work with AI for an hour and you'll see it. The model produces a wall of plausible code in seconds. That feels like enormous progress — you went from blank file to something that runs in no time. What the feeling hides is the rest of the loop: reading the generated code, catching the part that's subtly wrong, re-prompting, reconciling it with the bit of the system only you understand, fixing the integration the model couldn't see.

The fast part is loud and the slow part is quiet. Generation is visible and instant; review and correction are invisible and grind. So your sense of speed anchors on the burst of output and discounts the cleanup — and on a complex task in code you know well, the cleanup is where the time actually goes. You're not lying to yourself. You're just measuring the wrong half of the work.

It's not "AI coding is fake" — it's "measure it"

Here's the honest other side. METR revised that study in early 2026: the original sample skewed toward developers who benefit least, and a cleaner cohort landed at roughly a 4% slowdown — basically flat, with the range crossing into positive. Their updated read was that AI "likely provides productivity benefits in early 2026." So the lesson is not that these tools are useless.

The lesson is that the real number is modest and conditional, while the felt number is large and constant — and the two keep diverging. Across the industry the same shape shows up: ~93% of developers use AI tools, and aggregate productivity has barely moved. Near-universal adoption, faint signal. That's not a contradiction; it's exactly what you'd expect when everyone feels a 20% boost that nets out to a few points.

Where the gain is real, and where it's a tax

The studies point at a usable rule. AI is genuinely fast on the work you don't already hold in your head: unfamiliar territory, boilerplate, the first draft of a thing, the language you half-know. There the model carries you, and the feeling and the fact agree.

The slowdown shows up on the opposite work: complex changes in a codebase you know cold. There the bottleneck was never typing — it was understanding, and you already had the understanding. The model adds a round trip. It generates, you verify, you correct, when you'd have been faster writing the twenty lines you could already see. Knowing which kind of task you're on is most of the skill now. Reach for the model where you're slow; be suspicious of it where you're already fast.

The bottom line

The danger isn't that AI coding doesn't help. It's that the feeling of help is detached from the amount of help, and the feeling is the one that's loud. Manage by vibe and you'll push AI into the exact tasks where it taxes you, because it always feels like a win in the moment.

So measure something. Cycle time, review load, defects, how long a task actually took versus how long it felt — anything real, even rough. The whole problem is that AI feels like a speedup even when it isn't, so the only way to know is to look at a number instead of a feeling. The tools are good. Your sense of how good is the part that lies.