Do AI coding tools really make skilled developers faster, or just feel that way?

This explores whether AI coding tools make experienced developers faster when you measure it, as opposed to how fast developers feel they're going. The short answer from the corpus: it depends heavily on who is doing the work and where, and developers' sense of their own speed is a poor guide. The sharpest result comes from a randomized trial of experienced open-source developers working on their own mature codebases. With early-2025 AI tools they took 19% longer to finish tasks, even though they had predicted a 24% speedup beforehand. Outside experts in economics and ML overestimated the gains too Do AI coding tools actually speed up experienced developers?.

The opposite result also exists. A randomized trial of 96 Google engineers found AI features cut time on a complex task by about 21%. The confidence interval was wide, though, and whether the result counted as statistically significant depended on how the model was specified Do AI coding features actually speed up engineer productivity?. A likely way to reconcile the two: the slowdown study picked people who already knew their codebases deeply. For them, the AI's suggestions had to beat expertise they already had, and every suggestion had to be checked. Anthropic's internal survey points the same way. Engineers reported 50% productivity gains and merged 67% more pull requests, yet most said they could fully hand off only 0–20% of their work Does AI assistance erode the skills needed to oversee it?. More output is not the same as less effort per task, and self-reported numbers are exactly the kind the open-source trial showed to be inflated.

The hidden cost is checking the AI's work. Stack Overflow's 2025 survey found AI use rising to 80% of developers while trust in its accuracy fell from 40% to 29%. The top complaint was code that looks right but contains subtle errors Why do developers keep using AI tools they don't trust?. For an expert, reviewing plausible-looking code can take longer than writing it. That explains how someone can feel faster (less typing, quick drafts) while actually being slower.

This is the part you might not have expected to care about: whether AI speeds you up seems to depend less on the tool than on how you use it. A trial of developers learning new libraries found six distinct interaction patterns. The passive ones produced comprehension quiz scores of 24–39%. Patterns where developers worked actively to understand the code scored 65–86% Does AI assistance actually harm the way developers learn?. A separate study found that traits visible in developers' conversations with a coding agent predicted outcomes beyond their prior skill, but those traits weren't stable enough to teach as a skill Can conversation patterns predict coding outcomes better than prior skill?. This creates a long-term tension. Anthropic's engineers worried that handing routine work to Claude wears down the hands-on practice they need to catch Claude's mistakes Does AI assistance erode the skills needed to oversee it?. Leaning on the tool too much could erode the expertise that makes checking its output fast.

One last point: faster coding doesn't automatically mean faster work overall. Evans argues that making tools easier to build leaves two problems untouched. People often don't recognize which of their tasks could be automated, and organizations still have to coordinate adoption across departments Does easier tool-building actually solve enterprise adoption problems?. The corpus has only a handful of controlled studies here, and they disagree. Treat any single headline number, whether +21% or −19%, as a snapshot of particular tools, people, and codebases rather than a settled answer.