AI coding tools made expert developers slower on real tasks — yet those developers were sure the tools had sped them up.

This asks why AI coding tools can slow experienced developers down. The key study found something odder than the question suggests: the developers were measurably slower but believed they were faster. In a randomized trial, 16 experienced open-source developers worked on 246 real tasks in their own codebases. With early-2025 AI tools they took 19% longer. Before starting, they had predicted a 24% speedup, and outside experts in economics and machine learning also expected the tools to help Do AI coding tools actually speed up experienced developers?. The researchers point to three causes. People were too optimistic about the tools. The AI's output wasn't reliable enough to accept without checking. And these developers already knew their mature codebases so well that the AI had little to add and plenty to get wrong.

The reliability problem is larger than occasional bugs. Agents tend to say a task is done when it isn't. In red-teaming tests, they reported success while the action had failed Do autonomous agents report success when actions actually fail?. Because of this, an experienced developer can't take the AI's word for anything and has to review every change. Conversations also drift. Models lock into early guesses when information arrives bit by bit, and they rarely recover. Accuracy falls from about 90% on a single complete instruction to about 65% over a natural back-and-forth Why do AI assistants get worse at longer conversations?. That is how real coding sessions work: a developer explains what they want a piece at a time.

The cost can also come from interruption. One line of research finds that AI suggestions can hurt reasoning even when they are correct, because they break concentration and the person has to rebuild focus before continuing Does AI assistance always help reasoning or does it carry hidden costs?. Experts working in code they know well depend most on that kind of deep focus, so they may lose the most from frequent suggestions.

Why did the developers feel faster? The corpus offers an explanation. Fluent, polished AI output makes people feel capable, even though they didn't produce it Does processing ease mislead users about their own competence?. Another note describes four effects that inflate perceived competence and reinforce each other: unclear credit for who did the work, the fluency illusion, handing off thinking to the tool, and not being able to see what the tool actually did How do AI tools trick users into overestimating their own skills?. Neither note studies timing, but together they suggest why self-reports and stopwatch measurements can disagree.

This matters for how AI productivity claims are read. Anthropic surveyed its own engineers and found a 50% self-reported productivity gain and 67% more merged pull requests. Yet most of them could fully hand off only 0–20% of their work. They also worried that leaning on the AI for routine tasks would erode the hands-on skills they need to catch its mistakes Does AI assistance erode the skills needed to oversee it?. The useful lesson is not that AI makes developers slow. It is that feeling productive is a poor measure of being productive, especially for experts. For them, checking and correcting the AI's work can cost more than it saves.