Studies on AI coding assistants' productivity gains reach conflicting conclusions

On November 7, 2026, DEV Community's AI section ran an overview of studies measuring how much AI coding assistants raise productivity, concluding that the research does not give one number and that the findings contradict one another. The piece walks through what each study measured, why the conclusions diverge, and how teams can turn that evidence into scheduling and effort-estimation decisions, using data drawn from earlier work rather than new experiments.

Three studies are put side by side. In 2022, GitHub ran 95 developers through a controlled experiment in which they built a JavaScript HTTP server; the group using Copilot finished 55% faster. Field experiments inside companies showed a smaller but measurable gain, roughly 26% more completed tasks. The best-known randomised trial, however, put experienced developers to work in their own large codebases and found they were 19% slower with AI. The article also notes that Google's DORA research has found higher throughput at the team level, though no figure is given for it. The pattern the piece draws is that results shift with the setting: tightly defined tasks show big gains, while open-ended work in existing codebases can slow developers down. As of this report there has been no follow-up from the studies' authors, and the contradiction stands as presented.