A National Bureau of Economic Research working paper from May 2026, "Writing Code vs. Shipping Code", analyzes more than 100,000 GitHub developers. The authors are Mert Demirer, Leon Musolff, and Liyuan Yang. The study compares productivity before and after adoption of three generations of AI coding tools, from 2022 to 2026.
What the research found
The main result: gains in code written are large, but they shrink at each later stage of the process. The more the tool does on its own, the more of the gain is lost before software is actually shipped and used.
The figures differ between versions of the paper, so this article does not repeat them. The official abstract is here: NBER, Working Paper 35275.
Why the gain is lost
The authors interpret the drop with a weak-link model: a process moves only as fast as its slowest stage. Writing code got faster, but review, integration, testing, product decisions, and approval still depend on people.
According to the study, human effort and AI only partly substitute for each other. AI does not remove the human bottlenecks that come after writing.
A counterpoint from 2025
In a randomized trial with experienced open-source developers, the research group METR found that participants believed AI had made them faster, but they were measured as slower. Secondary source: LeadDev.
What this means for teams that hire or manage software
- Measure delivery, not code volume: releases, usage, and production issues.
- Map the bottlenecks after writing: review, testing, and approval.
- Treat AI as an accelerator for one stage, not for the whole process.
Frequently asked questions
Doesn’t AI increase developer productivity?
The study shows real gains in code production. The problem is turning that code into software that is shipped and used, which depends on other stages.
Is the study conclusive?
No. It is a working paper that has not been peer-reviewed, and its conclusions may change in later versions.




