AI Wrote 741% More Code and Shipped Just 30% More, Says a 100,000-Developer Study
An NBER study of over 100,000 GitHub developers found AI agents lifted code volume by 180% and finished releases by only 30%, because humans still decide what ships.
Economists matched more than 100,000 GitHub developers with their real AI usage and tracked the work from keyboard to shipped product. Conversational assistants cranked out 741% more lines of code, yet finished releases rose only about 20% to 30%, because review, approval, and human judgment never sped up. The same paper checked four app marketplaces and found a flood of new apps with no rise in anyone actually using them, just a growing pile of software nobody downloads. The garbage fountain runs at full pressure, and the productivity revolution evaporates at the one step where a human still has to care. Quartz on the GitHub coding study
More from Labor
Bosses Fired Workers for AI, It Flopped, Now They're Quietly Rehiring
Bosses fired workers for AI, the AI flopped, and now they are quietly rehiring at a loss.
Gen Z Rated AI a Minus 44
Gen Z looked at the product sold as their replacement and rated it minus forty-four.