An NBER working paper tracking more than 500,000 GitHub developers finds AI coding agents sharply raise commits but lift actual software releases far less, pointing to human bottlenecks after code is written.
GDP measures market production, so free AI tools show up mostly as costs rather than benefits. New BEA research and Stanford surveys estimate the gap in two different ways, and their numbers cannot simply be added.
U.S. productivity rose at a 1.4% annualized rate in the second quarter while hours worked also grew, and most AI-using firms report no AI-driven employment change. The shifts showing up are in who gets hired.
Census research finds 64% of AI-using firms made no organizational adjustments to use it, and only about 15% developed new workflows. Here is what the evidence does and does not say about redesigning work first.
