AI coding tools haven't made us ship faster, and researchers are figuring out why.

A collection of recent academic papers and talks is examining a persistent paradox in software development: despite the proliferation of advanced AI coding assistants like GitHub Copilot and their promise to accelerate development, teams aren't actually shipping software to users any faster. The research, presented at venues like ICSE and in NBER working papers, investigates the 'activity-outcome gap'—the disconnect between increased coding activity (more lines generated, more commits) and meaningful project outcomes (features delivered, bugs reduced).

For developers, this means the promised productivity revolution remains elusive. The studies suggest bottlenecks have simply shifted from writing code to other parts of the process, like code review, integration, testing, and decision-making, which AI tools don't yet fully address. The sharp takeaway: buying an AI coding tool for your team might make individual developers *feel* more productive, but it won't shorten your release cycles unless you also fix the surrounding organizational and process constraints.