More code, less progress. That is the uncomfortable truth emerging from a recent Harvard study of 718 firms, which found that AI coding assistants are boosting code volume without a corresponding increase in completed work. The bottleneck is human: AI-generated code shifts the burden to later stages like code review, where teams must untangle, validate, and integrate output that is plentiful but not necessarily finished. This finding should give every organization pause, especially those racing to adopt AI tools without rethinking their workflows.
The study, powered by the Jellyfish analytics platform, confirms what many developers have suspected. Writing code faster does not mean shipping features faster. When AI generates more lines, the downstream cost of review and debugging grows, often canceling the initial speed gains. This is a productivity paradox that no amount of additional AI agents can solve on their own. As we have explored in Discover How AI Amplifies Developer Productivity Without Burning Out Teams, the real leverage comes from how AI is integrated into the human process, not from how many lines it can produce. Meanwhile, tools like the one described in A practical tool from Google DeepMind expands how teams can identify AI-created content across platforms highlight a growing need for transparency. If we cannot easily distinguish AI-generated code from human-written code, how can we responsibly manage the review process that the Harvard study identifies as the new choke point?
The practical implication for teams is clear: adopting AI coding assistants without adapting your development pipeline is a recipe for hidden debt. You will ship more code, but you may not ship more value. The fix is not to abandon AI, but to measure what matters. Track completion rates, not commit counts. Invest in automated testing and structured review processes that can handle the volume. And be honest about the trade-offs. As the How over 10 million users are using AI to transform shopping and productivity story shows, AI adoption scales fastest when it directly improves user outcomes. The same principle applies to developers. If the output does not make their day easier, the tool is not working.
The most concrete takeaway from this study is a warning against treating AI as a volume multiplier. It is a complexity multiplier. The open question for every engineering leader is this: are you tracking the downstream cost of AI-generated code, or are you mistaking activity for progress?