Third-party dispatches, not graded receipts
Study: AI coding agents generate more code but not more shipped software
A study reported by Ars Technica finds efficiency gains from AI coding agents are absorbed by a human review bottleneck, so output volume rises without more finished software. This is direct empirical context for claims about AI writing most of the world's code.
“Study finds coding efficiency gains get “absorbed” by human review “bottleneck.””
Bears on: amodei-ai-writes-90pct-of-code-by-2025 · superhuman-coder-2027