Engineering leaders must balance AI autonomy against business risk and competitive moat, as AI-generated code widens the gap between code volume and team understanding.

A 2025 METR randomized controlled trial found experienced developers forecasted 24% faster task completion with AI, estimated 20% faster completion, but actually completed tasks 19% slower—revealing a broken feedback loop between perceived and actual productivity.

Cognitive debt (the gap between code volume and comprehension) compounds silently: engineers using AI when learning new tools scored 17% lower on comprehension tests than those coding by hand, with steepest drops in debugging ability; 83% of essay-writing participants using LLM assistance could not quote a sentence from essays they had just written.

When AI-generated code reaches production without full team understanding, incident response fails: engineers called to 2 AM outages lack mental models of why systems were designed a particular way, what they connect to, or edge cases under load—and AI can explain what code does but not why it was designed that way.

CodeRabbit's analysis of real-world pull requests found AI-authored changes contain up to 1.7× more critical and major defects than human-written code; code review and team oversight are bottlenecks that catch vulnerable code, and AI makes it easy to remove them.