Addy Osmani puts a name to something I’ve been feeling for a while: Comprehension Debt.
When we lean on AI tools (Copilot, Claude Code, Cursor) to generate code, velocity metrics look great. The PRs are clean, the tests pass. But we quietly stop understanding why our systems work the way they do.
AI generates code far faster than humans can evaluate it. What used to be a quality gate is now a throughput problem… Surface correctness is not systemic correctness.
PR review used to be a bottleneck, but a useful one. It forced you to understand design decisions and architecture. Now a junior engineer can generate 1,000 lines of syntactically perfect code faster than a senior can audit it. An Anthropic study found that engineers who passively delegated to AI scored 17% lower on comprehension quizzes than those without AI. The interesting bit: engineers who used AI to ask questions and explore tradeoffs kept their understanding intact.
Tests and specs don’t save you here either. Tests only cover behaviors you thought to specify. And when AI changes implementation behavior and updates hundreds of test cases to match, your safety net is no longer trustworthy. Only actual understanding catches that.
What makes this worse than regular technical debt is that nothing in your metrics captures it. Velocity is up, DORA looks fine, coverage is green, and comprehension is hollowing out underneath.
Osmani’s point is clear: as generating code gets cheaper, the engineer who actually understands the system — the load-bearing behaviors, the architectural history, the context — becomes the scarce resource everything depends on.
I don’t think we should stop using AI to write code. But we do need to stop pretending that passing tests means we understand what shipped.