What are code reviews even for?

AI coding tools are generating code faster than humans can review it — at Meta, lines of code per diff grew 106% and diffs per developer rose 51%, mostly from agentic AI. But the real problem predates AI: teams were already struggling with large PRs, poor change descriptions, and reviewer overload. Research shows code review was never just about defect detection; only 14% of review comments are defect-related. The rest serves knowledge transfer, shared ownership, and architectural understanding. Meta's RADAR system offers a model — automating low-risk diffs while routing higher-risk ones to humans, achieving a revert rate one-third of normal and a production incident rate one-fiftieth as high. The recommended approach: fix review fundamentals first (small PRs, meaningful descriptions, automated routine checks), then design AI assistance around human judgment rather than replacing it, and actively protect the knowledge-building function that human review provides. Automating away review risks accumulating 'cognitive and intent debt' — a growing gap between what a system does and what the team understands about why.