There is more to code review than (automatable) detection
Amid the rise of AI-generated code and automated reviewers, engineers are reexamining what code review is actually for. Many argue that while bots and linters can catch syntax issues, basic bugs, and style violations, they fail at higher-level concerns like architecture, business fit, long-term maintainability, and knowledge sharing across a team. Others counter that human review won’t scale with AI-accelerated output and must be supplemented or displaced by stronger automated testing, design reviews, and new processes for distributing system understanding.
Purpose and Value of Code Review
- Seen as much more than bug-finding: knowledge transfer, shared system understanding, architectural defense, validation of tests and deployment plans, and even legal/organizational “due diligence.”
- Some argue PRs are the final, critical gate where architectural mistakes, misunderstandings, and tech debt can still be stopped.
- Others say if big problems are routinely caught at PR time, earlier stages (design, standups, pairing, requirements review) are failing.
Style Nitpicks vs Automation
- Complaints that senior reviewers often only comment on formatting and punctuation.
- Strong support for automated linters/formatters and a one-time style-guide decision to eliminate bikeshedding.
- Debate over bottom-up consensus vs top‑down enforcement of style to avoid endless discussion.
Humans vs AI in Code Review
- Many note LLMs are decent at local concerns: syntax, simple bugs, performance hints, tests, and style.
- Widely perceived weaknesses:
- Architectural fit and reuse of existing patterns.
- “Meta” issues like maintainability, project conventions, and repo organization.
- Business logic, acceptance criteria, and “do we even want this feature?”
- Recognizing and reducing overengineering and large, unreviewable AI‑generated blobs.
- Some report AI review tools as “now good enough” to replace most human review; others find popular tools shallow, noisy, and unreliable.
Impact of AI on Teams and Process
- Concern that heavy AI use erodes developers’ deep understanding, making cost/benefit estimation and safe change harder.
- One organization described a “crisis” where nobody felt safe changing code; they reintroduced strict human code review, WIP limits, and mandatory two‑human signoff.
- Frustration that human feedback is often just fed back into agents, seemingly training tools rather than people.
Future of Code Review
- Some predict human review will mostly disappear for non‑critical code; human focus will shift to design reviews and governance.
- Others argue relying on human review to gate AI‑generated code at AI speeds is impossible; call for AI‑driven testing, meta‑checks, and even full subsystem regeneration instead.
- Counterview: current AI output is often messy, duplicative, and corrosive to architecture; skipping human review is seen as reckless tech‑debt accumulation.