The Cognitive and Empirical Limits of Human Review on AI-Generated Code
Relying on manual code reviews for massive AI output is structurally unsustainable. Empirical software engineering research reveals strict thresholds for human review efficiency: maximum 1-hour sessions and 400 lines of code per hour.
Impact: Medium
Why it matters
You should adjust your engineering team's vibe-coding delivery expectations, knowing human review capabilities cannot scale to match instant LLM outputs.
TL;DR
- 01Human review efficiency drops exponentially after 1 hour or when exceeding 400 LOC per hour.
- 02Treating LLM code assistants as interns who just need manual human reviews is mathematically unviable for scale.
- 03Reviewers of AI code are demonstrably more confident but objectively less effective at catching bugs.
Key facts
- Max effective review speed
- 400 LOC/hour
- Max review session duration
- 1 hour
The 400 LOC Bottleneck
Software engineering literature establishes that a professional reviewer cannot reliably evaluate more than 400 Lines of Code per Hour (LOC/H). If a coding assistant generates 4,000 lines of boilerplate, tests, or application logic, a human reviewer would require 10 full hours of hyper-focused attention to inspect it properly. Since most engineers only have a few high-quality concentration blocks per day, infinite code generation hits a strict human cognitive wall.
The Illusion of Quality
Empirical research highlights a dangerous psychological asymmetry when humans inspect AI code: reviewers often report high confidence that they caught all bugs, yet objective metrics show they catch fewer errors in LLM-generated code compared to human-written PRs. This overconfidence combined with high output volume drastically decreases the overall maintainability of a codebase.
What to do today
- Structure your pull requests to never exceed 400 lines of AI-generated code per review block.
- Limit manual engineering code-review blocks to a maximum of 1 hour to prevent fatigue-induced bugs.
Sources