What AI catches first
Across 50,000 review sessions: AI is significantly better at spotting off-by-one errors (93% catch rate vs 61% for human reviewers), SQL injection patterns (88% vs 54%), and time complexity issues (77% vs 69%). AI is faster — average first comment in 2.1 seconds vs 4 hours for async human review.
What humans catch better
Humans significantly outperform AI on architectural concerns (91% vs 44%), naming and readability issues that require domain context (88% vs 51%), and 'this doesn't match the spec' issues that require business logic understanding (79% vs 33%).
The optimal prompt
Don't prompt AI to 'review my code.' Instead: 'Review for correctness and time complexity first, then flag any security issues, then comment on naming and style. Focus on the algorithm in the hot path.'
typescript
// Our Claude API call
const result = await anthropic.messages.create({
model: 'claude-sonnet-4-6',
max_tokens: 1024,
messages: [{ role: 'user', content:
`Review this ${language} code for bugs, complexity, and security.\n\n${code}`
}]
})