Code review has never really been about catching bugs. It's how teams share context, enforce standards, and build collective understanding of the codebase. The bug-catching is a side effect of people actually reading each other's work.
AI-generated code complicates every part of that.
What Changes
When a chunk of a PR comes from a model, the author's relationship to the code is different. They directed it more than they wrote it. That's not inherently bad, but it changes what review needs to catch.
The things that are harder to spot: subtle logic errors where the code looks clean but mishandles an edge case. Confident-but-wrong patterns — the model picks a reasonable-looking approach that doesn't account for something specific to your system. Naming and structure that's technically fine but inconsistent with your conventions in ways that accumulate over time.
What I've Changed
I review AI-assisted PRs differently now. I spend less time on syntax and formatting — the model handles that well. I spend more time on behavior. Does this actually do what the ticket describes? What happens on the unhappy path? Are the test cases covering the real risks or just the obvious ones?
I also ask more questions in reviews. Not "why did you write it this way?" but "what did you prompt for, and what did you change after?" That conversation reveals whether the author understood what the model produced or just accepted it.
The Bigger Question
We're recalibrating QA processes against outcomes instead of specific syntax and semantic review. Line-by-line code review is falling out of favor. What's replacing it is a focus on specified behaviors and test suites that express expectations clearly. Does the system do what it's supposed to? Do the tests prove it?
The open question I keep coming back to: does the review skill itself atrophy if you're not careful? If you're reading code you didn't struggle to write, do you engage with it deeply enough? I don't have a clean answer yet. But I think the engineers who stay sharp are the ones who keep asking that question.