Code Review Was Never The Bottleneck

Reviewing a colleague's pull request starts from trust. They understood the problem, wrote the fix, and probably tested it before pushing. My job is to catch what they missed, not to re-derive whether they understood the assignment in the first place.

Reviewing a model's diff starts from nothing. It didn't understand the problem so much as pattern-match toward something plausible. It didn't test it unless told to. And it will say it's done in the same confident sentence whether it's actually done or just looks done. That's the whole shift. Not worse code. A different kind of author, one that never learned to say it isn't sure.

what actually goes wrong

It's rarely syntax. Models are good at syntax. What slips through is scope, confidence, and claims of correctness nobody earned.

Ask for a bug fix and get the fix plus three refactors nobody asked for, each reasonable in isolation, none reviewed the way the original fix was. Ask for a CSS change and get told it's fixed, when fixed meant a rule was added, not that anyone looked at the rendered page. I had exactly this happen building my own site. Two buttons, active and hovered, were both meant to look selected, and a fix landed that only worked until you hovered the button you'd just clicked, which is the one thing anyone actually does right after clicking it. The diff looked complete. It wasn't complete until someone opened a browser and inspected the actual computed style and saw the text color match the background.

The same thing happens with facts, not just code. A thesis draft I was writing carried a claim about two ride-hailing apps and an OTP flow, stated with total confidence, and it turned out backwards on the first real check. Confidence was never the signal, not even with people, but a person's confidence at least correlates loosely with whether they checked. A model's doesn't correlate with anything.

the discipline that replaces trust

With a colleague, you skim the parts you trust and slow down for the parts you don't, and that instinct is usually right because it's built on actually knowing them. With a model, that instinct has nothing to calibrate against. Every line needs the scrutiny of the first line, because there's no history to lean on for which lines are safe.

Nothing is done until it's verified the way a stranger would verify it, run it, look at it, read the actual output, not the description of the output. A screenshot of the change beats a sentence describing the change. A computed style beats a claim that a rule was added. A checked fact beats a fact stated with confidence. It's slower than trusting the summary. It's also the only way to know the summary was true.

the bottleneck moved

Writing code got dramatically faster. Reviewing it didn't, because reviewing was never a typing-speed problem. It was a judgment problem, and judgment doesn't scale just because the model does.

The teams that ship faster with AI-generated code aren't the ones generating more of it. They're the ones who rebuilt review around the assumption that nothing is true until it's checked, and got fast at checking.

That's the actual skill now. Not writing code. Reading it like it hasn't earned your trust yet, because it hasn't, until you've looked.

what actually fixes it

Verifying harder isn't the whole answer, it still leaves someone reading and running every line. The cheaper fix is needing less of it. A short standing set of rules, don't invent this pattern, don't touch that file, always verify this way, turns a hundred small judgment calls into one decision made in advance instead of caught after the fact. An open-ended request produces an open-ended diff too, so asking for one bounded piece of work at a time keeps what needs reviewing small enough to actually hold in your head. I've stopped a task mid-way before, three steps into a direction I hadn't actually asked for, and the fix wasn't reviewing the output more carefully once it landed. It was cutting the task back to what I meant before more of it got written.

More on actually fixing this soon, it deserves its own piece.