I don't review every line of AI-written code

There's a comforting idea going around: the safe way to use AI is to read every line it writes. It sounds responsible. I don't do it, and I don't think you should expect anyone who ships at a real pace to either.

A serious build is thousands of lines. Reading all of them is slow, and worse, it gives false confidence — you can stare at code that looks perfectly fine and still miss the one thing that falls over in production. Line-by-line review is mostly theatre. It feels rigorous; it rarely is.

Where the risk actually lives

Bugs that matter don't usually hide in some random function — they live in the shape of the system. So that's where I spend my attention:

• Architecture and the data model. Get these right and most of the individual lines genuinely don't matter. Get them wrong and no amount of line-reading saves you.
• The risky surfaces. Anything touching money, auth, personal data, or a nasty edge case — I read those closely. The routine CRUD and glue code, far less so.
• Tests and behaviour. I'd rather the software prove it works than read every function and hope. Tests catch a regression every single time; a pair of eyes catches it once, on a good day.

Trust, then verify — at the right level

I treat AI like a fast, capable colleague: I give it clear direction, I check the work that matters, and I verify the result by running it — not by babysitting every keystroke. That's not "rubber-stamp whatever comes out," and it's not "audit every character." Both of those are how you either ship junk or never ship at all. The job is knowing which 5% deserves a hard look, and being disciplined about that part.

What this means if you're hiring someone

If a builder tells you they review every line of AI code, they're either not shipping much or not being straight with you. The better questions are: how do you test it? What do you monitor once it's live? And how do you decide what's worth reading closely? That's where real safety comes from — not from pretending a human re-types the machine's work in their head.