Do you even read the code anymore?
Friends who aren't developers ask me this a lot and I always answer with "it depends." Not all software carries the same size, complexity or risk.
That answer can sound like a dodge, so here's what I mean by it.
Think about buildings
A shed, a house, an apartment building and a skyscraper all need different plans, permits and engineering.
Nobody would hold a backyard shed to the same process as a skyscraper, and nobody would build a skyscraper the way they build a shed. The process scales with what's at stake.
Software works the same way.
The shed: a script I write for myself
A shed is a script I write for myself. If it breaks, I fix it. I barely read the code.
I'm the only one who depends on it, so if it goes wrong, I'm the one who finds out and fixes it. Spending a lot of time reviewing it wouldn't buy me much.
The house: a small website
A house is a small website. I click through it and spot-check the parts that matter.
The check here is mostly using it the way a visitor would, and looking closer at the pieces where a mistake would actually show. More people see it than a script on my laptop, but it isn't something people depend on to do their jobs.
The apartment building: a business app other people rely on
An apartment building is a business app other people rely on. Now there are automated tests, and I approve the AI's system design before it writes anything.
This is where the checks start to change shape. Other people are relying on it now, so I don't want to find problems by clicking around after the fact. Automated tests check the behavior every time, and approving the system design first means I've agreed with the plan before any code exists.
The skyscraper: a payment system moving millions of dollars a day
A skyscraper is a payment system moving millions of dollars a day. That means design reviews, tests on every change, a human reading every diff, a slow rollout and alerts when something looks off.
Each of those is another check between the code and the people it affects:
- Design reviews before the work starts.
- Tests on every change, not just the big ones.
- A human reading every diff, so nothing ships that nobody has looked at.
- A slow rollout, so a problem reaches a few people before it reaches everyone.
- Alerts when something looks off, so someone knows quickly when something goes wrong.
At this level, yes, someone reads the code. Reading it is just one check out of several.
The pattern
Look at the four buildings together and the pattern is simple. As the risk goes up, the checks go up with it. At the shed, the check is me fixing it when it breaks. At the skyscraper, there are checks before the work starts, while it's being written, as it ships and after it's live. None of that is about distrusting the AI. It's the same thinking you'd apply to any builder: the bigger the building, the more inspections it needs.
So, do I read the code?
AI writes more of the code now. The question isn't whether you read it. It's how much risk the code carries, and how many checks you put between the AI and production.
If you're not a developer, that's the question worth asking the people building software for you: which building is this, and what checks are in place for it? A good answer should sound different for a shed than for a skyscraper.
What would you put at the top of the list for the skyscraper?
Chase Parks writes about putting AI to work at WorkSpec AI.