Would you ship code you don't fully understand?
I rebuilt our entire marketing site with AI doing most of the coding, and there are chunks of it I've honestly never read line by line.
For a marketing site, whatever. Worst case a page breaks and I regenerate it.
But I can feel the line moving. A year ago I reviewed everything, now I review the parts that touch data or money, and I'm not sure where this trend stops.
The pushback I'd give myself: you already ship mountains of code you've never read, it's just other people's. Dependencies, frameworks, the whole stack. Maybe AI code is no different.
Where's your personal line right now? And is it moving?
35 views
Replies
Great topic,@maxmusing.
I believe AI-assisted development has its benefits and can push you beyond your current level of comfort as a solo developer. That said, I definitely draw the line at letting Cursor (my current favorite) fix Tier 1 bugs. The reasoning isn't that it can't do it, but that it likes to waste tokens guessing. However, telling it what to fix is a lot different - providing an output of a network HAR file and then uploading that network log into Cursor, it can find and fix the bug in about 3 minutes.
That being said, I will let Cursor have full "creative freedom" on our marketing site and there's thousands of lines of code on it I don't even recognize, but I make sure to proofread the output and make changes before pushing into Prod.
Basedash: AI data analyst
@austinbuhl yeah this is basically where I’m at too. I’m comfortable giving AI a lot of latitude on low-risk surfaces, but for T1/debugging I still want a proper human review before letting it ship anything.
my line has stopped being about whether i understand the code and started being about whether i understand the blast radius if it's wrong. we let agents write plenty of stuff nobody reads line by line, but only in places with fast rollback and cheap failure, staging configs, internal tooling, generated tests. anything that touches billing or auth gets read slowly no matter who wrote it. feels less like a new problem and more like the same judgment call we already made with ORMs and frameworks, just moving faster
Basedash: AI data analyst
@galdayan even better way to put it than how I did. Low blast radius + fast rollback changes the calculus a lot.
For me, the real question isn't whether I understand every line—it's whether I understand enough to maintain it with confidence. If the answer is no, I'll review the logic , ask teammates, or create a small proof of concept before moving forward.
@maxmusing Yep, definitely. As a designer, I don't understand any code. I tried to vibe with code and learn what AI does, but after talking to senior devs and understanding what's actually happening behind the scenes, I stopped looking at the code totally.
If a problem's solved or a feature's built, I ask another AI agent to check if the build's correct or can be optimised.
Basedash: AI data analyst
@roopesh_donde thankfully AIs are great at understanding and explaining code!
Would I ship code I don't fully understand? Yes, unless it's a critical system. And even then, might still be yes, if there's enough tests and tooling to find issues before they arise.
From my experience AI is on par or better than a fair share of human developers. Over the years I've seen so much, even deliberately negligent human development, that I trust AI more, especially with quality instructions and plan.