There is a tradeoff I’ve been thinking about quite a bit when we let an AI write code for us. For lack of a better term, we could call this insight debt.
A common misconception is that software development is mostly about producing new code. That’s half-true at best. In practice, most engineering work happens far from greenfield territory. We’re not primarily adding capability; we’re negotiating with an existing system and the debt it has already accumulated.
No matter how well code is written, every new line narrows the system’s future options and makes some changes harder than they were before. As decisions accumulate, projects lose flexibility. This is a fundamental fact. It’s not an indication that the code is “bad”. It’s just a manifestation of the fact that if you build a car, you don’t suddenly turn it into a boat three months down the line.
What I mean by insight debt is based on an observation that, in the real world, understanding the existing code is equally valuable as, if not more valuable than, producing it. Let’s say we produce 20 thousand lines of code using some state-of-the-art LLM. Now we have a bunch of new functionalities, but at the same time, we have produced a huge amount of insight debt. AI code is no exception in terms of lost flexibility. It will inevitably feature decisions and patterns that will affect anything that is built upon it. How could we account for them if we do not know what those decisions are? In other words, how can we build on a codebase we don’t understand?
I don’t think we’re quite at the point of discussing this collectively. Many are still starstruck by the fact that with vibe coding you can generate huge amounts of code that “works” in the sense that it compiles and may even pass a test suite. However, this is a narrow view, and highly insufficient in terms of real-world use cases. It’s essential that we can maintain what we build, and maintenance is impossible with large amounts of insight debt.
There is also a more psychological aspect to this. For many people it’s important to experience ownership. You know? The feeling that “I made this”. It’s something that we lose when it’s all AI-made.
It will be an interesting couple of years, when the rapidly produced systems people are reportedly making will reach the point where someone has to maintain that thing.
