Comprehension Debt - the hidden cost of AI generated code

I agree in part, but disagree with where the focus seems to be in the article. The issue is not in the ability of a developer to retain understanding of the codebase at all times. I think the issue is in the ability of the developer to ensure that the codebase and all other relevant artifacts (commit messages, review discussions, documentation, comments, mailing list threads, forum threads, etc.) remain in a state where a competent developer can resynthesize comprehension as-needed.

I think I hold this belief because I'm persuaded to varying degrees by the parable of Chesterton's fence and the concept of Bus Factor: it is useful to know why the system is the way it is before you consider changes, and so it is useful for the means to gain that understanding to be broadly available and durable.

I think the proliferation of generative versions of code and these kinds of useful artifacts will make striving to reach this bar of comprehensibility harder and harder. I worry that I will be debugging an issue years down the line, and be frustrated at every turn:

• The code looks idiomatic and "well documented", but it feels off, and a particular detail seems incorrect.
• Oh I know, I'll look at the blame and commit messages! Oh wait, this was written with a generative model and there is no intent to be found.
• Never fear, I can check the review discussion. Oh wait, this was reviewed by generative models, any "discussion" is not intentional either.

I guess there will be some "spec" and a chat-style log of the session which culminated in it. I fear I will not be able to stomach that being the new normal in what passes for development artifacts.