I suspect the code being generated ends up falling into a local maxima, and only much later will the cost of extracting be apparent.
Right now, management has velocity so it’s happy.
Earlier they got the code understanding as a default, and it was never something that made a difference to their metrics.
The success of the LLM Code model then relies entirely on the ability of tech debt to be cheaper to solve in the future.
What you're missing in your calculus is that by the time the local maxima is reached, the business has either pivoted to something else or is throwing money at the problem to make it go away.
My biggest issue with AI is not really technical debt, it's engineers losing sight of the system's internal working. This can make it difficult to reason about the code, fix bugs or security issues.
I just don't think we have the tools right now for engineers to accept software becoming a black box.