LLMs are quite good at writing codemods from specs though, so certain specific kinds of tech debt are fairly automatable (think migrating between major library versions and removing deprecated API dependencies). If that’s the kind of tech debt they mostly have to deal with, this seems realistic.
God, I'd be a lot more accepting of AI if they just stopped with all the exaggerations and lies for once. It's not a magical thinking machine, it's just statistical next-word prediction. It could be useful for a few things, but it's not worth destroying society over.