> At a surface level they look good but if you actually scrutinize them they fall apart.
This is overwhelmingly true for AI generated code in my experience.
FWIW it makes me highly discount the perspectives of internet commenters who argue that LLMs generate "better than human" or even "mostly working" code.
The question to ask is, better than which human?
The top 25% percentile of coders in my org definitely code better than most agents. The rest? I trust the output of an LLM to be far more consistent when adding/deleting features across service layers than a human that can create accidental typos. Same thing with bog-standard React components or Docker build scripts.
This is true for a huge amount of AI output in my experience.