you gotta take into account that the conceptual design phase has been completely eliminated from the process. And lots of the core logic would've been ported over wholesale. one example you can find is searching about the code in the katamari ports, where there's parts of the code left behind that don't do anything, and point to ps2 libraries that don't exist in Unity.
while there's nothing revealed here on how much is just copy/pasted over, I'm sure they'd be motivated to bring over as much as they can when they were initially estimating the work needed for the project, and then test the hell out of it. A majority of the graphics engine would probably have to be remade if they didn't aim to emulate it, and would be one of the major challenges in a mobile port of a higher-end console/pc game, but the assets/resources would be copied over and no work on that beyond ensuring it looks fine on the target resolution and framerate; maybe even with the use of automated AI upscaling/downscaling to reduce that workload even further. I find it safe to assume a straight port doesn't usually require the same or more amount of work as the original
I'm not a fan of games that are designed by committee, and I fear AI-generated games would take that to the 11th degree.
Given that, I feel very specific aspects could still be vastly improved by AI, like games that implement procedural generation; I feel like his mention of procedural fears more of everything becoming procedural, instead of it supplementing the pre-existing applications of it. Those kinds of games hit a plateau at a certain point in the gameplay loop - the limitations of the tiles or combinations of assets starts to become very predictable and doesn't achieve the purpose it sets out to at that point.
Also to take into account, AI needs a dataset to train it, and to avoid the homogenization he fears would involve producing datasets for specific tasks, and differentiating them from one another; to me, devs producing these unique datasets to sell is inevitable, and there's definitely going to be a lot of "shovelware"-quality datasets being thrown around. The ethics of the data contained in a lot of those kinds of ones will definitely be questionable.