Gotta say though, while the DLSS 5 character face "improvements" look worse on every game shown except for Starfield, the AI lighting improvements on game environments are crazy good:

I actually much prefer the left even though the right is more photorealistic. It looks like someones taken a film frame and made it look more documentary like

Yeah, whether or not each gamer decides to turn on DLSS 5 (or similar techniques that are sure to be coming to AMD's FSR and Intel's XeSS and Sony's PSSR shortly afterward), seems like it will come down to stylistic vs photorealistic, which one each gamer likes more. Personally, on games that were originally aiming for photorealistic, like i would say AC Shadows was, it seems like it might be worth turning on (not sure what it does to the faces on this one though), but obviously on a game that is aiming for stylism to start with, cell-shaded anime games, Kingdom Hearts, Borderland, Fortnite, etc., this doesn't seem like it is something one would want to turn on since it will change the original artistic vision pretty radically. 

Not gonna lie.... in most cases DLSS 5 looks better.
i'm not some "it loses the original artistic identity" guy. I would pick the one that looks better, and sometimes DLSS5 looks way better!
This realy looks great, IMHO, too good to be real.

This is dogshit slop, Digital Foundry should be ashamed of themselves for promoting it.

What fresh hell is this?

I wouldn't say it falls flat. When all available character art & renders look completely different from the DLSS 5 rendition, you've got to question to what extent the outcomes of the technology are turning into their own beast even with developer input. This already happens with upscaling technologies and its criticised there when it changes textures, removes particle fx etc but this is on a different level. Capcom is obviously happy to show this off and this it is experimental, but I don't think that dismisses the very obvious disparity.

What makes an upscaler trained on game data and a bunch of different NeRF-likes trained on that same exact data different from an ethical perspective? I have been seeing the take that John Linneman from Digital Foundry should be against this since he is generally against unethical AI, but from a training perspective DLSS 2 and 5 are both using the same data, which is data that is produced or procured  consensually and not scraped from the internet. 

Is it because it more directly displaces work done by artists? In that case, can't the same be said of raytracing vs. baked lighting? 

If it is an issue with quality, then that will almost certainly improve over time, just like upscaling. 

You should see what DLSS5 ultra looks like