> What about generating foliage or varied ground textures? What about generating buildings? Or NPCs?
This isn't how NVidia DLSS or AMD FSR works at all. DLSS/FSR can't create new buildings or foliage, or NPCs, so hard to foresee problems of the kind Valve are concerned with. Same for varied ground textures- the entire point of the technology is to sharpen and upscale the original image, or in case of DLSS3 inject new matching frames.
The only "risk" to DLSS is in Nvidia's own training data, but there is no "risk" of another company's existing IP leaking into final frame - again if there was, gamers wouldn't want to use it, as its destroying the original frame! If the resulting frame isn't a near perfect match for the original, DLSS has failed. Thankfully it does nearly perfectly match the original in use and alongside AMD's FSR 2.0 stuff has been one of the best advancements in gaming technology of recent years - effectively significant FPS boost "for free" on same hardware.
> https://www.nvidia.com/en-us/geforce/technologies/dlss/
> https://www.amd.com/en/technologies/fidelityfx-super-resolut...
While the line may be gray for other AI technologies in gaming, such as using it create new original textures or models, DLSS/FSR is just a really good upscaler - no "new" content being created and therefore no risk of IP infringement. To be really blunt; FSR and DLSS are in almost every new game for the last couple of years on both PC and console across literally hundreds of games now - if there was IP infringement issues, we would know by now - we are already onto second/third generations of these upscalers.