I meant world and even artwork.
You and the others are all true in theory, but it really depends on how it's done. Sorry for the length of this post and possible ramblingness, but it is something I know a bit about. Generally, as soon as you hand tweak the output, the output can no longer be replicated by whatever algorithm/procedure that was used to originally make it. Updating the algorithm/procedure is a possibility, but you will quickly come to a point where it is no longer valid, and you just need to save out each data point. There's just no real way around it.
This is just off the top of my head, but if you have a 1km^2 area, with 1 vertex per meter (on the lower end of mesh density), you'll have 1001 points since you need a start and end point, that squared is 1,002,001 points. Even if you eliminate x and z values, and only store the y (height), you're looking at 4 bytes for a float, you've got about 4 MB right there. 5000 areas, for the sake of argument, of those would be around 20,000 MB (20 GB). That's JUST terrain height, not counting all the other stuff that goes into terrain, like multiple UVs (usually x,y floats), low res versions of the terrain (to render at far distances), some kind of texturing indexing or texture blending map/data.
But let's say you have 100 areas, and you either have a masterful formula (let's say 10 kB), which either gives the desired output, or you suddenly increase that cost 40,000 times as you're having to save individual data points (from 1kB to ~400 MB). Still pretty reasonable, though it's still not including all the other data that has to go with it, and 100 isn't many areas, if you're going for raw size (ignoring gameplay/social impact).
True, it can be optimized, but it can't be discounted. You can either store it locally on the player for faster loading, or stream it (hope you've got good & cheap bandwidth), but it has to come from somewhere.
You won't be generating the terrain on the fly either, unless your AI is built to accommodate it. Most games use a nav mesh (simplified terrain mesh) to tell them how to navigate around the area, and you want your nav mesh to be tested and optimized.
As for art, you can usually expand your art options with the procedural generation, but you need the original (good) art to combine it. The more original art you have, the more combinations you can get out of it. But just doing a full set permutation usually sucks (you wouldn't put a dog's head on a person's body... unless it's suppose to be a dogman). That has to be stored on the local client too, or again streamed, but the combinations could be generated at runtime or stored on the server and sent to the players involved, so you don't have too much additional install cost there.
Without human intervention, it looks some kind of muddled messed. And with that intervention comes a storage cost.
As for content, the random dungeon/fight combinations are cool when well done, but they still have to be hand tweaked. They do something like "Ok, environment A can have Set 1, 2 or 5 of mobs, environment B can have Set 2, 3 or 5, etc..." same with loot sets. That's data that has to be stored somewhere as well. Usually server side, so it's true it won't affect the user install.
It is, by no means, a magic bullet if you want quality.
So very cool stuff, but it hasn't be done on a MMO to the scale suggested.
As a little aside, according to Steam's
latest hardware survey, 34% had 255-499 GB total space, and 25% of steam users with 100-249GB free disk space. So I suppose you could fit a massive game in there.