I think you may have to expand a little more on how the players would interact with each other's worlds, and how you plan to actually generate those worlds in the first place.
But related to what you're talking about, and an idea that may be useful to you; I have long been thinking of an idea based on community 'script' development and a loose AI director.
The point would be in a procedurally generated RPG would be able to string together 'scripts' or 'plot lines'. The game could either be run in 'strict' mode that follows a detailed long plot line from start to finish that would be written in the style of traditional cRPGs, or in a 'string' mode, that will tie shorter scripts together for a more Rogue like feel.
Scripts would belong to various levels of a hierarchy, and use standardized tags to identify things. So at the top level someone could write a script about the interactions of two "Big Bad Gangs", and how "Character A" is a triple agent, betraying first one gang, then the other, and coming out as a mastermind. But the script itself doesn't need to define much of anything about either gang, or Character A. Low level scripts would define details about character, stats, precomposed lines of dialogue script, etc. And then a few layers in between based on how complex your system is.
So when running in loose mode the game would start pulling together scripts from various sources and using them to guide a player through a game.
The elements that I see making such a system work well (And it is an insanely complex one that I haven't devoted serious time to developing due to its scope) would be community management methods. You ideally would have a detailed ranking system for the value of the scripts, as well as metadata describing what kind of script it is and what kind of stories it is suitable for. An automated blind peer-review system would be idea, having players score and provide feedback to authors, but not be able to see any consistent identifiers. (This is to help avoid trolling. Trolls can't work together as easily to push through deliberately 'bad' content, nor focus attacks on other people's good content because they will have no control over which user's content they're rating. And then automated ranking systems should help bring vetted authors up to the top of the list.)
Players then would configure their game generation through filters based on community standardized key-wording. So they have a loose control over what kind of plot lines and scripts the system will select to throw at them. (Or could choose to select from the hard written modules instead.) You can also include a feedback system from the players as well, having them score their enjoyment of the game when they finish. That info could be tied back into future game generation in some way, or spread points around the scripts the game used. After awhile you could build up a system and core dataset that can generate a huge range of really compelling games that combine good elements from hand crafted game worlds with the unpredictability and replayability of procedural content.
In theory you could also extend such a system to game assets, and add additional elements to your visual or audio aspects of your game.