The initial image is always a resource for AI. But Generative Fill, Generative Expand, and Remove tool, are not perfect. It often takes several tries to get close to what you want. This is why an AI tool like Generative Fill, offers you several options to choose from. And if you don't like those, you can keep hitting generate until you get close to what you want/need. When you use fill, it's taking its prompt from your image and also drawing from it's own algorithmic image base. Often, you still have to go back to "old school tools" like clone stamp and brushes to finish the image.
AI does not create from nothing. It needs your resource photo to do what it "thinks" you are asking. It will replicate, or get close to replicating, your colors, textures, shapes etc. There are times when the response is perfect - there are times where it seems to pick "answers" out of its ass. That's the best way I can put it since there are times when AI leaves me scratching my head in disbelief.
Keep in mind that these tools are "conveniences" - they help you with quick results. Things that might take you an hour in research and manipulation will take less than a minute. True creativity still comes from the user. There are people here on this forum using CS 5 version and getting incredible results.
BTW, I've used both the beta and full versions. I don't see many differences. But I'm not testing it technically. I'm just looking at results.
So to answer your questions, "..does it restrict itself to source pixels in the image? Or does it grab pixels from other images that aren't mine? Or create them from nothing?"
The honest answer is yes. 
