That's exactly what they are: a visual interface to tags and tagged social graph nodes.
I've always heard that only 5% of email users use Gmail, but I don't actually have a source. Does seem to be accurate based on my (non-techie) friends.
34.5% yahoo.com
24.0% other
13.9% aol.com
9.9% gmail.com
2.6% comcast.net
2.2% msn.com
My mistake.
That's not how it works. You can put people into as many circles as you like.
Even simply the most contacted people in the last week - and sharing to those people the most, even though you might have a friend you haven't seen for a long time, and they might be a great friend, that doesn't automatically mean you want to share with them. To assume that people want to share more with better and older friends and less with newer friends is wrong.
Users should get a menu of content automatically provided based on prior contact and clique analysis, and pick the items and people they find interesting to ensure more permanent viewing, which in time can fade over time if that content isn't liked. Even tracking what is clicked can automate this without having to manually like or +1 something - it's like listening to a person speak.
The last thing I want to do is setup a G+ account and start customizing friends lists, circles, clicking like and dislike buttons - it should be automated, and Google should delve into my GTalk and Email to see people I've been in contact. Whose emails am I replying? What people am I having conversations? I don't quite know who all my online friends are - G+ should work that out for me and cement those relationships without my input. Online friend dynamics are too liquid to be manual.
The problem isn't circles, it's the way they are created, edited and devised.
I'd imagine this will be expanded and soon find its way into Google+.
A person can belong to more than one circle (that's why you have a list of checkboxes, one per circle).
It's also cognitively tiring to work out the consequences of sending a message to a circle larger than what you might type into an email "to" field.
You can work out the consequences of a broadcast announcement to the world/all your friends/your whole company. And you can work out the consequences of a private message to a few friends.
But in between is hard, and kind of unnatural. Maybe people will learn, but I think most people will only use a few circles (Public/Private) at most.
Random thought: the main reason for Circles is to reduce the negative consequences of sharing an identifiable message with someone inappropriate. If that could be predicted via machine learning, that'd be interesting.