Oh and duh

I'm writing on this one again -- and probably to the exclusion of Tumblr -- because after seeing those initial scenes with Livejournal in "The Social Network" I got nostalgic for how much fun it was to externalize my thought processes.

Thinkin'

So there are two general categories of AI that seem to make sense. Others probably do, but here's where I'm at right now.

There's the kind of AI that attempts to do what a human can do, but unattended, automated, repetitively, reliably. It acts like somebody with a reeeeallly clear performance plan.

Then there's the kind of AI that interacts with a person, acts like a robot butler or exoskeleton.

The second kind is a heck of a lot harder to build, right? Because that's the one where you need to understand how people think and act, rather than how a task needs to be divided and conquered. The first kind could generally be addressed with a recipe, or a recursive algorithm. The second has to be on pins and needles, constantly insecure that it might lose the user's attention, consideration, interest. It has to act like an improv performer, or a European-style clown.

What would go into an expert system that didn't act "expert-like", condescending, dismissive like a Call Centre employee? How would you build an expert system that acted like a bartender?