Pandorabots is doing something interesting which shows a lot of promise for the future. All conversations of all 60,000 Pandorabots have been collected over the last couple of years and compiled (somewhere in the neighborhood of six million conversations, I believe they said). They're being analyzed and parsed using Zipf's Law, and the objective is that eventually they'll have a knowledgebase where you can enter a topic, and have a file returned to you of, say, the top 100 most frequent queries on that subject. This should make training AI to converse intelligently about specific topics a lot easier.