http://www.cleverbot.com/

Fun to play around with! 

From a discussion thread:

The bot learns from talking to people. Same principle as jabberwacky; in fact, same creators.

You are talking to people, in some sense. The responses are human responses. Everything it says is based on something someone once said to it. It chooses a response it feels most likely to fit in the context of the conversation, though there's a fair bit of randomness involved, too.

It's not in real time. The responses may have been in the database for a month.

Users aren't responding to what you are saying either, but rather statements that are judged similar to what you are saying, unless you say something that someone else has exactly said before, but even then if there isn't enough data it will also look at similar statements to help pick a response.

Looks like Cleverbot is passing the turing test.