It reminds me of chatting with scammers who don't speak good English on dating sites. My first test is whether they can answer questions accurately such as "How's it going?" or "What are you up to?"
They can have responses to those. Thinkbot for example can. But they don't understand the topic of the current conversation usually so their next response wouldn't make any sense.
There should be some meta-markov states on top of the conversation states so as to not randomly deviate into an unrelated topic.
