The complexity/frequency ratio (As in, complexity of the bot over frequency of spam) seems a little low to spend time writing something like this. A manual solution would probably be better.
Not really, the goal is to have lots of emails responding to the scammers, so they don't know who to pay attention to, so that's hard to have a manual process for that (and there's already 419eaters).
As I commented elsewhere in the topic, this runs the risk that the scammers themselves start to use bots to parse and analyze the responses, yielding to a net increase in spam traffic, although no human would read the messages.
Yes, but then the scam business would be less profitable for them since they would have to buy or build those bots to analyze answer. Plus it could create some kind of war on automated conversation bot which would be a net win for the AI in general.