This is fun to think about. If I ever have time, I'm going to write exactly the bot you described - it will run some sort of machine learning algorithm to learn what features of an article lead to success. My guess is that the correlation between payouts and authorship is extremely strong - @dantheman and @roelandp posts are clearly near the top of the list.
That would be an example of a "selfish" bot - it would be voting with the sole purpose of maximizing its curation rewards. And you're right - if too many whales run this type of bot, it becomes circular. Each bot would be upvoting because the bots are upvoting, and we could lose the real correlation between quality and payout. This is something like your Scenario 2.
On the other hand, I could write a bot that doesn't learn what features make an article high-paying; rather, it gauges each submission by some more objective standard. For instance, I could write a bot that checks 1) spelling and grammar, 2) plagiarism, 3) presence of images, and maybe some other things. This bot would act as a first check on post quality; it wouldn't be perfect, but it also wouldn't have any risk of creating the feedback loop that my first bot makes. This bot wouldn't garner as high curation rewards as the first bot, but it would probably be quite good for the Steemit ecosystem.
Make sure to add me in git if you open a repo - I'm FelixA on github.
Interestingly, nearly at the same time I wrote my article there was another one that offers a bug bounty for the second bot you described https://steemit.com/steem/@cryptoctopus/usd500-bounty-we-need-this-now
Actually I am pretty sure that a selfish bot would be used very fast by very many people. I actually expect that some bots like this are already being used. I also have to admit I would use it, since in a smaller scale it is not selfish, only if it becomes viral. IMHO this is a prisoners dilemma in Steem.
This makes me quite afraight of bots.
Interesting. That does look like a prisoner's dilemma. I wonder how we could quantify the system's resistance to bot-overrun? Could a whale make a robocop bot to flag content that's had too much bot activity?
I wish I had more time to devote to this stuff. I'll look into Xeroc's piston.steem thing; that would be a pretty fun project.