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 - Hive account@dantheman and Hive account@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.