A few reasons come to mind. One is simply that if you have any metrics (ranking, win/loss ratio, greater site access..) it's going to feel nice to see them improve. Another is that losing at anything can be ego-hurting (similar reason good players sometimes sandbag with new accounts / lower ranks they can't possibly lose to, they need to 'win' more). Or reverse sandbagging/trolling with a bot might be amusing. Another is the cheater may justify it as a self-teaching game, and might not always play the strongest move but see if their move is even in consideration or try to improve ability to see the better moves by having them always pointed out -- but why not just play the bot, or save that for post-game analysis? I like to run my go games through gnugo's annotated analysis at the end (as I'm very weak I assume even the weak gnugo can teach me things), it'd be too troublesome to use it in a live game.