I'm not sure how you would detect those clusters? The language models wouldn't give it away for sure, and getting residential IPs has never been easier than in the present.
Social graph metadata. The GGP said they would all be "talking to each other."
I think by 'detect' they mean 'detect by people who might fall for it' not 'detect by national security agencies.'
I don't see any way for the average social media user to detect whether they're engaging with a botfarm unless the social media company exposes their analytics, and they don't, and they aren't likely to start unless forced.
I actually was thinking the people who'd do the detecting would be the social media site operator (e.g. Twitter), who would then shut down the inauthentic activity.