Given the quality of a lot of Facebook comments any of the Loebner Prize bots would have a reasonable chance of passing as human.

(http://www.loebner.net/Prizef/loebner-prize.html)

A bit of tweaking - some learning and some Markov chains - and you'd have something reasonably solid for that narrow gap.

Obviously there are trivially easy ways to trip up bots, if you suspect they are bots.