Luke, "we do not know how many humans are being turned away by the difficulty": actually we sort of do, that paper tells this as well. It's where their study came from, and gives recommendations on what captcha techniques are best for balancing efficacy with difficulty for humans. We don't seem to be following any (except waving, which, they say, shouldn't be used alone). Then, I'm not qualified to say if their recommendations are the best and I've not searched other studies, but it's not correct to say that we start from zero or that we have to study by ourselves (an unreasonable requirement that implies we'll never change anything until we'll be forced to make our wikis read-only due to spam, as many MediaWiki users before us).
Nemo