lucy24 - 8:03 pm on Aug 31, 2011 (gmt 0)

Oh, hey, I remember reCAPTCHA [recaptcha.net]. (It's the same link as above, by another name.) Had to go all the way back to 2007 to find its discussion. Can't post the link that I originally read, because you have to log in (I checked with a different browser) so I'll quote the key point.

Uninformed consensus is likely to be wrong. Consider e.g. long-s: 99% uninformed readers would read f. And this is the easy case. Outdated spelling is another easy case.

(He goes on to talk about analogous language-related problems, such as choosing between final o and e in Italian.) To say nothing of yogh and thorn...

Anyway, g### has an awful nerve promoting the idea, when they're widely known to have the worst scans-- and consequently the worst OCR-- in the business ;)

Come to think of it, I've been on places where they offer multiple-choice tests of the "2+2" form. Catch is, there's always an x% chance of randomly guessing right, so you have to do several of them to get the probability low enough to meet your security needs. .25^{some number}, .2^{some number}, like that. Even if you know what they're doing, it always feels as if you answered wrong and they're giving you another chance.