Human validation with an image

[Peter Hansen]

Aahz wrote:
> 
> With thousands of registrations, you're going to have at least a few
> blind or visually-impaired users.  Bad move.  Instead, ask a question
> that requires intelligence to answer.  (E.g. "What color is the sky?" or
> "What is the third word of the second paragraph?" or "Give me the first
> letter of each word of the following phrase: spam and eggs")

I was thinking this problem shouldn't be very hard, given how
far computers are from passing the Turing Test...  on the other hand,
one difficulty with any of the above is that if the test is fairly 
static, anyone could quickly write a program that would respond 
appropriately and get past this.

One advantage of the image approach is that it is at least more difficult
to get past, both initially for the programmer, and each time from
a computational point of view.  Unfortunately, it really doesn't 
require intelligence per se.

Has no one yet written a nice "engine" which can with relatively
little inconvenience to the user, establish that user's intelligence
(and, most likely, command of one particular spoken language), in
a way that would stand up to programmatic attacks in a more robust
fashion?  What would it take?

It wouldn't be enough just to vary the "third word" part, since once
the pattern is apparent, parsing the text becomes pretty easy.  A 
fixed database of generic questions, such as "what color is the sky"
quickly becomes amenable to what is effectively a dictionary attack.

Other options?  Or pointers to the right keywords to let someone 
investigate the state of the art/literature more closely?

-Peter