Accessible Captchas

I work at a major university, which means that websites I produce are required to conform to Section 508 standards for accessibility. One of the things we often want to use with campus services are captchas, which are a way of displaying a human-readable (but hopefully not machine-readable) image and having the user key in the letters that appear in the image.

Most sites, when faced with the accessibility issue, offer a spoken version of the captcha as well so that users who are visually disabled or who can’t make out the letters have an alternate way to access the information. This still leaves out many elderly people who don’t see or hear well, users who are using a computer in a public place, and people who are both deaf and blind and use tactile screen readers.
Another way of handling the issue, which I’ll be implementing on both my customer’s sites and on my work sites in the near future, is to have a phrase that means nothing to a computer but would be easily complete-able by a human.
Eventually, a computer could be *taught* all of the phrases, as they are readable by computers. But phrases requiring reason or allegory and have a large dictionary to pull from are nearly impossible to guess.
Good examples:

• Q: “What’s the symbol of valentines day?”
  A: Must include the word ‘heart’.
• Q: “_________ is to cowboy as crown is to princess.”
  A: Hat.
• Q: “The Quick Brown Fox Jumped Over the _____ Dog.”
  A: Lazy

Other questions include easily googleable trivia, links that a user has to read something from to get the question right, and other questions that might require a judgement call. Bad questions would include something that can easily be counted by a computer — i.e. provide a link, and have the user count the number of “and”s on the page linked to. A computer would actually find it easier to do that. Turing questions aren’t necessarily the best as computers can be trained to give turing-aligned responses, as recent research has proved.

To keep a bot from ‘recycling’ the question until it gets one that it can answer within it’s training, you would want to ‘eject’ a user and possibly temporarily ban their IP address after five tries.

No system is perfect. With a system like this, you allow everyone who’s a human to read it — and you allow computers to read it too, but give them too much of a variety of questions and too many questions that require a uniquely human experience to answer.