Remember when there was that university study that pretended like it was big news that people could read sentences when all but the first and last letters of words were scrambled? You could probably combine that effect with Warp's idea to make something reasonably difficult to attack with a computer.
So, first, come up with 20 different ways of saying the instructions:
- Hello potential user, would you please write xyzzy in the captcha box?
- xyzzy is what you should definitely type into the following box if you want to register.
- Welcome, please type xyzzy in the box should you desire to join these forums.
- Etc.
Then pick one of the phrases at random, and scramble the letters of all the words (except for the code word) randomly: "Hlelo pteotianl uesr, wulod you pelase wtire xyzzy in the catphca box?" It might be even less susceptible to attack by a bot if the code word was nothing more than a scrambled real word, as well. Of course, this scheme breaks the "shouldn't have to understand English" rule, but oh well.
