Which of these is a colour: monkey, bank or purple?

Nothing mattered other than that the last word was purple.

Change “a colour” to “an animal” and it will still give the same exact result about the color purple.

So far the computational knowledge engine is a failure. But, it will improve as it gets more knowledge and relationships.

I think you are right that CAPTCHA’s are not meant as an absolute means for protecting an application, but a way to slow down spammers and whomever else. I like the idea of text CAPTCHA’s making it easier for people to decipher and from an academic standpoint, playing more with sentence structure for increased strength could be interesting. However, the effort that may go into that might not be practical for your cause? Like you said, CAPTCHA’s are meant as an obstacle, not anything more.

180 million questions is impressive! I look forward to seeing how the methodology for creating questions evolves. Good luck!

Rendering the questions as an image is not very helpful — this negates the reason to use logic at all (why not a word), you need to provide an audio alternative and is not much of an obstacle as OCR is easy.

I’ll have to have a think about the question construction, but it is a delicate line between making questions that are simply too confusing (or take too long) for everybody to actually understand and solve. I have thought about trying to grade question strength — and allowing people to specify a question strength when they ask for questions: stronger questions would be harder to break but probably more difficult for real users to understand, and the decision would depend on your audience.

I posted a comment on the Hacker News thread about that. Questions will come in certain sentence structures that become predictable. I think that, if text CAPTCHA’s are to be harder for a bot to guess, the subject matter has to be much more abstract. For instance, instead of asking “what is the second color in the sequence blue, apple, banana, yellow, green, orange?” (which can easily be learned and guessed by a bot) you need to ask something along the lines of “how many people will have warm hands if you hand out 3 hats, 5 pairs of gloves, and 2 scarves?”…. Or at least, maybe that gets closer?

Maybe a list of requirements/criteria need to be established against which each text CAPTCHA question is vetted. Then again, it’s just a CAPTCHA, so does it have to work 100%?

There’s another very important criterion for CAPTCHAs: they must be easy to generate computationally. If they’re not easy to generate, then you get a small number of questions that the bot can cache.

It seems to me that a big limitation of text CAPTCHA is the limited range of query templates. You can write parsers for all of those templates pretty easily, defeating the system. Even if you only get 20% CAPTCHA hit rates, that just reduces the number of accounts you can get per IP address in a given amount of time. But you’re still “in”m fundamentally.