What I think is cool are the captchas that make fake words that actually look like they could be real words (as opposed to a random string of text). Makes it easier for a human to read and figure out, but no easier for a bot. I dont know how they do that.
Imagine picking letters with the right frequencies. Now, instead of doing that, pick pairs of letters, with the right frequency, so that each pair "chains" with the previous. If you have good pair frequency data, you can do longer than pairs and get even closer to English.
