For sufficiently motivated adversaries, even a custom UI isn't enough of a roadblock.
Is it really realistic to expect advanced attacks on CAPTCHAs with visual recognition or Mechanical Turk? I think you'd have to be a very high value target. I wrote a little PHP captcha when I was 15 and I've been using it ever since, and I never got any spam on my (admittedly very low-traffic) sites.
I actually wanted to implement visual perturbations based on perlin noise, but I really never got around to doing that. Would be an exciting little project.
Also, if a site asks me to use someone else's credentials to comment, I probably won't bother. Why should my hobby-coding account on Github, or my spam-catching account on Gmail, be linked to your blog about philately or bird-watching?