I mean, the reason this works is generally because there is some information which Google does not have, that's why they are asking you to provide it. So, for i

I mean, the reason this works is generally because there is some information which Google does not have, that's why they are asking you to provide it. So, for instance, in the case of words, they used to present two words which were scanned in, one which they knew was correct and one which they didn't know. As long as you enter the one they know correctly, they assume you've also answered the one they don't know correctly. (Maybe they do some variant of this, but that was the general idea.) I assume they're doing something similar with the image segmentation tasks, so if you randomly select two tiles and click "next" there are some images where it actually doesn't know if you are correct or not, so it ends up storing your response and lets you through.