There's also the issue that it will lie to you if the alogrithm decides it simply doesn't like you. Which means you'll end up doing at least a couple of rounds before it decides to let you through.
I always envisioned their devious model to be something like:
- You want to train on an unlabeled dataset, label it along the way.
- You have a set of untrusted validators, some with no history, some with known credibility and accuracy scores. And you have a lot of them.
- You do kind of a zero-knowledge proof by showing the unlabeled dataset to validators that you know you can trust because of their historical high success rate, which you've already established through asking them to label a dataset that you already have high confidence on.
Kind of like how a blue-green colorblind person could find out which pen is blue, which pen is green if he is surrounded by people he can't fully trust. Ask people around you and maybe even show the same person the same pen (or a really dead-easy captcha) twice in a row. If they lie to you both times, they are not to be trusted.