IIUC they mix known data in with unknown data.

If you get a lot of the known data wrong, they throw out your answer and make you answer the CAPTCHA again. If you get the known data right, they assume that your answers on the unknown data are correct and use that to label them, then start adding them in as the known data for other people. They might also show the same images to multiple people and see if they give the same answers - if so, it's probably correct and can be labeled, if not throw them out and make them answer the CAPTCHA again.

It mixes knowns and unknowns. It only judges you based on the knowns.

They probably require some minimum level of consensus to determine knowns as well.

For instance NYPL’s “building inspector” crowdsourcing program accepts a piece of data as verified if it’s been checked by 3 or more participants and at least 75% of them are in agreement. If 75% hasn’t been reached, it will keep showing it to more participants until it is.

When they still did the text captcha they asked for two words the first was known and the second unknown. You could opt out of their data collection by just giving the first word.

For the images I assume its something similar.

Its sampled together with other people. If all of them provided the wrong information, you could label something else as a traffic light or whatever.