As others have said, yes the server already knows what it means, or which images are correct. However it is based on confidence, that that's how neural networks work. It's a statistical model to determine the right answer. So even though its already confident that "pictures 1, 3, and 8 have a bicycle", when you say that those images have bikes, you are reinforcing and training it further.

https://www.geeksforgeeks.org/machine-learning/what-is-reinforcement-learning/

There's usually a couple of pictures that the captcha knows what it is, and some that it's uncertain about. So you have both verification ones to check you're actually doing it right and some used for training.

It was a lot more obvious with the book/text recaptcha where one word was a sully stylised one, and the other was scan from a book. You could usually put in whatever you wanted on the book text.

Text recognition? Heard about a google company called "Waymo"? Ever thought about what you do in these silly captchas and why they let you fail for no reason?