Besides, Google doesn't need help in picking out pictures of motorcycles or buses or bridges.
I expect the idea is the same with images, if there are 4 buses I expect 2 it knows are correct and used as a sanity test, 1 it's pretty sure it's correct (from majority of users selecting it), and 1 has some heuristics that might be a bus but it is currently testing to see whether users believe is or not. Similarly with the 5 not a bus images.
It doesn't need to know that your selections are "correct", only that they match responses given by most other users.
My intent was to disabuse the notion that captchas are how Google trains for image recognition.
https://httptoolkit.com/blog/apple-private-access-tokens-att...
You know, I hardly ever see those things.
They come up as part of a challenge-response when the server thinks your traffic is sus. I generally avoid them by staying logged in, not blocking cookies, and not using Tor or another route obfuscator.
(BTW, the "server" is very often a reverse proxy hosted by Cloudflare that decrypts your connection to the actual server by design. In all likelihood, there's more data mining done there than we're made aware of :)
Which sounds great until you leave your cozy little corner of the first world and discover that millions of people are behind CGNAT.