It has to be more sophisticated than that. Otherwise users could easily taint the datasets by giving wrong answers on purpose.
It probably checks your answer against the current model’s best guess and if it’s close enough, you get a pass and your input is added to the training data for the next iteration. The more wrong you are, the more challenges you get.
I vaguely remember 4chan figuring out something to do with which was the control and which the variable and deciding to spam solving the control correctly but the variable with some kind of nonsense (knowing 4chan probably a slur) until the system got enough confirmation that it got moved to the control group and would accept I it there
AFAIK, the first one is the real check, the second one is too train their image recognition AI.
It has to be more sophisticated than that. Otherwise users could easily taint the datasets by giving wrong answers on purpose.
It probably checks your answer against the current model’s best guess and if it’s close enough, you get a pass and your input is added to the training data for the next iteration. The more wrong you are, the more challenges you get.
That was in text captcha days
I vaguely remember 4chan figuring out something to do with which was the control and which the variable and deciding to spam solving the control correctly but the variable with some kind of nonsense (knowing 4chan probably a slur) until the system got enough confirmation that it got moved to the control group and would accept I it there