hCaptcha engineer here: this is an issue with Google's approach, not all visual challenges.
Our challenges have average solve accuracy within 1% when comparing users in major African countries vs US users in our user testing. Nigeria does slightly better than the US, in fact.
American here. I had to say a mailbox was a parking meter yesterday to satisfy one of these terrible things. Your measurement of success may be biased by user acceptance of incorrect information in the process.
And as a non-American, I neither know nor care what American parking meters look like in the first place. They're training their model on my model of their model
Considering how horrifically broken hCaptcha is, slapping your hand on the keyboard would have an equivalent solve rate. I've had hCaptchas where one of the images was virtually identical to the example and it still didn't count as one.
Our challenges have average solve accuracy within 1% when comparing users in major African countries vs US users in our user testing. Nigeria does slightly better than the US, in fact.