I’m convinced that what’s measured isn’t the accuracy of what is picked, but the delay in picking. A bot will immediately start processing the image and selecting square after square at a certain rate, while the human has that pause of “wtf?”, then multiple hesitations.

You pass the test by clicking what the majority would click, not what’s correct.
The Captcha doesn’t know what’s correct, you are used as unpaid labor to train the AI.

I’ve been asked this, except it was a scooter in the picture. I correctly didn’t select any squares, but failed the test.

Scooters aren’t motorcycles!

I have a motorcycle and I still failed a motorcycle captcha.