This is funny and relatable, but from my anecdotal experience it doesn't really matter whether you click boxes that contain a little corner of the object you're supposed to identify. I think those are weighted less, or the scoring mechanism isn't that simple/binary, or something like that.
It often doesn't even matter to click wrong boxes to a certain degree. Green or red points are usually accepted as traffic lights (often no need to click the whole traffic light) and anything white painted on a street is usually accepted as a crosswalk and so on. Skipping/verifying the second task without even bothering to look at it (i.e. immediately after clicking "next" on the first one) usually works well, too. The AI based on the ML at work here seems not to be very sophisticated.
These captchas are used to crowdsource training data for semantic segmentation ML models. By shifting the image around, users statistically fill out the boundaries of objects by selecting which squares include the object. As a result, in many captcha instances, you see objects right at rectangle boundaries.
