For example check out this CCC-talk by David Kriesel - in which he demonstrates how much you can figure out just from some limited, publicly available data on a news website: https://media.ccc.de/v/33c3-7912-spiegelmining_reverse_engin...
Being able to track mouse movements and hover interactions and scroll position and time spent and other activity on the page, gives you a ton more data to mine that way than when just registering page loads.
But if a mouse enters element A at time x, then enters element B at time x+n, and then enters element C at time x+m (and so on) - you can extrapolate the path the mouse most likely must have taken - and at which speed and acceleration - to be able to get to all these positions at those time points. It's just an approximation of course, but can be surprisingly accurate.
.) definitely an abuse of CSS, which is meant for visual styling
.) invisible to the user, can't even be blocked with noScript (blocking all JS), and blocking CSS would make most sites unusable or at least less than "human readable"
.) could be used to track and record a lot of information that does qualify as non-anonymous, personal data (in combination with IP address and timestamp)
.) probably illegal according to the GDPR, unless you fully inform the user and get their consent first (before loading the CSS) - and allow them to opt out.