Every new data point partitions the statistical population size into two smaller parts, each with their own larger variance (ie the insurance company's risk). The insurance company could statistically combine the risk from each partition into the same original, but it's more likely they'll focus on the higher independent risk figures and raise premiums an outsized amount to cover each individually. And this effect is going to be more pronounced the more lopsided the partition is, leading to similar monoculture incentives as we see in the mortgage/housing bubble. For example, I'd bet there's somewhat of a correlation between insurance claims and whether a house is painted beige.
That is how rates can go up even when the extra data is fundamentally sound. But there can also be just enough extra information to be damning, but not enough to exonerate. For example after the "Do you have a trampoline" question, is there a follow up of "Do you let guests use it" ? Or perhaps a more formal "opt out of all liability coverage for the trampoline" ? Likely not.
Then of course there are places where the model is irrelevant or even outright wrong, because the thing being singled out seems like it rocks the boat. Like the driver surveillance devices that penalize focused acceleration due to perceived association with racing, when it's much more likely that the driver is actually paying attention to driving. Or penalizing people for going over the posted speed limit, when it's actually safer to go the prevailing speed of the road.
