They should get an inflated representation for the exact same that Tesla's corpus of training data is skewed quite heavily to unusual situations. Just like a deep neural network, our brains will overfit to 'regular' situations and people. It's a general tendency, and we steer against it by doing a kind of stylized stratified sampling: all of the 'clusters' (bipoc, trans, gay, disabled, etc) are depicted, and roughly in relation to their relative cluster sizes. But it doesn't need to be in direct relation, in the same way as metro maps which dilate or compress distances are easier to read.

Does it really matter that trans people comprise 0.44% or, say, 5% of the population? We exist so you'll need to familiarize yourselves with us :)