No, the "weights" of your model are trained from the input data.
What is usually AB tested are hyperparameters of the model, or different "flavors" of (model+input data).
What people are implying is still unsubstantiated though. The engineers on the Twitter Space say that this is to ensure that changes they make do not bias one category over another, they don't say that it's in order that they can make discretionary updates to bias towards Elon Musk.
Maybe after every update to the model, they check these stats to ensure that they haven't biased towards Elon Musk, and if so roll the change back.
