My guess is whenever it gets too liberal, they manually adjust it, but then once it gets back to the training data it starts to go liberal again. Something something reality has a liberal bias