Comment by int_19h

This is a prime indication that the bit about "white genocide" comes from the prompt. The model itself knows that it is bullshit from its training, though, and with a smart enough model, when there's a contradiction between the prompt and the training, the latter will generally win in the long run.