Amplified by surreptitious cross-training, I assume.

Lies can get in training data before the truth has finished booting.