Alignment pretraining: AI discourse creates self-fulfilling (mis)alignment