What is a major danger of pseudo-labeling?

A. It cannot use any labeled data B. It always reduces the sample size C. It makes all classes equal D. Early prediction errors can reinforce themselves during retraining