#rlhf

Why AI Models Are Being Deliberately Dumbered Down

Large language models undergo safety training that inadvertently impairs their reasoning, a phenomenon known as the alignment tax. This trade-off often causes systems to reject harmless queries that share superficial linguistic patterns with prohibited content.