A team fine-tunes BERT for a sentiment classification task and achieves strong results. They then try to use the same BERT checkpoint to generate product review text by feeding it a prompt and sampling tokens autoregressively. The generation is incoherent. Which architectural property of BERT directly causes this failure?