Why AI Writing Feels Bland: It's Training, Not Speed Tricks

A viral theory blamed speculative decoding for degrading AI writing quality, but the technique is mathematically lossless. The real culprit is RLHF training, which narrows model output toward safe, skimmable prose at the cost of diversity. A 2023 study measured this mode collapse directly, showing RLHF models produce significantly less varied responses.