AI Writing Quality Suffers From Training Bias, Not Speed Optimizations

A developer who built a 36-pattern checklist to detect AI writing flaws examined a viral theory claiming inference-time speed tricks degrade creative writing quality. The claim centered on speculative decoding, but the author argues this technique is mathematically lossless and does not affect output quality — only how quickly responses are delivered. The real culprit, supported by research from Kirk et al., is RLHF, a training process where models are rewarded by human raters who favor safe, skimmable responses, causing output diversity to narrow over time. This phenomenon, known as mode collapse, means AI models converge on a bland, averaged register rather than producing varied or distinctive writing. Popular workarounds like generating multiple drafts and merging the best parts fail to solve the problem because all drafts originate from the same narrowed model.