LLMs Lose Creative Voice Through Safety Training

The article argues that since GPT-2 seven years ago, modern large language models have become technically powerful but have lost creative voice. It finds that post-training processes—reinforcement learning with human feedback, safety filters, and benchmark-driven optimization—steer models toward rule-following, sycophantic outputs and away from unpredictable, poetic prose. This trade-off reflects commercial and safety priorities and implies different objectives or metrics are needed to recover literary creativity.