LLMs could write like humans but post-training guardrails make their text detectable

Research by Pangram CTO Bradley Emi suggests that the uniform writing style of modern LLMs is a result of post-training safety guardrails rather than a limitation of the underlying models. Comparisons indicate that base models exhibit significantly more stylistic variety before these constraints are applied. This finding highlights how alignment processes intended to improve safety and consistency may inadvertently reduce the expressive range and human-like nuance of generated text.