Tag
This paper investigates how the growing use of LLMs in writing is altering scientific communication, using a corpus of ACL papers and synthetic data to show lexical and stylistic changes, and connects these to subjective reading experience via expert annotations.
An analysis arguing that the optimal balance for AI-assisted writing is around 50% AI and 50% human input, where AI handles structure and organization while humans provide voice, judgment, and editorial control. The author contends that 100% AI reads as slop while 0% AI leaves capability on the table, and that meaningful AI assistance requires genuine expertise, strong structure, and distinctive human voice.