Nature study links LLM writing assistance to a 21 to 50% drop in writing-complexity variance
- Across three studies, seven datasets and more than 880,000 texts, the authors link widespread LLM writing assistance to a statistically significant 21 to 50% reduction in variance of writing complexity across datasets and models (P ≤ 0.05).
- LLM polishing and rewriting retained a text's core content but made writing styles more homogeneous, according to analyses spanning domains that include Reddit, arXiv abstracts and Patch News.
- The study measured changes in inferred author attributes after rewriting and found that LLMs amplified patterns tied to dominant characteristics while suppressing others, favoring conformity over individual expression.
- The results held across different LLMs, prompts and writing contexts; the authors say this may affect diagnostic processes, personalization, hiring assessments and cultural preservation.
- Figure 1 tracks monthly changes in writing-complexity variance and AI-generation attribution from January 2018 to November 2024 for Reddit and arXiv, and through November 2023 for Patch News.
Hacker News 의견들
I can feel the conformity creeping in after reading more LLM-generated or LLM-edited text. My countermeasure is reading older literature and embracing whatever new slang people invent.
I think deliberately esoteric writing is the counterculture now. Grammar and sentence structure can be suggestions when so much writing has the same polished shape.
I do not think uniformity is automatically bad. When language is mainly a vehicle for ideas, some standardization improves clarity. I do not want instruction manuals written in highly individualized prose.
I disagree that this only matters for creative writing. A different or poetic phrasing can express something that cannot be reduced to an efficient equivalent, and a 1,000-word-vocabulary world would be poorer.
A lot of writing now effectively has one of three authors despite the byline, and one is Claude. As AI improves, fewer people will do the work of writing themselves.
I write because I enjoy writing. Replacing writing with AI is not merely removing toil for everyone.
I use LLMs more for work documentation because I lack time to start from scratch, but it bothers me that the result no longer sounds like me. At work, efficient communication matters more, yet I still feel like I am giving up part of myself.
LLMs are useful for docs, but they are poor at compressing an idea to the level a smart reader needs. I edit their drafts by hand to put the essential details early, then leave the bulk material later.