--- title: "Polished by a model, the styles drew closer together." url: https://deadwritingtheory.com/cases/the-shrinking-range/ summary: "Across 880,000+ texts, writing-complexity variance fell after ChatGPT, and LLM polishing cut it by 21 to 50%. The strongest evidence for convergence so far, mostly from rewrite tests." published: 2026-10-09 updated: 2026-10-09 author: "Jack Stovell" publisher: "Adapt Progress Evolve Limited" language: en-GB --- # Polished by a model, the styles drew closer together. Across 880,000+ texts, writing-complexity variance fell after ChatGPT, and LLM polishing cut it by 21 to 50%. The strongest evidence for convergence so far, mostly from rewrite tests. **Case No. DWT-010: The shrinking range.** Verdict for the prosecution. Filed 9 October 2026. - **The question:** Is the spread of LLM writing assistants linked to a decline in the diversity of how people write? - **The sample:** Over 880,000 texts in seven datasets. Monthly series from January 2018: 80,238 arXiv computer science abstracts, 318,490 Reddit creative stories (both to November 2024) and 379,583 Patch community news articles (to November 2023); plus 1,000 pre-ChatGPT Reddit and arXiv texts each, polished by GPT-3.5, Llama 3 70B and Gemini Pro - **The study:** Zhivar Sourati, Farzan Karimi-Malekabadi, Meltem Ozcan, Colin McDaniel, Alireza Ziabari, Jackson Trager, Ala Tak, Meng Chen, Fred Morstatter, Morteza Dehghani (University of Southern California), "The shrinking landscape of linguistic diversity in the age of large language models", Nature Human Behaviour, 2026. https://doi.org/10.1038/s41562-026-02550-0 - **Whose study:** Third-party study. Not our study. Entered into evidence from Zhivar Sourati, Farzan Karimi-Malekabadi, Meltem Ozcan, Colin McDaniel, Alireza Ziabari, Jackson Trager, Ala Tak, Meng Chen, Fred Morstatter, Morteza Dehghani, University of Southern California. - **Peer review:** Peer reviewed ## Findings of fact 1. The variance of writing complexity fell after ChatGPT's release in all three real-world datasets. 2. On Reddit and arXiv, AI use predicted later drops in variance; on Patch community news it did not, which the authors suggest editorial workflows may buffer. 3. LLM polishing kept the meaning (87% of similarity scores above 0.95) but reduced writing-complexity variance by a statistically significant 21 to 50% across datasets and models. 4. LLMs amplified patterns associated with dominant characteristics and suppressed others. ## Caveats for the jury - The 21 to 50% comes from rewrite experiments: what an LLM does to a text, not what writers actually published. - The real-world part is a correlation in time, and it uses an AI detector (Binoculars) to estimate AI use. - It measures one composite of writing complexity, not vocabulary or punctuation. ## The court reporter's account This is the big one, and I mean that in the dull sense: lots of texts, lots of datasets, lots of moving parts. Bear with me. The authors are Sourati, Karimi-Malekabadi, Ozcan, McDaniel, Ziabari, Trager, Tak, Chen, Morstatter and Dehghani, of the University of Southern California. The paper is "The shrinking landscape of linguistic diversity in the age of large language models", Nature Human Behaviour, 2026, peer reviewed. Three studies, seven datasets, over 880,000 texts. Study 1a tracked the variance of writing complexity month by month, January 2018 to November 2024. The material was 80,238 arXiv computer science abstracts, 318,490 Reddit creative stories and 379,583 Patch community news articles (Patch only to November 2023). The authors set that variance against the share flagged as AI by a detector called Binoculars. Study 1b did something different: GPT-3.5, Llama 3 70B and Gemini Pro each polished 1,000 pre-ChatGPT human texts from each of Reddit and arXiv, using neutral prompts. Findings first. The variance of writing complexity fell after ChatGPT's release in all three datasets. On Reddit and arXiv, AI use predicted later drops in variance. On Patch news it did not, and the authors suggest editorial workflows may buffer the effect. Hold that thought. In the polishing experiment, meaning was kept: 87% of similarity scores were above 0.95. But writing-complexity variance fell by a statistically significant 21 to 50% across datasets and models. So the machine leaves the message alone and quietly narrows the style. LLMs also amplified patterns associated with dominant characteristics and suppressed others, which is a worrying thing to find in a polishing tool. Now a contrast for the court. The Fitterer case (DWT-004) found no lexical flattening in edited news. This study finds the range of styles narrowing, but most clearly where no editor stands in the way. Those two don't quite collide. One counted words in edited news; the other measured a composite of complexity, and its weakest link to AI use was in the edited news. This is about convergence, whether writing became more alike, not just about how much text involved AI. That's the question the theory really turns on. Caveats, and they matter. The 21 to 50% comes from rewrite experiments, so it's what an LLM does to a text, not what writers actually published. The real-world part is a correlation in time, nothing stronger. And it measures one composite of complexity, not words or punctuation. So the court gets a strong finding with a clear boundary around it. The range is narrowing, in the places the study could see, and the cleanest numbers come from the lab rather than the wild. I'd call that persuasive, not conclusive. Verdict: for the prosecution. ## The exhibits entered in this case ### For the prosecution - **Exhibit P-13**: LLM polishing narrowed the range of writing styles. **21 to 50%** less variance in writing complexity after LLM rewriting. GPT-3.5, Llama 3 70B and Gemini Pro polished 1,000 pre-ChatGPT Reddit and arXiv texts each. Meaning was kept: 87% of similarity scores above 0.95. - **Exhibit P-14**: Style variance fell after ChatGPT in all three corpora. **3 / 3** datasets with lower writing-complexity variance after ChatGPT. 80,238 arXiv abstracts, 318,490 Reddit stories, 379,583 Patch news articles from 2018. AI use predicted later drops on Reddit and arXiv, not on Patch news. ### For the defence - No exhibit entered for the defence in this case. ## The finding Across 880,000+ texts, writing-complexity variance fell after ChatGPT, and LLM polishing cut it by 21 to 50%. The strongest evidence for convergence so far, mostly from rewrite tests. ## Related - [The fingerprints are there. The flattening is not.](https://deadwritingtheory.com/cases/the-news-that-didnt-flatten/) - [The stories grew alike. Each one got better.](https://deadwritingtheory.com/cases/better-stories-more-alike/) ## Sources - [Sourati et al., "The shrinking landscape of linguistic diversity in the age of large language models", Nature Human Behaviour, 2026 (the original study)](https://doi.org/10.1038/s41562-026-02550-0) - [Sourati et al., the authors' version on arXiv](https://arxiv.org/abs/2502.11266) ## Questions people ask ### Is AI making writing more alike? This study says yes, in part. Sourati et al. (USC, Nature Human Behaviour, 2026) found the variance of writing complexity fell after ChatGPT on Reddit, arXiv and Patch news, and LLM polishing cut it by 21 to 50%. ### Does this contradict the news study in case DWT-004? Not directly. Fitterer, Gangl and Ulbrich counted vocabulary in edited news and found no flattening. This study measures a composite of writing complexity, and its weakest link to AI use was in edited community news. ### What is the main weakness? The biggest number, 21 to 50%, comes from experiments in which LLMs rewrote texts. The real-world evidence is a correlation in time. ### Is this a ScriptGrain study? No. It is a third-party, peer-reviewed study, credited to its authors and linked to the original.