**Original paper:** [https://arxiv.org/abs/2608.12630v1](https://arxiv.org/abs/2608.12630v1)

# Summary: 2026-08-12_22-32-39Z_Novelsgeneratedbylanguagemodelsshowcompressedforma.md
Saved: 2026-08-13 22:31
Source: 2026-08-12_22-32-39Z_Novelsgeneratedbylanguagemodelsshowcompressedforma.md
Model: None

---

## Summary  
The paper investigates whether large language models can sustain the full range of formal variation that characterizes human‑written novels, rather than merely producing text that is detectable as AI‑generated. By comparing six corpora—AI‑generated and human‑written novels across different styles—the authors demonstrate that repeated model outputs compress formal features such as sentence structure, readability, punctuation density, and lexical diversity. Their contribution is a nuanced distinction between two types of overclosure: variance overclosure (limited overall range) and correlational overclosure (individual novels resemble humans while the collection does not). This work advances the understanding of AI‑generated literary style and its implications for authenticity assessment.

## Key Contributions  
- [Finding 1] Repeated generations from GPT‑5.5 Thinking in a nineteenth‑century British realist style, Qwen3‑14B in that same style, and their zero‑style variants produce novels with markedly reduced sentence‑structure variation compared to human counterparts.  
- [Finding 2] Across all measures (MATTR‑500, Shannon entropy, average sentence length, readability scores, punctuation rate) the AI corpora exhibit compression, indicating a narrower formal range than observed in human literature.  
- [Finding 3] While lexical diversity is similarly compressed, Qwen Zero‑Style MATTR shows an exception, suggesting that certain model configurations retain more lexical freedom.

## Methodology  
The authors constructed six corpora: (1) twenty novels generated by GPT‑5.5 Thinking in a nineteenth‑century British realist style; (2) twenty novels generated by Qwen3‑14B in the same historical style; (3) twenty novels generated by each model in a contemporary zero style; (4) 205 nineteenth‑century human‑written British novels; and (5) sixty‑five contemporary human‑written zero‑style novels. For each corpus they computed six quantitative metrics: MATTR‑500 (a stylometric similarity index), Shannon entropy, average sentence length, readability (Flesch‑Kincaid), punctuation rate, and lexical diversity. These measures were aggregated at the document level to capture overall formal variation.

## Results  
All AI‑generated corpora showed lower Shannon entropy and higher average sentence length than human corpora, reflecting compressed syntactic patterns. Punctuation rates were uniformly reduced, indicating fewer complex sentences. MATTR‑500 values clustered tightly within each model’s style group, confirming variance overclosure. Lexical diversity remained low except for Qwen Zero‑Style MATTR, which exhibited a modest increase. No strong cross‑measure correlation emerged between GPT and Qwen outputs, underscoring the lack of a stable pattern across models.

## Significance  
These findings reveal that AI language models, despite producing individually coherent novels, generate corpora with dramatically less formal variation than human authorship. This matters for authenticity detection, literary criticism, and any application where stylistic diversity is a key quality metric. The distinction between variance overclosure and correlational overclosure provides a conceptual framework for interpreting AI‑generated content beyond simple “AI vs. human” binaries.

## Related Concepts  
- Formal variation in literature (sentence structure, readability, punctuation)  
- Overclosure: the phenomenon where generated outputs occupy a narrower range than the source domain  
- Variance overclosure vs. correlational overclosure  
- Stylometric metrics (MATTR‑500, Shannon entropy)  
- AI language model generation pipelines (GPT‑5.5 Thinking, Qwen3‑14B)