---
title: "Stylometric Approach to AI-generated Texts. — Pawłowski 2026"
description: "This study asks whether classic stylometry can tell human literary prose apart from LLM imitations, and whether models reproduce individual authors'…"
authors: ["Adam Pawłowski", "Tomasz Walkowiak"]
year: 2026
venue: "Proceedings of the 10th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature 2026"
source: "https://aclanthology.org/2026.latechclfl-1.21/"
doi: "10.18653/v1/2026.latechclfl-1.21"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/pawlowski-2026-stylometric-approach-ai-generated-texts/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/pawlowski-2026-stylometric-approach-ai-generated-texts/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/9/) / Stylometric Approach to AI-generated Texts. An Analysis of Contemporary French-Language Literature

# Stylometric Approach to AI-generated Texts. An Analysis of Contemporary French-Language Literature

**Authors:** Adam Pawłowski, Tomasz Walkowiak · **Year:** 2026 · **Venue:** Proceedings of the 10th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature 2026 · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.latechclfl-1.21/) · [PDF](https://aclanthology.org/2026.latechclfl-1.21.pdf) · [DOI 10.18653/v1/2026.latechclfl-1.21](https://doi.org/10.18653/v1/2026.latechclfl-1.21) · [OpenAlex](https://openalex.org/W7140118011)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14)

## Summary
This study asks whether classic stylometry can tell human literary prose apart from LLM imitations, and whether models reproduce individual authors' styles. The authors built a paired corpus from twentieth-century French-language fiction by European and African writers: 744 human samples of about 15,000–20,000 characters each. Each of five LLMs first condensed every sample into a 500-word summary. It then wrote a new story from that summary, prompted to imitate the named author's style. Texts were represented as TF-IDF-weighted most-frequent-word vectors, visualized with PaCMAP, and classified with k-NN for three tasks: detecting human text, identifying the source (human vs. each model), and attributing authorship. Human texts separated clearly from machine texts despite the rich prompts. Each LLM also formed its own clusters, which suggests model-specific "idiolects." Texts by the same human author were more alike than texts a model produced for that author, so LLM output looks like one collective voice imitating many writers. The work matters for AI-text detection and for literary studies of machine creativity.

## Key points
- The final corpus has 3,999 samples: 744 human-authored and 3,255 machine-generated, short of the 3,720 expected because some API calls failed. The models were GPT-4o, GPT-4o-mini, DeepSeek-V3, c4ai-command-r-plus and c4ai-command-a.
- The method uses TF-IDF on the most frequent words, PaCMAP projection, and a k-NN classifier (k = 5). Similarity was measured as the mean distance from each text to its nearest neighbour with the same author and source.
- Authorship attribution worked well for human authors even with few features. For AI-generated texts it became reliable only at 1,000 or more features.
- Human texts had the lowest nearest-neighbour distances at every feature count, for example 0.366 at 500 MFW and 0.568 at 5,000 MFW. DeepSeek-V3 had the highest, at 0.486 and 0.748. GPT-4o and GPT-4o-mini outputs clustered close together.
- Many first drafts were too short, so the authors added follow-up prompts asking for longer texts. Because the models treated earlier output as part of the prompt, this broke the ceteris paribus condition, and each model's texts split into two clusters.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) · [History & humanities](https://md.research.elamur.ai/free/en/shelves/history/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{pawlowski2026stylometric,
  title = {{Stylometric Approach to AI-generated Texts. An Analysis of Contemporary French-Language Literature}},
  author = {Pawłowski, Adam and Walkowiak, Tomasz},
  booktitle = {Proceedings of the 10th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature 2026},
  year = {2026},
  doi = {10.18653/v1/2026.latechclfl-1.21},
  url = {https://doi.org/10.18653/v1/2026.latechclfl-1.21}
}
```

- [← previous Style Transfer as Bias Mitigation: Diffusion Models for Synthetic Mental Health Text for Arabic](https://md.research.elamur.ai/free/en/papers/mankarious-2026-style-transfer-bias-mitigation-diffusion/)
- [next → Subject to change: Agreement and case](https://md.research.elamur.ai/free/en/papers/lingbulk-2026-rimon-subject-change-agreement-case/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/9/)

## References

12 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Maciej Eder, Jan Rybicki, and Mike Kestemont. 2015. Stylometry with r: A package for computational text analysis. The R Journal, 8:107–121. [↗](https://doi.org/10.32614/rj-2016-007)
2. Ismail Kassou Sara El Manar El Bouanani. 2014. Authorship analysis studies: A survey. International Journal of Computer Applications, 86(12):22–29. [↗](https://doi.org/10.5120/15038-3384)
3. Efstathios Stamatatos. 2009. A survey of modern authorship attribution methods. J. Am. Soc. Inf. Sci. Technol., 60(3):538–556. [↗](https://doi.org/10.1002/asi.21001)
4. Yingfan Wang, Haiyang Huang, Cynthia Rudin, and Yaron Shaposhnik. 2021. Understanding how dimension reduction tools work: An empirical approach to deciphering t-sne, umap, trimap, and pacmap for data visualization. Journal of Machine Learning Research, 22(201):1–73. [↗](https://doi.org/10.48550/arxiv.2012.04456)
5. Patrick Juola. 2006. Authorship attribution. Found. Trends Inf. Retr., 1(3):233–334. [find](https://scholar.google.com/scholar?q=Patrick+Juola%2E+2006%2E+Authorship+attribution%2E+Found%2E+Trends+Inf%2E+Retr%2E%2C+1%283%29%3A233%E2%80%93334%2E)
6. Mike Kestemont. 2014. Function words in authorship attribution. from black magic to theory? In Proceedings of the 3rd Workshop on Computational Linguistics for Literature (CLFL), pages 59–66, Gothenburg, Sweden. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Mike+Kestemont%2E+2014%2E+Function+words+in+authorship+attribution%2E+from+black+magic+to+theory%3F+In+Proceedings+of+the+3rd+Workshop+on+Computational+Linguistics+for+Literature+%28CLFL%29%2C+pages+59%E2%80%9366%2C+Gothenbu)
7. Frederick Mosteller and David Wallace. 1964. Inference and disputed authorship : The Federalist. Addison-Wesley series in behavioral science : quantitative methods. Addison-Wesley, Reading Mass. [etc. [find](https://scholar.google.com/scholar?q=Frederick+Mosteller+and+David+Wallace%2E+1964%2E+Inference+and+disputed+authorship+%3A+The+Federalist%2E+Addison%2DWesley+series+in+behavioral+science+%3A+quantitative+methods%2E+Addison%2DWesley%2C+Reading+Mass%2E+%5Betc%2E)
8. James O’Sullivan. 2025. Stylometric comparisons of human versus ai-generated creative writing. Humanities and Social Sciences Communications, 12. [↗](https://doi.org/10.1057/s41599-025-05986-3)
9. Adam Pawłowski, Ewa Kalinowska, and Tomasz Walkowiak. 2025. Au-delà de l’attribution d’auteur : la stylométrie permet-elle d’identifier l’« identité collective » des textes littéraires et le sexe des auteurs ? analyse comparée de la fiction narrative d’europe et d’afrique. Romanica Cracoviensia, 25:203. [↗](https://doi.org/10.4467/20843917rc.25.037.22750)
10. Karol Przystalski, Jan K. Argasiński, Iwona Grabska-Gradzińska, and Jeremi K. Ochab. 2026. Stylometry recognizes human and llm-generated texts in short samples. Expert Systems with Applications, 296:129001. [↗](https://doi.org/10.1016/j.eswa.2025.129001)
11. Simon Päpcke, Thomas Weitin, Katharina Herget, Anastasia Glawion, and Ulrik Brandes. 2022. Stylometric similarity in literary corpora: Non-authorship clustering and Deutscher Novellenschatz. Digital Scholarship in the Humanities. Fqac039. [↗](https://doi.org/10.1093/llc/fqac039)
12. Irina Rabaev, Marina Litvak, Vladimir Younkin, Ricardo Campos, Alípio Mário Jorge, and Adam Jatowt. 2023. The competition on automatic classification of literary epochs. In Proceedings of the IACT - The 1st International Workshop on Implicit Author Characterization from Texts for Search and Retrieval held in conjunction with the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2023), Taipei, Taiwan, July 27, 2023, volume 3477 of CEUR Workshop Proceedings, pages 49–56. CEUR-WS.org. [find](https://scholar.google.com/scholar?q=Irina+Rabaev%2C+Marina+Litvak%2C+Vladimir+Younkin%2C+Ricardo+Campos%2C+Al%C3%ADpio+M%C3%A1rio+Jorge%2C+and+Adam+Jatowt%2E+2023%2E+The+competition+on+automatic+classification+of+literary+epochs%2E+In+Proceedings+of+the+IACT+%2D+T)