Elamur Research⌕
☰
home / papers / 2026 / The Style of Machines: A Stylometric Study of LLM Generation and Translation

The Style of Machines: A Stylometric Study of LLM Generation and Translation

Paper · 2026 · Natália Resende and Sheila Castilho

Summary

The study asks whether a large language model has its own consistent stylistic fingerprint, like those stylometry attributes to human authors and translators, and whether that fingerprint holds when the model moves from writing to translating and across languages. The authors built a small, balanced parallel corpus of twelve literary texts in English, Brazilian Portuguese and Spanish. Human writers and ChatGPT-4o each wrote a short story from the same narrative brief. Selected human stories were then translated both by human translators and by ChatGPT-4o, using a minimal zero-shot prompt. Five surface feature groups were computed for every text: lexical density, lexical richness (MATTR), vocabulary sophistication via Zipf frequency, mean sentence length, and punctuation ratios. In all three languages, GPT output was lexically richer and denser, used less frequent vocabulary, and favoured the em-dash. The same profile appeared in translations but was weaker than in original writing, which suggests the source text partly constrains the model's style. Although exploratory, the findings matter for AI-text detection and for viewing machine translators as active stylistic agents.

Key points

  • Corpus design: twelve text categories crossing language (EN, PT-Br, ES), task (authorship vs. translation) and agent (human vs. ChatGPT-4o). Texts run roughly 1,000–2,000 words each.
  • Translation directions:
    • The English human-authored story was the source for the PT-Br and ES translations.
    • The PT-Br human story was the source for the EN translations.
    • The authors justify this setup by English's dominance as a source language in literary translation.
  • Feature extraction:
    • MATTR used a 50-word sliding window.
    • Vocabulary sophistication was the mean Zipf frequency of content words, computed with the wordfreq library (v3.1.1).
    • Punctuation measures were per-token ratios of em-dashes, single quotation marks and ellipses.
  • Lexical density results: GPT originals exceeded human originals in every language (EN 0.549 vs. 0.474; PT 0.540 vs. 0.485; ES 0.495 vs. 0.466).
  • Lexical richness in translation: GPT translations still scored higher MATTR than human translations, though by a smaller margin (e.g., EN 0.745 vs. 0.724; PT 0.788 vs. 0.748).
  • Limitations:
    • A single model, with one author and translator voice per language.
    • No replication runs and no inferential statistics.
    • Only surface-level features were analysed.

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, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

Cite

Cite as Natália Resende and Sheila Castilho. 2026. The Style of Machines: A Stylometric Study of LLM Generation and Translation. Proceedings of the First Workshop on Style in GenAI-Translated Content (StyGenAI).

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

@inproceedings{resende2026style,
  title = {{The Style of Machines: A Stylometric Study of LLM Generation and Translation}},
  author = {Natália Resende and Sheila Castilho},
  booktitle = {Proceedings of the First Workshop on Style in GenAI-Translated Content (StyGenAI)},
  year = {2026},
  url = {https://aclanthology.org/2026.stygenai-1.5/}
}

References

6 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 2025. Delving into llm-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11(27):eadt3813. find
  2. [Baker1993] Baker, Mona. 1993. Corpus linguistics [Mauranen and Kujamäki2004] Mauranen, Anna and and translation studies: Implications and applica-Pekka Kujamäki. 2004. Translation universals: Do tions. In Baker, Mona, Gill Francis, and Elena they exist?, volume 48. John Benjamins Publishing. Togn…
  3. [Baker2000] Baker, Mona. 2000. Towards a methodol- [Saldanha2011] Saldanha, Gabriela. 2011. Translator style: Methodological considerations. The Transla-ogy for investigating the style of a literary translator. tor, 17(1):25–50. Target: International Journal of Translation Studies, 12(2):241–266. [S… find
  4. Bernardini. 2006. A new approach to the study of translationese: Machine-learning the difference between original and translated text. Literary and Lin- [Speer2022] Speer, Robyn. 2022. rspeer/wordfreq: v3.0, September. guistic Computing, 21(3):259–274. find
  5. [Burrows2002] Burrows, John F. 2002. ’delta’: a mea- [Stamatatos2009] Stamatatos, Efstathios. 2009. A survey of modern authorship attribution methods. Jour-sure of stylistic difference and a guide to likely au-nal of the American Society for Information Science thorship. Lit. Linguistic Comput., 17:… find
  6. Index translationum: [CSA Research2020] CSA Research. 2020. The state [UNESCO] UNESCO. World bibliography of translation. https: of the linguist supply chain: Translators and inter-//data.unesco.org/explore/dataset/ preters in 2020. Technical report, CSA Research. tran001/information/. Accessed: 05/… find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/resende-2026-style-machines-stylometric-study-llm/