---
title: "The Cost of Perfect English — paper by Liu and Zhu (2026)"
description: "The paper asks what is lost when L2 English learners let large language models \"polish\" their essays. 24 references."
authors: ["Ao Liu", "Shanhua Zhu"]
year: 2026
venue: "arXiv:2605.13055"
source: "https://arxiv.org/abs/2605.13055"
doi: "10.48550/arxiv.2605.13055"
arxiv: "2605.13055"
licence: "arXiv — non-exclusive license to distribute"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/liu-2026-cost-perfect-english-pragmatic-flattening/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/liu-2026-cost-perfect-english-pragmatic-flattening/)

[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/45/) / The Cost of Perfect English: Pragmatic Flattening and the Erasure of Authorial Voice in L2 Writing Supported by GenAI

# The Cost of Perfect English: Pragmatic Flattening and the Erasure of Authorial Voice in L2 Writing Supported by GenAI

**Authors:** [Ao Liu](https://md.research.elamur.ai/free/en/authors/ao-liu/) and Shanhua Zhu · **Year:** 2026 · **Venue:** arXiv:2605.13055 · **Licence:** [arXiv — non-exclusive license to distribute](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2605.13055) · [PDF](https://arxiv.org/pdf/2605.13055) · [DOI 10.48550/arxiv.2605.13055](https://doi.org/10.48550/arxiv.2605.13055) · [arXiv 2605.13055](https://arxiv.org/abs/2605.13055) · [OpenAlex](https://openalex.org/W7161176703)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
The paper asks what is lost when L2 English learners let large language models "polish" their essays. The authors took 106 argumentative essays by B2-level Chinese students from ICNALE, covering Mainland China, Hong Kong and Taiwan and split evenly between two prompts. Each essay was rewritten by four commercial LLMs through their official APIs at temperature zero, giving 530 parallel texts including the originals. Using Brown and Levinson's Politeness Theory, they counted interpersonal metadiscourse such as engagement pronouns, hedges and boosters, and combined this with qualitative coding.

The models fixed grammar and kept the propositional content. However, they reshaped the interpersonal layer, a pattern the authors call the Semantic Preservation Paradox. Dialogic engagement collapsed in every model. Epistemic stance changed differently by model: some stripped hedges, while one added more of them. Variation between writers also shrank sharply. The authors read this as homogenization toward Anglo-American academic norms and call for Critical AI Literacy in writing classrooms.

## Key points
- Data: 106 ICNALE essays by CEFR B2 Chinese university students. The essays split evenly across the PTJ prompt (part-time jobs) and the SMK prompt (smoking ban). Each was rewritten by four LLMs, giving 530 texts in total.
- Engagement pronouns dropped from a human mean of 17.16 to between 1.92 and 4.55 after rewriting, a drop seen in every model.
- Epistemic stance diverged by model:
  - Gemini reduced both hedges and boosters.
  - DeepSeek v3.2 raised hedge frequency to 8.68, which the authors call "algorithmic caution".
- Diversity between writers collapsed. The standard deviation of engagement was 22.86 in the human texts and fell for all models, reaching 3.16 under Gemini.
- Acknowledged limitations:
  - The rewrite prompt pushed toward "elite academic discourse", which may confound the results.
  - The study lacks a neutral-prompt control.
  - It covers a single genre and a single learner population.

## Interesting ideas
- Build a parallel corpus of originals and model rewrites, then compare stance-marker counts per text, which separates "meaning kept" from "voice kept" as two distinct measurements.
- Use spread across authors, not just mean frequency, as a sign of homogenization: if rewritten texts cluster tightly, the system is imposing one norm.
- Treat the interpersonal layer as two separate dimensions, engagement versus epistemic stance, because a system can behave very differently on each.
- The authors' own self-critique says a directive rewrite prompt needs a neutral-prompt control group to show where the distortion actually comes from.
- Fixing generation settings (temperature zero, official APIs) makes model comparison reproducible.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (back): The text → graph → text compressor could be scored both on proposition recovery and on whether hedges, boosters and engagement pronouns survive the round trip.
- **I2** (learn): Evaluation of generated or silver outputs could track the per-text variance of stance features, flagging a model that drifts toward a single flat style.
- **I3** (frames): The prag crate could attach engagement markers (reader pronouns, rhetorical questions) and epistemic markers (hedges, boosters) to frames as separate, explicit stance annotations rather than discarding them.
- **I4** (learn): When LLM teacher tiers label or rewrite data, a neutral-instruction control run would serve as a negative control showing which biases come from the prompt rather than the teacher.
- **I5** (learn): Any LLM-teacher annotation pass should be run deterministically and logged, so that disagreements between witnesses reflect the models rather than sampling noise.

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

## Institutions

- [Southeast University](https://md.research.elamur.ai/free/en/universities/southeast-university/) — University, Dhaka, Bangladesh · 8 works · 160 citations

📚 Shelves (3): [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [📚 Literature and Literary Theory](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/literature-and-literary-theory/) · 519 works

🧭 Topics (1): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works

## Cite

Cite as: Ao Liu and Shanhua Zhu. 2026. The Cost of Perfect English: Pragmatic Flattening and the Erasure of Authorial Voice in L2 Writing Supported by GenAI. arXiv:2605.13055.

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

```bibtex
@misc{liu2026cost,
  title = {{The Cost of Perfect English: Pragmatic Flattening and the Erasure of Authorial Voice in L2 Writing Supported by GenAI}},
  author = {Ao Liu and Shanhua Zhu},
  year = {2026},
  doi = {10.48550/arxiv.2605.13055},
  eprint = {2605.13055},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2605.13055}
}
```

- [← previous The Conundrum of Trustworthy Research on Attacking Personally Identifiable Information Removal Techniques](https://md.research.elamur.ai/free/en/papers/ochs-2026-conundrum-trustworthy-research-attacking-personally/)
- [next → The Development of Spectral and Temporal Encodings in Speech Sounds](https://md.research.elamur.ai/free/en/papers/tan-2026-development-spectral-temporal-encodings-speech/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/45/)

## References

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

1. Al Hosni, J. (2025). Preserving authorial voice in academic texts in the age of generative AI: A thematic literature review. Arab World English Journal, 16(3), 244– 258. https://doi.org/10.24093/awej/vol16no3.14 [↗](https://doi.org/10.24093/awej/vol16no3.14)
2. Barattieri di San Pietro, C., et al. (2025). How inclusive large language models can be? The curious case of pragmatics. Frontiers in Education. https://doi.org/10.3389/feduc.2025.1619662 [↗](https://doi.org/10.3389/feduc.2025.1619662)
3. Brown, P., & Levinson, S. C. (1987). [Politeness](https://md.research.elamur.ai/free/en/papers/brown-1987-politeness-some-universals-language-usage/): Some universals in language usage. Cambridge University Press.
4. Chen, Y., et al. (2024). L2 pragmatic development in constructing and negotiating contextual meanings. Applied Linguistics. https://doi.org/10.1093/applij/amae046 [↗](https://doi.org/10.1093/applij/amae046)
5. Darvin, R. (2025). The need for critical digital literacies in generative AI-mediated L2 writing. Journal of Second Language Writing, 67, 101186. https://doi.org/10.1016/j.jslw.2025.101186 [↗](https://doi.org/10.1016/j.jslw.2025.101186)
6. Darvin, R., & Hafner, C. A. (2022). Digital literacies in TESOL: Mapping out the terrain. TESOL Quarterly, 56(3), 865–882. https://doi.org/10.1002/tesq.3161 [↗](https://doi.org/10.1002/tesq.3161)
7. Godwin-Jones, R. (2025). ChatGPT, second language pragmatics, and critical AI literacy. In Researching Generative AI in Applied Linguistics (pp. 274–286). Iowa State University Digital Press. https://doi.org/10.31274/isudp.2025.211.14 [↗](https://doi.org/10.31274/isudp.2025.211.14)
8. Hockly, N. (2023). Artificial intelligence in English language teaching: The good, the bad and the ugly. RELC Journal, 54(2), 445– 451. https://doi.org/10.1177/00336882231168504 [↗](https://doi.org/10.1177/00336882231168504)
9. Holliday, A. (2006). Native-speakerism. ELT Journal, 60(4), 385– 387. https://doi.org/10.1093/elt/ccl030 [↗](https://doi.org/10.1093/elt/ccl030)
10. Hyland, K. (2005). Metadiscourse: Exploring interaction in writing. Continuum. [find](https://scholar.google.com/scholar?q=Hyland%2C+K%2E+%282005%29%2E+Metadiscourse%3A+Exploring+interaction+in+writing%2E+Continuum%2E)
11. Ishikawa, S. (2013). The ICNALE and sophisticated contrastive interlanguage analysis of Asian learners of English. In S. Ishikawa (Ed.), Learner corpus studies in Asia and the world (Vol. 1, pp. 91–118). Kobe University. [↗](https://doi.org/10.24546/81006678)
12. Jiang, L., & Gu, M. M. (2022). Toward a professional development model for critical digital literacies in TESOL. TESOL Quarterly, 56(3), 1029– 1040. https://doi.org/10.1002/tesq.3138 [↗](https://doi.org/10.1002/tesq.3138)
13. Kecskés, I., & Dinh, H. (2025). ChatGPT for intercultural pragmatic learning? Potentially, but not yet. Intercultural Pragmatics, 22(2), 369– 398. https://doi.org/10.1515/ip-2025-2008 [↗](https://doi.org/10.1515/ip-2025-2008)
14. Kurt, G., & Kurt, Y. (2024). Enhancing L2 writing skills: ChatGPT as an automated feedback tool. Journal of Information Technology Education: Research, 23, 24. https://doi.org/10.28945/5370 [↗](https://doi.org/10.28945/5370)
15. Li, M., et al. (2025). The role of generative AI and hybrid feedback in improving L2 writing skills: A comparative study. Innovation in Language Learning and Teaching. https://doi.org/10.1080/17501229.2025.2503890 [↗](https://doi.org/10.1080/17501229.2025.2503890)
16. Mizumoto, A., & Eguchi, M. (2023). Exploring the potential of using an AI language model for automated essay scoring. Research Methods in Applied Linguistics, 2(2), 100050. https://doi.org/10.1016/j.rmal.2023.100050 [↗](https://doi.org/10.1016/j.rmal.2023.100050)
17. Mizumoto, A., et al. (2024). Testing the viability of ChatGPT as a companion in L2 writing accuracy assessment. Research Methods in Applied Linguistics, 3(2), 100116. https://doi.org/10.1016/j.rmal.2024.100116 [↗](https://doi.org/10.1016/j.rmal.2024.100116)
18. Moon, Y. (2025). The effects of ChatGPT on politeness in L2 university students' email requests to faculty. Modern English Education, 26(1), 352– 363. https://doi.org/10.18095/meeso.2025.26.1.352 [↗](https://doi.org/10.18095/meeso.2025.26.1.352)
19. Saldaña, J. (2021). The coding manual for qualitative researchers (4th ed.). SAGE Publications. [find](https://scholar.google.com/scholar?q=Salda%C3%B1a%2C+J%2E+%282021%29%2E+The+coding+manual+for+qualitative+researchers+%284th+ed%2E%29%2E+SAGE+Publications%2E)
20. Strobl, C., et al. (2025). Collaborative writing based on generative AI models: Revision and deliberation processes in German as a foreign language. Journal of Second Language Writing, 67, 101185. https://doi.org/10.1016/j.jslw.2025.101185 [↗](https://doi.org/10.1016/j.jslw.2025.101185)
21. Wang, J., et al. (2024). The impact of using ChatGPT on academic writing among medical undergraduates. Annals of Medicine, 56(1). https://doi.org/10.1080/07853890.2024.2426760 [↗](https://doi.org/10.1080/07853890.2024.2426760)
22. Yang, L., & Li, R. (2024). ChatGPT for L2 learning: Current status and implications. System, 124, 103351. https://doi.org/10.1016/j.system.2024.103351 [↗](https://doi.org/10.1016/j.system.2024.103351)
23. Zhao, C. G., & Wu, J. (2024). Voice and voicing strategies across native and second language writing: Extending the interactional metadiscourse framework. Applied Linguistics. https://doi.org/10.1093/applin/amae021 [↗](https://doi.org/10.1093/applin/amae021)
24. Zou, B., et al. (2025). Comparing the effects of ChatGPT and automated writing evaluation on students' writing and ideal L2 writing self. Innovation in Language Learning and Teaching. https://doi.org/10.1080/09588221.2025.2454541 [↗](https://doi.org/10.1080/09588221.2025.2454541)

*Updated 1 November 2026*