Quantifying the Risks of LLM- and Tool-assisted Rephrasing to Linguistic Diversity

Mengying Wang, Andreas SpitzEMNLP 2025

TL;DR

This paper investigates whether writing assistants (WATs) and Large Language Models (LLMs) reduce linguistic diversity in text when used at scale. The study finds that LLMs significantly alter text semantics and vocabulary, potentially leading to a "shrinkage" of linguistic diversity, while traditional WATs have a less pronounced effect.

What Problem Does This Paper Address?

As writing assistants and LLMs become ubiquitous tools for text creation, there's a growing concern about their potential impact on the richness and diversity of language itself. While previous research has focused on the effectiveness and accuracy of these tools for individual users, little is known about the aggregate effect of their widespread adoption on the language we use daily. This paper aims to quantify this risk by measuring semantic and vocabulary changes introduced by rephrasing tools across a diverse corpus of human-generated text.

Methods

The researchers employed a quantitative approach to measure the impact of rephrasing tools. They used a combination of word-based and vector-based metrics to analyze changes in texts before and after rephrasing.

  • Word-based measures:
    • Paragraph length: Measured by the change in word count.
  • Jaccard similarity: Quantifies vocabulary overlap between original and rephrased texts.
  • Levenshtein distance: A fine-grained measure of text difference at the character level.
  • Vocabulary Size: The number of unique words in a corpus of rephrased texts.
  • Vector-based measures:
    • Semantic similarity: Calculated using cosine similarity between sentence embeddings (from Sentence-BERT and ALBERT models) of original and rephrased paragraphs.
  • Conicity: Measures the dispersion of token vectors in latent embedding space, indicating semantic variation.

Datasets

The study compiled a corpus of 819 English texts written before 2010 to avoid pre-existing AI influence. These texts were drawn from eight diverse domains: literature, academic papers, encyclopedic texts, instruction manuals, news, social media posts, interview transcripts, and speeches. The data is available in a public code repository.

Key Insights and Ideas

  • LLMs are not drop-in replacements for WATs: LLMs induce significantly more semantic and vocabulary changes than traditional writing assistants, warranting caution in their indiscriminate use.
  • LLMs tend to summarize: LLMs consistently shorten texts, even with neutral prompts, suggesting a potential bias towards summarization learned during their training.
  • Vocabulary shrinkage is a concern: Both WATs and LLMs reduce lexical diversity, with LLMs showing a much more pronounced effect. This suggests a potential shift towards the tools' internal default vocabularies, leading to a less varied linguistic landscape.
  • Domain matters, but LLMs are broadly impactful: While some domain-specific variations exist, LLMs generally cause stronger divergence across most text types.
  • Commercial LLMs show more stability: Commercial LLMs (GPT-40 mini, Gemini) exhibited more consistent behavior across different prompts and domains compared to open-weight models, which showed greater variability.

Key Results

  • Paragraph Length: On average, WATs had minimal impact on paragraph length, with Wordtune shortening texts by 9.3% and Rephrase extending them by 8.6%. LLMs, however, consistently shortened texts, with Aya-23 reducing length by 24.8%.
  • Vocabulary Size: WATs slightly decreased vocabulary size (except Rephrase), while LLMs also consistently reduced it, with some open-weight LLMs causing reductions of up to 29%. Commercial LLMs like Gemini and DeepSeek, however, increased vocabulary size.
  • Vocabulary Overlap (Jaccard Similarity): Most tools significantly altered vocabulary. Grammarly and Wordtune retained higher overlap (0.90 and 0.71 respectively), while other tools had scores below 0.6. LLMs generally showed more pronounced vocabulary change than WATs.
  • Semantic Similarity: LLMs consistently showed stronger semantic divergence from the original text than WATs, with the exception of Rephrase.
  • Conicity: LLMs generally increased the semantic spread of texts, indicating less similarity among outputs, while WATs caused close to no change.

Limitations and Future Work

The authors acknowledge several limitations:

  • Prompting: The study primarily used neutral prompts for LLMs and did not explore the impact of style-specific prompts (e.g., simplification, domain adaptation), which could lead to different outcomes.
  • Corpus Scope: The corpus, while diverse, is not exhaustive. Future work should expand to more domains and languages.
  • Experimental Scale: The manual rephrasing for WATs limited the corpus size. A larger scale study, especially for LLMs, is needed to fully quantify vocabulary shift.
  • Qualitative Analysis: The study focused on quantitative measures of change, not the qualitative nature of those changes. A linguistic analysis of the styles of change would be valuable.
  • WAT Competence Assumption: The study assumed WATs only suggested necessary changes. Proactive suggestions from WATs could influence vocabulary change.

Future work could explore mitigating vocabulary loss through prompt engineering or explicit lexical constraints for LLMs. Investigating the impact of domain experts using WATs and expanding the study to non-English languages are also suggested.

Summary updated

BibTeX
@inproceedings{wang2025quantifying,
  author = {Mengying Wang and Andreas Spitz},
  title = {{Quantifying the Risks of LLM- and Tool-assisted Rephrasing to Linguistic Diversity}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure