Register alignment of ChatGPT-generated academic texts

IF 2.1

Applied Corpus Linguistics Pub Date : 2026-04-01 Epub Date: 2025-12-02 DOI:10.1016/j.acorp.2025.100174

Nur Yağmur Demir, Jesse Egbert

{"title":"Register alignment of ChatGPT-generated academic texts","authors":"Nur Yağmur Demir, Jesse Egbert","doi":"10.1016/j.acorp.2025.100174","DOIUrl":null,"url":null,"abstract":"<div><div>The rise of Artificial Intelligence (AI) tools such as ChatGPT has transformed language pedagogy and assessment. Despite their growing use in academic contexts—from classroom materials to standardized testing—questions remain about the register appropriateness of the texts they produce.</div><div>The humanlikeness of AI language must be defined not only by fluency or coherence, but by register appropriateness—functional language use that aligns with the situational characteristics of registers. This study investigates whether ChatGPT-generated academic texts mimic human-authored writing in two academic genres (journal articles and textbooks) across two disciplines (biology and history).</div><div>Using multi-dimensional analysis, we analyzed 200 texts (100 AI-generated and 100 human-authored) along three linguistic dimensions: (1) specialized information density vs. non-technical synthesis, (2) definition/evaluation of new concepts, and (3) author-centered stance. Our results reveal a mixed picture: while ChatGPT exhibits moderate success in mimicking register distinctions found in journal article registers, its performance is notably less aligned with textbooks. ChatGPT-generated textbook excerpts in biology, for instance, often resemble the dense, technical style of journal articles, as a result failing to match the simplified, pedagogically oriented discourse found in human-authored textbooks.</div><div>Our findings indicate that while ChatGPT can largely reproduce human-like register patterns in journal article writing, it struggles to achieve the same in textbook contexts, particularly within biology. Overall, the results suggest that ChatGPT-generated texts often lack sufficient functional appropriateness. We therefore recommend further quantitative linguistic analyses of AI-generated language and urge caution when using ChatGPT for content creation.</div></div>","PeriodicalId":72254,"journal":{"name":"Applied Corpus Linguistics","volume":"6 1","pages":"Article 100174"},"PeriodicalIF":2.1000,"publicationDate":"2026-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Applied Corpus Linguistics","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2666799125000565","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"2025/12/2 0:00:00","PubModel":"Epub","JCR":"","JCRName":"","Score":null,"Total":0}

引用次数: 0

Abstract

The rise of Artificial Intelligence (AI) tools such as ChatGPT has transformed language pedagogy and assessment. Despite their growing use in academic contexts—from classroom materials to standardized testing—questions remain about the register appropriateness of the texts they produce.

The humanlikeness of AI language must be defined not only by fluency or coherence, but by register appropriateness—functional language use that aligns with the situational characteristics of registers. This study investigates whether ChatGPT-generated academic texts mimic human-authored writing in two academic genres (journal articles and textbooks) across two disciplines (biology and history).

Using multi-dimensional analysis, we analyzed 200 texts (100 AI-generated and 100 human-authored) along three linguistic dimensions: (1) specialized information density vs. non-technical synthesis, (2) definition/evaluation of new concepts, and (3) author-centered stance. Our results reveal a mixed picture: while ChatGPT exhibits moderate success in mimicking register distinctions found in journal article registers, its performance is notably less aligned with textbooks. ChatGPT-generated textbook excerpts in biology, for instance, often resemble the dense, technical style of journal articles, as a result failing to match the simplified, pedagogically oriented discourse found in human-authored textbooks.

Our findings indicate that while ChatGPT can largely reproduce human-like register patterns in journal article writing, it struggles to achieve the same in textbook contexts, particularly within biology. Overall, the results suggest that ChatGPT-generated texts often lack sufficient functional appropriateness. We therefore recommend further quantitative linguistic analyses of AI-generated language and urge caution when using ChatGPT for content creation.

查看原文本刊更多论文

chatgpt生成的学术文本的寄存器对齐

ChatGPT等人工智能（AI）工具的兴起改变了语言教学和评估。尽管它们在学术环境中越来越多地使用——从课堂材料到标准化测试——但它们所产生的文本的注册适当性问题仍然存在。人工智能语言的人类相似性不仅必须通过流利或连贯来定义，还必须通过语域的适当性来定义，即与语域的情境特征相一致的功能性语言使用。这项研究调查了chatgpt生成的学术文本是否模仿了两种学术类型（期刊文章和教科书）中跨越两门学科（生物学和历史学）的人类写作。使用多维分析，我们沿着三个语言学维度分析了200个文本（100个人工智能生成和100个人类撰写）：(1)专业信息密度与非技术合成，(2)新概念的定义/评估，以及(3)以作者为中心的立场。我们的研究结果揭示了一个复杂的情况：虽然ChatGPT在模仿期刊文章寄存器中发现的寄存器差异方面表现出适度的成功，但它的表现与教科书的表现明显不一致。例如，chatgpt生成的生物学教科书节选通常类似于期刊文章的密集、技术风格，因此无法与人类撰写的教科书中简化的、以教学为导向的论述相匹配。我们的研究结果表明，虽然ChatGPT可以在期刊文章写作中很大程度上重现人类的语域模式，但在教科书环境中，尤其是在生物学领域，它很难达到同样的效果。总的来说，结果表明，chatgpt生成的文本往往缺乏足够的功能适当性。因此，我们建议对人工智能生成的语言进行进一步的定量语言分析，并敦促在使用ChatGPT进行内容创建时保持谨慎。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊