心理学
情绪分析
语言学
多元分析
多元统计
社会心理学
自然语言处理
认知心理学
人工智能
尼泊尔语
作者
Sarah J. Mansfield,Michael Laudenbach,Alex Reinhart,Kachatad Pantusen,David West Brown
出处
期刊:Ampersand
[Elsevier BV]
日期:2026-11-03
卷期号:17: 100294-100294
标识
DOI:10.1016/j.amper.2026.100294
摘要
Writing instructors increasingly report an intuitive “feeling” that something is wrong with LLM-generated academic writing—a sense of inappropriate tone, excessive positivity, or what researchers have termed “sycophancy.” However, this intuition, while valuable, leaves both instructors and students without concrete language or systematic methods to identify, discuss, or remedy these differences. This study introduces multivariate sentiment analysis as both a research methodology and a pedagogical tool for making visible the affective dimensions that distinguish human academic writing from LLM outputs. By applying the GoEmotions sentiment analysis model to a balanced corpus of human and AI-generated academic texts, we demonstrate how computational methods can quantify what experienced writers intuitively recognize, transforming impressionistic observations into teachable knowledge. With a focus on the particularly stark differences in academic registers, our findings show that GPT-4o uses significantly more approval and admiration language than human writers (approximately 2x more common) and expresses less neutral language (64.7% vs. 73.2%). Importantly, different LLM models exhibit distinct emotional patterns that vary from human writing in different ways, underscoring that these register differences will persist as models evolve. These findings provide instructors with concrete, data-driven explanations for why LLM outputs often feel inappropriate for academic contexts, supporting the development of critical awareness about generative AI tools in academic writing. Instructors may adapt these findings and find accessible examples that show students how AI-generated texts fail to capture appropriate human-like emotional expression in specific contexts.
科研通智能强力驱动
Strongly Powered by AbleSci AI