Page 689 - ISC PROCEEDINGS 21.4

EVALUATING THE IMPACT OF BEHAVIORAL NUDGES ON STUDENTS’ USE OF
ARTIFICIAL INTELLIGENCE IN LEARNING AND SCIENTIFIC RESEARCH:
AN ECONOMIC APPROACH

Nguyen Thi Thu Huong* , Do Minh Dan Anh 2
1

1 Hanoi Open University, Hanoi, Vietnam.
2 University of Sydney, Australia.
(*E-mail: huongntt.kt@hou.edu.vn)

ABSTRACT

As large language models (LLMs) become deeply embedded in university students'
academic workflows, concerns about uncritical AI dependence have intensified. This study
employs a behavioral economics framework to experimentally evaluate two categories of
nudges - verbalized uncertainty and cognitive forcing - as potential interventions against
AI overreliance among Vietnamese undergraduate students. Contrary to expectations,
neither nudge reduced overreliance. Verbalized uncertainty showed no statistically
significant effect on accuracy, while cognitive forcing conditions yielded markedly lower
performance than the control. Critically, participants who had initially selected the
correct answer disproportionately revised toward the AI-endorsed incorrect answer
following exposure to LLM output without a corresponding shift in self-reported
confidence, suggesting that behavioral compliance can occur independently of genuine
persuasion. These patterns are interpreted through the lenses of cognitive dissonance
and Social Comparison Theory, with AI functioning as an authoritative social referent that
destabilizes rather than supplements students' independent judgment. The findings
challenge the adequacy of nudges alone and call for interventions that directly address
students' perception of AI authority in educational settings.
Keywords: Large language models (LLMs); cognitive forcing; verbalized uncertainty;
cognitive dissonance; self-confidence calibration.

1. Introduction
Artificial intelligence, though not a novel concept, has only risen to mainstream use
through the popularization of large language models (LLMs) such as OpenAI’s ChatGPT,
Google Gemini and Claude. From the basis of natural language processing (NLP), LLMs are
extensively developed and enhanced through pre-training, supervised fine-tuning,
human- based reinforcement learning and parameter fine-tuning (Min et al., 2023;
Minaee et al., 2024). The integration of LLMs into research and education has been
debateable, for the rapid development in AI-generated responses’ accuracy and
immediacy has proved to be a convenient tool for students - particularly tertiary-level
ones - in researching, self-learning or even answering quizzes and completing
assignments.
Convenient and accurate as it may seem, problems with LLMs are prevalent,
including but not limited to hallucination, failure to follow instructions and sycophancy
(Alansari & Luqman, 2025; Geng et al., 2025; Cheng et al., 2025). Hallucination in LLMs is
characterized by the provision of fabricated and/or inaccurate information, which occurs
due to the model’s tendency to prioritize grammatical proficiency and coherence over

688