Source-linked AI summary
Trust in Generative AI among students: An Exploratory Study
Matin Amoozadeh, David Daniels, Daye Nam, Aayush Kumar, Stella Chen, Michael Hilton, Sruti Srinivasa Ragavan, Mohammad Amin Alipour
TL;DR
Students’ trust in GenAI is important because it shapes how novice programmers use AI-generated code, yet educational evidence about that trust is limited. The paper surveys 253 students at two universities and finds varied trust levels, with trust associated with self-reported motivation and confidence improvements. The findings support further work on calibrating trust across student populations and educational settings.
Problem
Educational evidence is limited on whether novice programming students can appropriately calibrate trust in GenAI tools, despite trust shaping adoption and use.
Method
The authors surveyed 253 undergraduate and graduate students at the University of Houston and IIT Kanpur about GenAI use, trust, confidence, motivation, and perceived effects on programming.
Results
Students reported varied trust in GenAI: approximately 16% expressed distrust, 36% were neutral, and 47% reported trust; trust also correlated with self-reported motivation and confidence improvements.
Takeaways & Limitations
The findings highlight the need to understand trust across student populations so educators and developers can support appropriately calibrated GenAI use.
Takeaways & Limitations
Because the survey had a small sample relative to the overall student population, the generalizability of its results is limited.
Abstract
from arXiv · showhide
Generative artificial systems (GenAI) have experienced exponential growth in the past couple of years. These systems offer exciting capabilities, such as generating programs, that students can well utilize for their learning. Among many dimensions that might affect the effective adoption of GenAI, in this paper, we investigate students' \textit{trust}. Trust in GenAI influences the extent to which students adopt GenAI, in turn affecting their learning. In this study, we surveyed 253 students at two large universities to understand how much they trust \genai tools and their feedback on how GenAI impacts their performance in CS courses. Our results show that students have different levels of trust in GenAI. We also observe different levels of confidence and motivation, highlighting the need for further understanding of factors impacting trust.
1 INTRODUCTION
This study examines how students use and trust GenAI in programming, addressing limited educational evidence about whether novice programmers can calibrate trust appropriately. A survey of 253 students found varied trust levels and significant associations between trust, motivation, and confidence.
- Students may overtrust GenAI and accept suggestions blindly, or undertrust it and abandon potentially useful assistance.
- The survey investigates students’ GenAI use, trust, perceived benefits and drawbacks, and perceptions of GenAI in programming.
- 47% of participants reported trusting GenAI, while 36% were neutral and approximately 16% expressed distrust.
- Trust was significantly correlated with self-reported improvements in students’ motivation and confidence in programming.
- The findings have implications for developers designing effective tools and educators fostering appropriate trust during learning.
2 RELATED WORK
Prior research has examined GenAI programming tools, their educational applications, and trust in AI, but has largely emphasized tool effectiveness or professional users. Student trust in educational GenAI remains comparatively understudied.
- Research on GenAI programming assistants has evaluated generated code, explanations, usefulness, and user experiences.
- GenAI in education raises concerns about plagiarism, bias, inaccurate results, poor habits, and challenges for existing pedagogy.
- Educational studies have explored code generation, explanations, and debugging assistance, while noting that benefits vary with task complexity.
- Trust research defines trust as expectations about how AI can help under uncertainty or vulnerability and recognizes risks from both insufficient and blind trust.
- Despite broader AI-trust research, little is known about student trust in educational AI tools and its implications.
3.1 Data Gathering: Survey Study
The authors surveyed undergraduate and graduate students at the University of Houston and IIT Kanpur using a mixed set of demographic, experience, trust, confidence, and opinion questions. The questionnaire combined mostly closed-ended items with one open-ended component and was refined through piloting.
- The survey recruited students from large public universities in the US and India to gather perspectives from diverse participant groups.
- Questions covered demographics, GenAI programming experience, programming confidence, trust operationalized through a prior survey, and AI-tool experiences.
- The questionnaire used five sections to distinguish GenAI users and non-users and collect users’ trust, motivation, confidence, and opinions.
- The analysis was predominantly quantitative, with open coding used for responses to the open-ended question.
- The survey was piloted with teaching assistants, revised based on feedback, and excluded pilot data from the final analysis.
- Participation was voluntary, including recruitment from OS and CS2 at UH and a broader student email list at IITK.
3.2 Data Analysis
The study combined descriptive and correlational quantitative analyses with thematic open coding to examine students’ GenAI use, trust, and perspectives. Two authors collaboratively coded qualitative responses and iteratively reviewed prior coding.
- The authors used descriptive statistics, distribution comparisons, and correlations to identify factors related to students’ trust in AI.
- The qualitative analysis coded 73 optional responses into 15 data-induced themes about students’ attitudes toward AI in programming.
- Table 2 summarizes whether students had heard of GenAI, used it, and the tasks for which users selected GenAI tools.
- Two authors collaboratively assigned one or more categories to responses and revisited earlier comments whenever new codes emerged.
- The thematic coding supported insights into the complex range of student perspectives and emotions surrounding programming AI.
4 RESULTS
Students widely use GenAI for programming help and code understanding, while trust varies across participants and demographics. Trust is generally neutral and positively associated with perceived motivation and confidence, with stronger associations among first-generation students.
- RQ1: Exposure and Adoption of GenAI: Most students had used GenAI, primarily for programming help and understanding code rather than writing new code.Students also used GenAI for non-programming tasks; the survey allowed multiple task selections.
- RQ2: Students’ trust in GenAI: Students’ trust was generally neutral, averaging 3.07 with a median of 3, and respondents were more likely to reject than accept outputs they could not verify.Among 216 participants, 103 disagreed with believing uncertain outputs, while 45 agreed.
- RQ2: Students’ trust in GenAI: Trust differed across geography, class level, and gender, including higher trust among UH CS2 students than OS students and among UH female than male users.UH CS2 and OS averages were 3.23 and 2.95; UH female and male averages were 3.31 and 3.07.
- RQ2: Students’ trust in GenAI: Trust showed a moderate positive correlation with students’ perceived improvements in motivation and confidence when using GenAI.The study reports no correlation between years of programming experience and trust in the overall data.
- RQ2: Students’ trust in GenAI: Among first-generation students, trust correlated more strongly with improved motivation and confidence than among continuing-generation students.The reported correlations were 0.57 and 0.4 for first-generation students versus 0.37 and 0.35 for continuing-generation students; causation was left for future work.
- Contribution to Learning: Open-ended responses highlighted distrust caused by perceived biases, errors, and limitations, alongside views that GenAI can help learning and programming.Respondents emphasized human supervision, while some described GenAI as a supplement rather than a solution.
5 THREATS TO VALIDITY
The survey’s validity is constrained by self-reported responses, limited generalizability, predominantly STEM participants, and self-selection bias.
- External validity: The small, self-selected, predominantly STEM sample limits generalization to the broader student population despite surveying two universities on different continents.The authors also identify memory bias in surveys, while reporting randomized options and researcher triangulation to mitigate some concerns.
6 DISCUSSION AND CONCLUSION
The discussion finds substantial variation in students’ trust in GenAI and links trust with reported motivation, confidence, and knowledge gains. It argues that educators and system designers need to understand and calibrate this trust for appropriate educational use.
- Although GenAI models are opaque and can behave unpredictably, 118 of 230 participants believed GenAI was transparent.This mismatch between model opacity and student perceptions underscores the need to moderate trust for informed use.
- As GenAI tools become more capable and widely used, students who distrust them may miss benefits and fall behind peers who use them.The authors therefore call for educators to identify appropriate trust levels and help students develop them.
- Students differed in trust across demographic groups, including class standing and gender at the American university.These differences motivate further investigation of which factors shape trust in different student populations.
- Understanding trust can inform educational interfaces, classroom pedagogy, and decisions about when to encourage or discourage GenAI use.The discussion assigns this need to both educators and GenAI system designers.
- Trust was positively correlated with students’ reported improvements in motivation, confidence, and knowledge, with a stronger correlation among first-generation students.The authors suggest judiciously designed and used GenAI could provide additional support for first-generation students, but present this as a possibility requiring further study.