Technical Summary: Academic Integrity and Emotional Responses to Inappropriate LLM Use in Software Engineering Education

Problem Statement

Academic integrity in higher education is increasingly defined by complex socio-technical environments where automated tools, such as Large Language Models (LLMs), blur the boundaries of authorship and acceptable assistance. While existing literature extensively documents the prevalence of academic misconduct and the technical capabilities of LLMs to generate code and text, there is a significant gap in understanding the affective and moral experiences of students who engage in what they perceive as academically inappropriate LLM use.

Prior research in software engineering education has largely focused on defining misconduct, identifying vulnerable assessment formats, or proposing technical detection mechanisms. However, it offers limited insight into how students emotionally process the violation of integrity norms when using LLMs. Specifically, it remains unclear whether students experience strong negative emotions (e.g., guilt, anxiety) upon recognizing they have compromised academic standards, or if these actions are met with indifference or rationalization. This study addresses the research question: How do software engineering students describe their emotional experiences after using large language models in ways that violate academic integrity expectations?

Methodology

The authors employed a cross-sectional survey design to investigate self-reported experiences of 116 undergraduate software engineering students. The study utilized an anonymous online survey (Qualtrics) structured into five thematic sections:

1. Experiences with Inappropriate LLM Use: Participants were asked to describe specific situations where LLM use conflicted with course rules.
2. Motivations and Decision Factors: Questions probed the reasons for use and justifications employed at the time.
3. Emotional Responses: Participants selected from a list of emotions (Relief, Guilt, Indifference, Satisfaction, Anxiety) and provided open-ended reflections on their feelings.
4. Affected Areas: Identification of which software engineering activities (e.g., programming, testing, design) were most impacted.
5. Perceived Consequences: Assessment of anticipated academic and professional outcomes.

Data Collection and Filtering:

• Recruitment: Participants were recruited via the Prolific platform and snowball sampling, filtered for current enrollment in undergraduate software engineering programs and prior LLM experience.
• Screening: From an initial 208 responses, 60 were removed for failing eligibility or attention checks. A subsequent manual review excluded 32 responses exhibiting signs of AI-generated or low-authenticity content, resulting in a final dataset of 116 valid responses.
• Analysis: The study combined descriptive quantitative statistics with iterative inductive qualitative coding. Open-ended responses were coded for recurring patterns regarding motivations, emotional reactions, and contextual factors. To systematically characterize affective responses, the authors applied the Self-Assessment Manikin (SAM) framework, analyzing responses along valence (positive/negative/neutral) and arousal (calm/stressed) dimensions.

Key Contributions

This study contributes to the field of software engineering education by shifting the focus from purely behavioral or technical aspects of academic integrity to the affective dimension.

• Empirical Evidence of Emotional Heterogeneity: It provides the first empirical data specifically linking LLM misuse in software engineering to a spectrum of emotional responses, challenging the assumption that integrity violations are uniformly accompanied by distress.
• Contextual Nuance: The research highlights that emotional responses are not uniform but are shaped by the specific software engineering activity (e.g., introductory coding vs. capstone projects) and the situational context (e.g., deadline pressure, unclear guidance).
• Integration of Affective and Contextual Analysis: By combining the SAM framework with qualitative narratives, the study offers a structured way to interpret how students reconcile the awareness of risks with their emotional state.

Results

The survey revealed emotionally heterogeneous responses, with indifference being the most dominant sentiment, even among students who recognized the risks to their learning and academic standing.

1. Emotional Distribution

• Indifference (36.2%): The most frequent response. Students often acknowledged potential consequences (e.g., failure, skill erosion) but described them in a detached, pragmatic, or minimized manner. In the SAM framework, this correlated with neutral valence and a range of arousal levels, suggesting a normalization of risk.
• Guilt (20.7%): Associated with negative valence and moderate arousal. This emotion was linked to moral discomfort and self-reproach regarding the undermining of personal learning and skill development.
• Anxiety (18.1%): Characterized by negative valence and moderate-to-high arousal. This was driven by fear of external sanctions, detection, or long-term professional repercussions.
• Relief (18.1%): Associated with positive valence but often accompanied by moderate-to-high arousal. This emotion arose primarily in deadline-driven contexts or situations with unclear guidance, reflecting a reduction in perceived stress rather than an absence of concern.
• Satisfaction (6.9%): The least reported emotion, linked to positive valence and low arousal, often citing efficiency or technological progress as justification.

2. Contextual Drivers

• Course Types: Inappropriate use was reported across the curriculum, with high prevalence in introductory programming (due to skill gaps and time pressure), theory/writing courses (due to insecurity or lack of engagement), and advanced design/capstone projects (due to conceptual complexity and high stakes).
• Perceived Consequences vs. Emotion: A critical finding is the dissociation between consequence awareness and emotional engagement. Many students articulated serious academic or professional risks (e.g., failing the course, suspension) yet maintained a neutral or indifferent emotional stance. Negative emotions like guilt and anxiety were typically situational rather than sustained.

3. Demographic and Structural Factors

• Guidance: Over half (54.3%) of respondents reported no formal training on responsible LLM use, and 53.4% noted that course expectations were communicated only in general terms.
• Academic Stage: Emotional responses were distributed across all years of study, indicating that indifference is not merely a result of early exposure but persists through senior levels.

Significance and Claims

The paper claims that academic integrity in the era of LLMs is increasingly understood as an ongoing process of sense-making shaped by evolving technologies and institutional responses, rather than a fixed standard of moral failure.

• Normalization of Misuse: The study argues that LLM-assisted cheating in software engineering is a normalized and affectively heterogeneous practice. It is sustained by routine academic pressures (deadlines, workload) and negotiated interpretations of acceptable use, rather than a lack of awareness of consequences.
• Limitations of Prohibition: The authors suggest that educational responses centered solely on prohibition or enforcement may be misaligned with student experiences. Since awareness of consequences does not consistently produce strong negative affect, simply increasing penalties may not deter misuse if the emotional response remains neutral or rationalized.
• Future Directions: The paper modestly claims that these findings highlight the need for research and educational practices that account for activity-sensitive perspectives. It suggests that future work should move beyond documenting prevalence to understanding how students interpret acceptable use across different stages of the software engineering lifecycle.

The authors conclude that emotionally muted or ambivalent responses are a critical aspect of how software engineering students experience academically inappropriate LLM use, challenging the assumption that integrity violations are necessarily accompanied by strong negative affect.