The AI guilt complex: Moral emotions and ethical dilemmas in academic technology adoption

Synthesis: Vassallo surveyed all 3,100 academic staff at a comprehensive public university in Malta in 2025; 109 completed the survey, a 3.5% response rate the paper treats as data about the state of ethical debate in academia rather than only as a limitation. On a four-item AI Guilt Index (Cronbach's α = 0.88) anticipatory guilt about professional standing substantially exceeded remorse reported after use, and k-means clustering produced four moral response profiles from Comfortable Adopters to Morally Distressed Avoiders. The guilt paradox is the counterintuitive result: non-users reported higher guilt than users, which the paper reads as anticipatory distress preventing the experience that might resolve it. It argues that AI integration in Higher Education needs emotional and identity work alongside technical training and policy.

Key Findings

1. Anticipatory guilt outran experienced remorse. Mean AI Guilt Index score was 2.39 (SD = 1.01), with 25 participants (22.9%) above the midpoint. Credibility worry drew the strongest endorsement (M = 2.68, SD = 1.32; 34.9% agreeing or strongly agreeing), then feeling like cheating when using AI (M = 2.46, SD = 1.23; 25.7%), feeling bad about automating previously manual tasks (M = 2.41, SD = 1.10; 17.4%) and remorse after AI use (M = 2.02, SD = 1.05; 9.2%), so fear of transgression outweighed its emotional impact.
2. Non-users reported more guilt than users. Those who never used AI for research papers or teaching preparation (n = 8) scored higher (M = 3.25, SD = 1.04) than any users (n = 101, M = 2.32, SD = 0.97), t(8.39) = 2.00, p = .078. The remorse item separated them, t(7.79) = 2.44, p = .041 (non-users M = 3.00, SD = 1.41; users M = 1.94, SD = 0.99), read as anticipated rather than experienced guilt.
3. Cluster analysis identified four moral response profiles: Comfortable Adopters (n = 29, 26.6%), Guilty Non-Users (n = 32, 29.4%), Cautious Users (n = 31, 28.4%) and Morally Distressed Avoiders (n = 17, 15.6%). ANOVA confirmed differences on guilt, F(3,105) = 22.13, p < .001, η² = 0.39, research use (η² = 0.57) and teaching use (η² = 0.56). Only Comfortable Adopters show low moral negotiation, so most respondents are negotiating something.
4. Guilt falls as career security rises. Career-stage differences were significant, F(3, 99) = 2.78, p = .045, η² = 0.078. Early-career academics reported the highest guilt (M = 2.71, SD = 1.22; 40.0% above midpoint), then mid-career (M = 2.50, SD = 0.95; 21.1%) and senior academics (M = 2.03, SD = 0.84; 13.5%). Tukey testing located the difference between early-career and senior academics (mean difference = 0.68, p = .037) and the linear trend was significant (p = .012). Gender showed no difference, t(100.2) = 1.24, p = .217.
5. Guilt tracked concealment rather than disclosure. Limiting AI use because of uneasiness correlated with guilt at r = .62 (95% CI [0.49, 0.72]), avoiding AI for peer-reviewed work at r = .51 ([0.36, 0.64]) and avoiding disclosure to colleagues at r = .50 ([0.34, 0.63]), all p < .001. Keeping different AI policies for oneself than for students correlated at r = .29 (p = .002), while formal disclosure in academic outputs showed no relationship (r = .08, p = .398).
6. Institutional clarity helped only weakly. Only 34 participants (31.2%) agreed that their institution provides clear guidelines, and clarity correlated negatively but weakly with guilt (r = −.25, 95% CI [−0.42, −0.06], p = .010). Perceiving AI as ethically problematic in one's field correlated at r = .26 (p = .007), negative colleague views barely at r = .14 (p = .152) and pressure to keep up with colleagues not at all (r = .02, p = .837).

How the study was conducted

The design was a cross-sectional mixed-methods survey, exploratory and hypothesis-generating because no prior empirical work on AI-related guilt in academia exists. The population was all academic staff at one comprehensive public university in Malta (N = 3,100) across 14 faculties. Recruitment ran through institutional email, no incentives were offered, and the survey was anonymous, with median completion time about 15 minutes. The 46-item instrument covered demographics, AI usage frequency, moral emotions and identity (10 Likert items), behavioural patterns, institutional context and three open-ended prompts: a time the respondent felt conflicted or guilty, how they reconcile AI use with the values they teach, and which parts of their professional identity feel most challenged. Analysis used R 4.3.1 for reliability analysis, exploratory factor analysis, t-tests and ANOVA with effect sizes, k-means clustering and Pearson correlations, and open-ended responses were analysed with Braun and Clarke's (2006) reflexive thematic analysis, with theme percentages given as approximate proportions.

The AI Guilt Index emerged from the moral emotion items. Factor analysis gave a unidimensional solution explaining 63.5% of variance (eigenvalue = 6.35, KMO = 0.93) and four retained items: credibility worry (loading 0.80), feeling like cheating (0.83), feeling bad about automating manual tasks (0.79) and remorse after use (0.81), giving α = 0.88 with item-total correlations of 0.68 to 0.80. The highest-loading item, feeling conflicted about using AI while teaching students about Academic Integrity (0.90), was excluded as conflating personal guilt with pedagogical role conflict. Power analysis indicated adequate power (0.80) for medium correlations (r = .30) but not for small to medium group effects.

Reading the guilt paradox

Anticipatory guilt is treated not as a milder form of remorse but as a different mechanism. Because non-users never test their fears, avoidance can preserve the distress it was meant to avoid, which the paper frames through technology acceptance research and technology threat avoidance theory (Liang and Xue, 2009), where perceived threat overrides utilitarian benefit. Two details temper this: the non-user group is small (n = 8 in the strictest comparison), and the cross-sectional design cannot distinguish guilt preventing adoption from low adoption sustaining guilt. The AI Guilt Complex is positioned as anticipatory guilt specific to cognitive augmentation, arising under normative ambiguity rather than clear norm violation, and echoing Watermeyer et al.'s (2025) triad of deskilling, inauthentication and dehumanisation.

Four moral response profiles

• Comfortable Adopters (26.6%, n = 29) report the lowest guilt (M = 1.77) and highest use across every activity; only one member (3.4%) exceeded the guilt midpoint. "I don't feel guilty as I use it mainly for finding relevant papers".
• Guilty Non-Users (29.4%, n = 32) are the largest group, with moderate guilt (M = 2.54) and selective use concentrated in teaching (3.38) rather than research (1.72), 28.1% above the midpoint. Their resistance reads as identity protection: "I refuse to use AI. I won't let my thinking be influenced by what is created by a machine".
• Cautious Users (28.4%, n = 31) hold low to moderate guilt (M = 2.12) with moderate, consistent use and 9.7% above the midpoint, negotiating boundaries rather than adopting or refusing wholesale.
• Morally Distressed Avoiders (15.6%, n = 17) combine the highest guilt (M = 3.68) with near-complete avoidance (research 1.18, teaching 1.53), and 70.6% scored above the midpoint. The author advises caution with this group while retaining four clusters over a three-cluster alternative, because high guilt with avoidance is distinct from moderate guilt with moderate use.

What academics said about managing moral discomfort

Ninety participants (82.6%) gave at least one substantive open-ended response: 60 (55.0%) answered the guilt question, 86 (78.9%) the reconciliation question and 60 (55.0%) the identity question. Four dilemma patterns emerged: a temporal guilt dilemma where early discomfort conflicted with perceived benefit (about 35% of responses); principled non-use under institutional and efficiency pressure (15%); pragmatic adoption through boundary setting (20%); and an authenticity dilemma about ownership, originality and intellectual labour (25%).

Five reconciliation strategies appeared: tool framing, comparing AI to calculators or spell-check (30%); emphasis on critical thinking and human evaluation (25%); boundary setting (20%); transparency (15%); and 10% asserting no reconciliation was needed. Identity challenges mirrored this: skill atrophy anxiety (30%), an authenticity crisis around originality (25%), career-stage differentiation (20%) and, notably, 25% reporting no identity threat at all. A temporal trajectory, where remembered guilt diminished with continued use, explains the paradox: "In general, I felt guilty about using it considerably in the beginning when we knew even less about it, it was a sort of experimentation phase".

Implications for practice

Technical training and usage policy will not resolve a problem that is partly emotional and identity-based. Because only 31.2% of respondents saw clear guidelines and the correlation between clarity and lower guilt was weak, the paper recommends that universities create spaces for open discussion of moral concerns and collective boundary-setting rather than relying on top-down mandates, and that academic development programmes treat guilt and identity concern as normal transitional responses rather than faults to correct. The temporal trajectory suggests structured, low-stakes experimentation as a route through anticipatory anxiety, and the career-stage result supports mentorship by senior academics who have integrated AI, since guilt falls hardest on those facing the greatest professional pressure.

Limitations

The 3.5% response rate limits generalisability, and the author is explicit that these findings describe a self-selected minority willing to engage with questions about AI and moral emotions; the 2,991 non-respondents are discussed as a contextual signal about ethical debate. The cross-sectional design prevents causal inference, and the single-institution, single-country setting limits transferability, although patterns were consistent across disciplines within the sample. The AI Guilt Index is new, so convergent and discriminant validity remains untested despite its internal consistency, and the four profiles are heuristic categories rather than stable typologies. All results are framed as hypothesis-generating, with longitudinal, multi-institutional and cross-cultural validation named as priorities.

Connected Concepts

• Academic Integrity — the norm academics feel they may be breaching when using AI themselves
• Anxiety and Stress — anticipatory guilt as an anxiety pattern preceding use
• Teacher AI Competency — emotional readiness as the dimension missing from faculty AI competency frameworks
• Workplace Learning — the paper's case for moving beyond technical training into moral discussion
• Well-Being — unaddressed moral distress as a professional wellbeing issue
• Ethics — normative ambiguity, not rule violation, as the source of AI-related guilt
• Trust — perceived colleague judgement and the weak reassurance of institutional guidelines
• Trust Calibration — fear of transgression exceeding experienced discomfort is a miscalibrated appraisal
• Educational Development — the academic development and mentorship response recommended
• Technology Adoption Models — the rationalist adoption model that moral factors qualify
• Teaching — teaching integrity while using the tools one restricts for students
• Self-Report Measures — the AI Guilt Index as a new instrument needing validation
• AI Use and Disclosure Statements — disclosure avoidance as the behavioural correlate of guilt
• Theory Development in AI in Education — anticipatory guilt in cognitive augmentation as a new construct

Connected Articles

• When faculty ask, 'what's the point of teaching?': GenAI as identity crisis, not skills gap — Faculty identity disruption under GenAI, the identity side of this guilt complex
• Stuck in a Spiral": Shame and Guilt as Social Regulators of AI Use in Computing Education — Guilt and shame linked to self-regulation strategies in AI-assisted learning
• From fear to innovation: A case study of transformative faculty development for ethical AI integration in higher education — Faculty development responses to ethical AI use
• From AI Anxiety to Strategic Regulation: How University Students Transform Generative AI into a Strategic Learning Resource — Anxiety and regulation strategies in AI-assisted academic writing
• Guarded adoption of generative AI in higher education: high-achieving students, successful-student identity — Guarded, boundary-limited adoption as a pattern beyond this sample
• ‘Resistance is futile?’: identity tensions and principled selectivity in AI-integrated teaching — Identity tensions when AI enters teaching practice
• AI Anxiety: A Comprehensive Analysis of Psychological Factors and Interventions — Prevalence and structure of AI anxiety in university populations
• Navigating the moral panic: encouraging appropriate use of GenAI in the classroom rather than condemning innovation as disruption — Moral framing of GenAI in academic settings

Citation

Vassallo, D. (2026). The AI guilt complex: Moral emotions and ethical dilemmas in academic technology adoption. Journal of Academic Ethics, 24, 53.