Attributes Associated with Emotional Discomfort and AI Guilt When Delegating Social Work Tasks to AI Llms

Introduction: Emotional discomfort and ethical concerns have been reported by social work students, faculty, and practitioners as AI large language models (AI LLMs) are increasingly adopted. Such concerns were less prominent in explaining adoption of previous information technologies, even though social workers were often slow adopters of those innovations. A recently emerging construct of “AI guilt,” has developed in higher education, reflecting students’ ambivalence about using AI LLMs for text generation in assignments. Given the extensive documentation demand in social work, AI guilt is likely relevant to discussions surrounding the adoption of AI LLMs in practice and education. This study aims to understand factors associated with higher emotional discomfort related to the perceived pressure to adopt AI LLMs for work.

Methods: A total of 121 respondents (faculty/staff = 30%, students = 53%, practicum supervisors = 17%) completed an online survey that lasted approximately 23 minutes. The average age was 37.4 years (SD=12.3), with 56.7% identifying as non-Hispanic White and 43.3% identifying as other racial/ethnic groups. AI guilt was measured using a 15-item Likert scale (1 = low AI guilt to 5 = high AI guilt), including12 items adapted from Chan (2025), one item measuring emotional discomfort related to environmental impact, and one item assessing concerns about compromising personal ethical standards. Unadjusted analyses included one-way ANOVA, independent t-tests, and Pearson’s bivariate correlations between AI guilt and covariates. Adjusted analyses were conducted using multiple regression, with the significance level set at 0.05.

Results: About one-third of respondents (31%) reported using AI LLMs more than once a week. Regarding familiarity and readiness, 43% were very familiar or more with AI LLMs, and two-third (67%) reported high confidence in using technology. However, only one-fifth (21%) had received AI-related training, and one-third (33%) reported having a paid AI subscription. The mean of AI guilt was 3.09 (SD=.63) indicating slightly higher than moderately felling such emotional discomfort with using AI LLM. Students reported the highest levels (mean=3.25; SD=0.65), followed by practicum supervisors (mean=2.93; SD=.65) and faculty/staff (mean=2.91; SD=.60). In unadjusted analysis, higher AI guilt was reported among those who used AI LLMs fewer than once a week (Δ = .49; p < .001), hadn’t received AI training (Δ = .33; p < .019) and didn’t have a paid subscription (Δ = .46; p = .001). Younger respondents showed higher AI guilt (r = −.46; p < .001). In adjusted analyses (adjusted R² = .250). AI guilt was higher among younger respondents (β = −.34, p = .002) and among those who used AI LLMs less than once per week (β = −.25, p = .011).

Discussion and Conclusions: Our study shows that younger respondents and less frequent or non-users of AI LLMs reported higher levels of AI guilt. These findings highlight the need for clearly defined ethical expectations to help reduce AI guilt, particularly among younger students, faculty and staff, practicum supervisors, and individuals with limited experience using AI, while promoting the responsible use of AI in the social work profession.