Researchers Develop Validated Scale to Measure Generative AI Addiction in Undergraduate Students

A research team has developed and validated a new psychometric instrument—the Generative Artificial Intelligence Addiction Scale (GAI-AS) (搜索)搜索该药品—designed to measure addictive patterns of generative AI (GAI) use among undergraduate students. Published in Frontiers in Psychology (搜索)搜索该词条, the study by Isbulan, Ergene, and Demirhan addresses a growing concern that the dynamic, personalized, and anthropomorphic nature of GAI systems may foster patterns of excessive or dysregulated use distinct from traditional technology addictions.

The GAI-AS was constructed through a systematic eight-step scale-development framework proposed by DeVellis (2017), grounded explicitly in Griffiths' (2005) components model of addiction within a biopsychosocial framework. The instrument assesses GAI addiction across four dimensions: salience (the extent to which GAI becomes central in daily thinking), excessive dependency (reliance on GAI for cognitive or instrumental tasks at the expense of one's own effort), mood modification (using GAI to cope with stress or regulate emotions), and functional impairment (loss of control and negative academic, social, or personal consequences).

Scale Development and Validation

The scale was developed and tested across three independent groups of undergraduate students. An initial 36-item pool was reviewed by a panel of eight experts—including two with GAI-related scale-development experience, three specializing in technology use, and three in addiction research—resulting in the removal of 12 items for redundancy or limited conceptual clarity. The remaining 24 items were administered to 429 students for exploratory factor analysis (EFA) and 490 students for confirmatory factor analysis (CFA).

EFA using principal axis factoring with direct oblimin rotation yielded an 18-item scale with a four-factor structure accounting for 64.32% of total variance, with factor loadings ranging from 0.569 to 0.833. The Kaiser-Meyer-Olkin (KMO) value of 0.904 and a significant Bartlett's Test of Sphericity [χ²(153) = 3,627.52, p < 0.001] supported sampling adequacy. CFA confirmed acceptable model fit, with χ²/df = 2.66, RMSEA = 0.058, SRMR = 0.051, CFI = 0.98, NFI = 0.97, GFI = 0.93, and AGFI = 0.90.

Convergent validity was established through composite reliability (CR) values ranging from 0.884 to 0.923 and average variance extracted (AVE) values between 0.605 and 0.753, exceeding the accepted thresholds of 0.70 and 0.50, respectively. Internal consistency was strong, with Cronbach's alpha coefficients ranging from 0.81 to 0.89 across subscales and 0.90 for the total scale, alongside a McDonald's omega of 0.91. Harman's single-factor test indicated that common method bias did not meaningfully affect the data, with the first factor accounting for only 37.70% of variance—below the 50% threshold.

Key Findings on Demographic and Usage Patterns

The validated 18-item GAI-AS was subsequently administered to 803 undergraduate students to investigate demographic and usage-related differences in addiction tendencies. The findings revealed no significant differences in GAI-AS scores according to gender across any dimension or the total scale score (p > 0.05). Similarly, no statistically significant differences emerged between students using only the free version of GAI tools and those using both free and paid versions.

Grade-level differences were observed only in the excessive dependency dimension, with freshman students (M = 2.73, SD = 0.90) reporting significantly lower scores than sophomore (M = 2.99, SD = 0.88), junior (M = 3.06, SD = 0.89), and senior students (M = 2.95, SD = 0.99), F(3,799) = 4.588, p < 0.01, η² = 0.016. The authors note the small effect size indicates this difference was modest.

Frequency of GAI use emerged as the strongest and most consistent predictor. Participants using GAI tools every day demonstrated the highest levels across all four dimensions and the total score. ANOVA results confirmed significant group differences for salience, F(3,799) = 21.295, p < 0.01, η² = 0.074; excessive dependency, F(3,799) = 15.689, p < 0.01, η² = 0.055; mood modification, F(3,799) = 11.275, p < 0.01, η² = 0.040; functional impairment, F(3,799) = 4.688, p < 0.01, η² = 0.018; and total score, F(3,799) = 17.665, p < 0.01, η² = 0.062.

Correlation analyses demonstrated moderate positive associations among all GAI-AS sub-dimensions (p < 0.01), including salience with excessive dependency (r = 0.530), mood modification (r = 0.521), and functional impairment (r = 0.536). These interrelations support the conceptualization of GAI addiction as an interconnected behavioral and emotional phenomenon consistent with Griffiths' components model.

Clinical and Conceptual Implications

The authors emphasize that the GAI-AS is intended as a research and screening tool rather than a diagnostic instrument. They caution that dysregulated GAI engagement may not yet meet conventional diagnostic criteria for clinical addiction, and may instead represent a compensatory coping strategy among students experiencing academic insecurity or low self-efficacy. The study highlights the need for future research to clarify whether GAI dependence develops into an independent behavioral pathology or remains better understood as a maladaptive, technology-mediated coping mechanism.

Several limitations are acknowledged, including the undergraduate-only sample, the cross-sectional design that precludes causal inference, the absence of criterion-related validity testing against existing scales, and reliance on self-report data subject to social desirability and recall bias. The authors call for longitudinal and cross-cultural studies, as well as multi-method approaches incorporating behavioral usage metrics and observational data, to strengthen understanding of GAI addiction trajectories.