Self-reported AI literacy and self-efficacy showed different relationships with student dependency in a 478-person survey

Maizel and colleagues examine why students with greater AI literacy do not all relate to AI in the same way. Their July 2026 Education Sciences article analyzes an online survey of 478 Israeli higher-education students recruited through a paid panel in August and September 2025. The sample included 228 college and 250 university students; 76 percent were women, the mean age was 25.8 years, and most were studying for a bachelor's degree. Fourteen of 492 initial respondents were excluded because their AI-dependency measure was missing.

The survey combined a five-item AI-dependency scale with the 20-item Meta AI Literacy Scale and measures of academic self-efficacy and learning-resource management. The dependency scale had a reported Cronbach's alpha of .85 and the overall AI-literacy scale an alpha of .91, although one ethics and emotional-regulation subscale was less reliable at .64. Resource-management measures covered time and study-environment management, effort regulation, peer learning and help seeking. The help-seeking items concern human sources, so they should not be interpreted as direct measures of asking an AI system for help.

In a regression with eight predictors, the model was statistically significant, F(8, 469) = 22.03, p < .001, with R squared of .273 and adjusted R squared of .261. Using and understanding AI was the strongest positive predictor of self-reported dependency, with a standardized coefficient of approximately .404. Detecting AI was also a smaller positive predictor. In contrast, AI self-efficacy, academic self-efficacy and effort regulation were associated with lower dependency. These coefficients describe relationships after accounting for the included variables; they do not establish which factor came first.

The authors also describe four profiles. High literacy with low dependency included 156 students, or 32.6 percent; moderate literacy with moderate dependency included 127; low literacy with low dependency included 110; and high literacy with high dependency included 85, or 17.8 percent. The high-literacy, low-dependency group reported the strongest academic self-efficacy, time and study management, and effort regulation. Help seeking did not differ significantly across the profiles. The coexistence of two high-literacy groups is more informative than a simple claim that literacy either prevents or causes dependency.

Important limits remain. The cross-sectional, self-report design cannot establish causal direction: frequent AI use may build technical knowledge, technical confidence may encourage more use, or unmeasured factors may affect both. The paid Israeli panel, high proportion of women and single survey period limit generalization. No chat logs, assignment records or independent learning outcomes were collected. The dependency scale also should not be equated with usage frequency, clinical addiction or proof that a student's learning has deteriorated.

For Hong Kong higher education, the study supports a broader AI-literacy design rather than a direct intervention claim. Courses can combine tool knowledge with academic self-efficacy, effort regulation, source checking and planned moments of human help. Evaluation should include multilingual and discipline-balanced samples, behavioral evidence and unaided performance over time. The practical question is not whether students know more AI functions, but whether they can decide when to use them, sustain effort and demonstrate understanding when assistance is removed.