Does Access to Generative AI for Learning Change High School Students’ Academic Self-Concept?

• In Event: AI and Youth Learning Across the Globe: Insights from Four Large-Scale Surveys

Abstract

Objective
The rise of generative artificial intelligence (AI) presents educators and researchers with a paradox: its personalized feedback and tutoring capabilities could strengthen students’ self-efficacy, yet over-reliance on AI might undermine authentic academic self-belief. The current study aims to examine domain-specific changes in high school students’ academic self-concept when generative AI becomes available and to identify the key predictors of such changes.

Theoretical Framework
According to social cognitive theory (Bandura, 1997; Schunk & DiBenedetto, 2020), when adolescents successfully complete challenging academic tasks with AI assistance, these experiences may contribute to enhanced perceptions of competence and increased willingness to engage with difficult material (Zimmerman, 2000). On the other hand, theoretical concerns emerge from self-determination theory and attribution theory regarding AI’s potential to undermine intrinsic motivation and authentic self-concept development (Deci & Ryan, 2000; Weiner, 2018). If academic success becomes heavily dependent on AI assistance, adolescents may experience what researchers term “cognitive offloading,” where essential learning processes are delegated to external tools rather than internalized through effortful practice (Risko & Gilbert, 2016; Storm & Stone, 2015).

Methods
We conducted a survey study with 7,739 (7,509 after attention checks, 61.5% female) high school students in Mexico. Academic self-concept was assessed under a 2 (subject: math and writing) × 2 (AI availability: without AI or with AI) within-subjects design. For each condition, five self-concept dimensions were measured using a Likert scale from 1 (not good at all) to 7 (extremely good). Additionally, we measured students’domain-specific AI use, general AI use, and AI literacy using validated scales (Bickham et al., 2024). All instruments were administered in Spanish and subsequently translated to English for analysis.

Results
Paired t-tests revealed that AI access influences students’ self-concept in math and writing (see Table 1). For both math and writing, students reported higher confidence in peer and subject comparisons, and surprisingly, lower expectations for their future performance with AI availability. At the same time, domain-specific effects emerged: self-assessed ability increased with AI access in mathematics only, whereas confidence in learning new material improved exclusively in writing.
Multiple regression analyses revealed that domain-specific AI use served as the primary predictor of self-concept change (Figure 1 and 2). Math self-concept change was most strongly predicted by math-specific AI use (β = 0.492, p < .001), and writing self-concept change was dominantly predicted by writing-specific AI use (β = 0.375, p < .001). General AI use emerged as an additional predictor in both models, whereas AI literacy showed no significant relation with changes in either domain. Interestingly, cross-domain relations proved asymmetric, with writing-specific AI use predicted math self-concept change (β = 0.058, p = .022) but math-specific AI use showing no relation with writing self-concept change (β = -0.021, p = .394).

Significance
Our study provides empirical evidence for educators and policymakers addressing AI integration in schools. The finding that domain-specific AI use predicts positive changes indicates that one-size-fits-all approaches are suboptimal, whereas the decline in performance expectations highlights the urgent need for pedagogical interventions that maintain students’ agency in the AI age.

Authors

• Echo Zexuan Pan, Harvard University

• Trisha Thomas, Harvard University

• Danny Glick, Oranim College of Education

• Ying Xu, Harvard University