Same tool, different work: patterns of generative AI use and academic outcomes

Synthesis: Same tool, different work: patterns of generative AI use and academic outcomes — a survey study of 157 undergraduates showing that how students use GenAI matters more for academic outcomes than how often they use it. Stamatoulis et al. (2026) identify two distinct patterns of use — evaluative integration (EI), using GenAI to support understanding, and low-verification uptake (LVU), using it as a shortcut — with opposite associations with performance.

The study argues that research on students' GenAI use has relied too heavily on frequency, which captures exposure but not how students engage. Using exploratory factor analysis of items developed from pilot data, it isolates two qualitatively different use patterns and shows that their associations with performance diverge sharply — while frequency predicts neither performance nor academic self-efficacy.

Method

• Design: Survey; exploratory factor analysis of GenAI-use items developed inductively from pilot data; path modeling.
• Sample: 157 undergraduates.
• Key variables: Evaluative integration (EI, use to support understanding), low-verification uptake (LVU, shortcut use), academic self-efficacy (ASE), academic performance; frequency included for comparison.

Key Findings

• Evaluative integration (EI) was positively associated with performance; low-verification uptake (LVU) was negatively associated with it.
• Mediation via academic self-efficacy: EI's association with performance was fully indirect through ASE (no remaining direct effect); LVU showed both a significant indirect association and a remaining direct negative association.
• Frequency predicted neither performance nor ASE — a striking null result supporting the study's core claim that patterns of use, not amount, matter.
• In a separate model, frequency strengthened the EI→ASE positive association, though the interaction was not significant when EI and LVU interactions were estimated jointly; no moderation was found for LVU.

What this means for practice

• Instructors. Design GenAI tasks that require evaluative integration — students documenting what they checked, changed, rejected, and retained — rather than offering unstructured access to the tool.
• Instructors. Stop judging GenAI use by frequency: how often students used the tool predicted neither academic performance nor self-efficacy in this sample.
• Instructors. Watch for low-verification uptake (uncritical shortcut use), which was negatively associated with performance and with academic self-efficacy, and treat it as the pattern to intervene on.
• Faculty developers. Build understanding-oriented use as a route to Self-Efficacy: evaluative integration's link to performance was fully mediated by academic self-efficacy, with no remaining direct effect.
• Researchers. Measure amount and pattern of use separately; the items are an initial empirical step, not a validated scale, and the two-factor structure still needs independent testing.

Limitations

• Items were refined after inspecting the initial factor solution, and the factor structure and outcome associations were tested in the same sample of 157 undergraduates, so both require independent replication.
• Evaluative integration may capture conscientiousness or general self-regulated academic behavior rather than a GenAI-specific pattern; discriminant validity from self-regulation and conscientiousness was not established.
• Social desirability may have inflated evaluative-integration reports and suppressed low-verification-uptake reports, and the restricted high-GPA distribution may have attenuated associations with performance.
• Excluding non-users restricts variation at the low end of frequency and limits findings to students with some academic GenAI use, and the modest single-institution sample limits generalizability.

Citation

Stamatoulis, C., Pyrovetsi, L., Mourikis, C., Ponnam, A., & Karayianni, I. (2026). Same tool, different work: patterns of generative AI use and academic outcomes. PsyArXiv Preprints.