On this page

Synthesis: A cross-sectional survey of 425 master's students at 19 Chinese universities, all of whom reported using AI regularly in their research, tested whether AI technology readiness raises research self-efficacy directly or through AI trust and supervisor academic support. Analyzed with partial least squares structural equation modeling, the direct readiness paths were not significant, but both indirect routes were: readiness facilitators predicted self-efficacy through perceived supervisor support and through AI trust. Supervisor support was the strongest direct correlate of research self-efficacy. Contrary to the TRI assumption that inhibitors suppress adoption, readiness inhibitors were positively associated with AI trust, which the authors read as critical scrutiny rather than technophobia. No path differed significantly between the humanities/social sciences and science/engineering/medicine groupings.

Key Findings

  1. Readiness does not raise research self-efficacy on its own. Neither facilitators nor inhibitors had a significant direct path once the mediators were in the model.
  2. Supervisor academic support is the strongest direct correlate of research self-efficacy, stronger than AI trust.
  3. Readiness facilitators work mainly through the supervisory relationship. Their indirect route through supervisor support was significant even though the direct route was not.
  4. AI trust is a second, smaller channel. Facilitators predicted trust strongly, and trust in turn predicted self-efficacy more weakly.
  5. Skepticism did not depress trust. Inhibitors were positively associated with AI trust and unrelated to supervisor support or self-efficacy.
  6. Disciplinary context did not change the structure of the model, so the same pattern held across humanities, social sciences, science, engineering, and medicine.

Readiness works through people, not around them

The study's central result is that a personal disposition — readiness to use AI — reached an efficacy belief only by passing through a social relationship. Under graduate education conditions, students with optimistic and innovative orientations toward AI reported more substantive supervisor support, and that support in turn tracked their confidence in designing studies, analyzing data, and writing. Readiness facilitators were associated with support at β = 0.460, and the indirect path to research self-efficacy was β = 0.201 — larger than the trust route.

The authors explain this with Social Cognitive Theory rather than with tool exposure: self-efficacy grows from mastery experiences, vicarious learning, and verbal persuasion, and a supervisor who redirects attention to conceptual and methodological problems supplies all three. AI may free supervisory time from low-level questions, but the efficacy gain depends on what the supervisor does with that time. Both constructs were self-reported in the same instrument.

Caution that builds trust instead of eroding it

The most surprising finding is that readiness inhibitors — discomfort with AI design, insecurity about reliability and over-reliance — predicted higher AI trust. The authors reject the reading that anxious students avoid AI. Their sample consisted only of regular AI users, so the sample was selected on use; within it, worries about inaccurate output, privacy, and plagiarism functioned as critical evaluation rather than refusal.

This aligns with trust calibration: justified trust rests on a realistic model of what a system can and cannot do, not on faith. Students who notice AI's fallibility may interrogate outputs more and end up with better-calibrated confidence. The authors counter that too much trust can degrade research reasoning, and that perceived confidence is not evidence of better practice.

Why a discipline-uniform model is a design opportunity

Respondents were split into a humanities/social sciences group and a science/engineering/medicine group on the soft–hard discipline distinction: interpretive, text-centered work with individualized supervision on one side, experimental, instrument-centered work with laboratory teams on the other. Measurement invariance was partial but sufficient for comparison, and no path difference was significant.

Two within-group patterns appeared: readiness facilitators predicted self-efficacy directly only in the humanities/social sciences group, and AI trust predicted it only in the science/engineering/medicine group. The authors label these exploratory, since the corresponding coefficient differences were not significant. For administrators, a single support model is defensible; targeted emphases — trust-building in lab disciplines, direct readiness work in text disciplines — remain hypotheses.

What this means for practice

  • Instructors: Treat AI readiness as an entry ticket to a conversation, not an outcome. The efficacy gain came from supervisors spending reclaimed time on research design and methods, so build those exchanges into supervision rather than assuming tool fluency transfers.
  • Administrators: Fund supervisor capability in AI-mediated research — methodological guidance, timely feedback, and the judgment to know when an AI-assisted draft is weak — since supervisor support carried the largest direct association with student confidence.
  • Designers: Build AI literacy training around output interrogation (accuracy, privacy, over-reliance) rather than around enthusiasm. In this sample, students who voiced those concerns reported higher, not lower, trust.
  • Researchers: Measure AI use behavior and research performance, not only self-reported confidence, before treating readiness programs as proven.

Limitations

  • Convenience sample of master's students in one country, restricted to those who already used AI regularly; both features limit generalization to non-users and to other systems.
  • Cross-sectional, entirely self-reported data cannot support causal claims about readiness, support, trust, or efficacy.
  • Supervisor academic support and research self-efficacy are conceptually close, which the authors concede may have inflated their association.
  • The discipline grouping is binary, hiding variation within the humanities/social sciences and within science/engineering/medicine.
  • Prior research experience, year of study, grades, digital literacy, and AI-use frequency were not modeled as covariates.

Citation

Luo, Y., Fan, Y., Wang, T., Xu, Q., & Zhao, W. (2026). AI Technology Readiness, Trust, and Supervisor Academic Support in Relation to Graduate Students’ Research Self-Efficacy. Behavioral Sciences, 16(8), 1313.

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.