AI-mediated research agency formation in higher education: Autonomy, self-efficacy and innovation in early-career scientific training

Synthesis: Han and Liu (2026) examined how generative AI reshapes the agency of early-career researchers in doctoral and postdoctoral training. They define AI dependence as reliance on intelligent tools as the primary source of cognitive support in judgment-intensive research tasks, and read its risk as epistemic delegation — the transfer of problem framing, methodological choice, and interpretive authority to the system. Among 420 astronomy doctoral and postdoctoral researchers in China, dependence was negatively associated with research autonomy and research self-efficacy, both positively associated with innovative research behavior; supervisory support weakened the negative associations with autonomy and innovation, while the self-efficacy interaction fell short of significance. The study shifts the debate from AI-enabled productivity to AI-mediated researcher formation.

Key Findings

1. AI dependence depresses research agency. Dependence correlated with lower autonomy (r = −.355) and lower self-efficacy (r = −.321), yet was unrelated to self-reported use frequency (r = −.051, p = .296); supervisory support correlated positively with both (r = .532 and r = .435).
2. Two mediated pathways link dependence to innovation. Indirect associations of −.115, 95% CI [−.159, −.076] through autonomy and −.069, 95% CI [−.104, −.038] through self-efficacy, with a non-significant direct coefficient (B = −.060, p = .158).
3. The autonomy pathway is stronger. In the item-level latent model, β = −.137 through autonomy versus β = −.085 through self-efficacy: perceived ownership of the research process is the principal route from dependence to innovation.
4. Supervisory support buffers selectively. The interaction was significant for autonomy (B = .077, p = .020) and innovative behavior (B = .092, p = .005) but not for self-efficacy (B = .066, p = .064).
5. Conditional indirect effects shrink under support. The autonomy-mediated association fell from −.149 at low support to −.077 at high support; the total indirect association fell from −.238 to −.125, about 47.5% in absolute magnitude.
6. Split-sample psychometrics. Parallel analysis retained one AI-dependence factor in one random half (n = 210) and one-factor CFA fitted the other (χ2(35) = 43.33, CFI = .990, RMSEA = .034); the five-factor item-level model fitted the full sample (χ2 = 818.95 (655), CFI = .979, RMSEA = .024).

AI Dependence as Epistemic Delegation

The paper's central move is to treat AI dependence not as tool use but as a redistribution of epistemic authority. Decomposing a question, choosing a method, and interpreting results are traditionally the researcher's own cognitive work; when an intelligent system becomes the primary source of cognitive support, those operations migrate to the machine — epistemic delegation, an educational risk because doctoral education must build, not outsource, independent judgment. Delegation differs from Cognitive Offloading and automation reliance, which describe transferred effort or reliance on machine recommendations; delegation concerns who frames a problem, decides what counts as evidence, and owns the final call.

Measuring Reliance in Astronomy Training

Astronomy was chosen as a critical case of data-intensive training, not a disciplinary endpoint. The anonymous questionnaire circulated through peer-based WeChat student, classmate, and academic communication groups — never through departments or supervisors — and collected no university, laboratory, or supervisor identifiers. Of 500 responses, the 420 classified as Normal were retained (84.0%), 25 were flagged for review and excluded, and 55 were removed; the sample comprised 237 PhD researchers and 183 postdoctoral researchers. No existing scale captured AI dependence in astronomy research training, so the authors built a 10-item dependence scale and seven-item measures of autonomy, self-efficacy, innovative behavior, and supervisory support, all on a seven-point Likert scale. Six experts rated item relevance (I-CVI = 1.00; S-CVI/Ave = .97), and the tests used 5000 percentile bootstrap resamples, controlling for stage, field, team size, experience, AI-use frequency, and formal AI training.

Ownership, Not Confidence, Carries the Association

A control-only model explained 5.0% of the variance in innovative research behavior; adding AI dependence raised it to 12.3% (ΔR² = .073, B = −.301), and adding the two agency resources produced the largest gain, R² = .478 (ΔR² = .355), with autonomy (B = .435) outweighing self-efficacy (B = .272), the dependence coefficient near zero, and support raising R² to .501. The item-level structural model reproduced this (χ2(656) = 858.33, CFI = .974, RMSEA = .027), with support positively associated with autonomy (β = .580), self-efficacy (β = .483), and innovation (β = .179). Two checks qualify it: an equal-parameter reverse-path model fitted slightly better, so the developmental ordering is not established, and no path differences appeared by academic stage (Δχ2(8) = 6.16, p = .629).

From Productivity to Researcher Formation

The contribution reframes the question: rather than asking whether generative AI makes early-career research faster, it asks how AI shapes who the researcher becomes, locating the actionable site of intervention in supervisory Scaffolding and Learner Identity formation.

What this means for practice

• Researchers. Read heavy reliance on generative AI as a research-agency signal, not just a workflow choice: decide which framing, method, and interpretation decisions stay with you and which have moved to the tool.
• Instructors. Build supervision around explanation and verification — require students to justify method choices and audit outputs — because support was the one condition that weakened dependence's negative links to autonomy and innovation.
• Faculty developers. Give doctoral supervisors protocols for epistemic supervision, since the autonomy pathway carried the stronger indirect effect and buffering was strongest where ownership and innovative action were concerned.

Limitations

• Cross-sectional design: mediation and moderation are modeled associations, and a slightly better-fitting reverse-path model means the direction from dependence to reduced autonomy and innovation is inferred, not confirmed.
• All measures are researcher self-report from peer-based WeChat groups among astronomy early-career researchers in China; the setting limits generalization and the absent institution identifiers prevent team-level analysis.
• Both focal measures were purpose-built: split-sample factor analysis and the expert review support a broad one-factor AI-dependence structure but call for independent-sample validation, and supervisory support measured general encouragement, feedback, and respect for independent judgment rather than AI-specific epistemic supervision.

Connected Concepts

• Learner Agency
• Generative AI
• Self-Efficacy
• Human AI Collaboration
• Higher Education
• Learner Identity
• Cognitive Offloading

Connected Articles

• From Research Assistant to Surrogate Supervisor: A Qualitative Study Exploring PGR Students' Diverse Uses of Generative AI — From Research Assistant to Surrogate Supervisor (doctoral students' GenAI uses)
• Academic Erasure: The Disappearance of Complexity Under AI-Supported Writing — Academic Erasure and complexity under AI-supported writing
• Smaller, Younger, and More Impactful: How AI-Assisted Writing Transforms Research Teams — How AI-assisted writing transforms research teams

Citation

Han, S., & Liu, P. (2026). AI-mediated research agency formation in higher education: Autonomy, self-efficacy and innovation in early-career scientific training. Computers and Education: Artificial Intelligence, 100674.