Are Students Dependent on AI in Writing Courses? Analyzing Factors Influencing Dependence on Generative AI Through the I-PACE Model

Synthesis: Liu, Zhuang, and Wang (2026) apply the I-PACE model of problematic use to generative AI dependence in undergraduate academic writing. In a mixed-methods study at two Chinese universities — a survey analyzed with structural equation modeling plus eight interviews — academic stress emerges as the strongest predictor of dependence, AI Literacy as a protective factor, and social influence as consequential only through perceived Trust. Students knew AI's risks yet kept using it strategically, citing efficiency, peer pressure, and ambiguous institutional rules.

Key Findings

• Academic stress is the strongest driver of dependence. Students under the heaviest workload pressure were most likely to lean on AI to produce writing, and had trouble separating the academic task from the assistance.
• AI literacy is a protective factor. Students who understood how AI works, its limits, and its social consequences were less prone to dependence, apparently because literacy supported critical distance.
• Trust, not usefulness, is the pathway to dependence. Students who could name AI's failures still trusted its output and depended on it more; awareness of error did not become skepticism in practice.
• Peer influence acts through trust, not directly. Social influence had no significant direct effect on dependence, but strongly predicted perceived trust, which in turn predicted dependence — conformity matters mainly by making AI feel credible.
• Academic self-efficacy does not by itself reduce stress. Contrary to prior work, confidence in one's academic ability was unrelated to stress until perceived usefulness entered the picture; the buffering effect appeared only for students who saw AI as useful.
• Students use AI strategically and call it defensible. Interviewees described mixing tools and rewriting output to evade detection, framed as responsible use, and blamed the absence of any clear institutional definition of compliant AI use.

Study Design & Method

A sequential mixed-methods design in a required research-methods writing course for journalism majors. In the quantitative phase, 266 undergraduates at universities in Fujian and Guangdong completed an online questionnaire; after screening for incomplete, straight-line, and implausibly fast responses, 229 were analyzed. The survey measured academic self-efficacy, AI literacy, academic stress, perceived trust, social influence, and perceived usefulness with adapted five-point Likert scales, back-translated and pre-tested. Relationships were estimated through structural equation modeling, with bootstrapping for non-normal data and a common-method-bias check.

The qualitative phase drew on open-ended survey responses and semi-structured interviews with eight volunteers selected for heterogeneity in academic performance. Interviews covered their experience of AI-assisted writing, motives for heavy use, and their sense of dependence; transcripts were analyzed thematically by two researchers who coded independently.

What this means for practice

• Instructors. Break long writing assignments into staged checkpoints with feedback, because academic stress was the strongest predictor of AI dependence and students under the heaviest workload pressure leaned most on AI to produce text.
• Instructors. Teach AI Literacy as critical distance rather than tool operation: students who understood AI's limits and social consequences were less dependent, whereas students who could name the tool's failures still trusted its output.
• Administrators. Publish an explicit, example-based definition of compliant AI use and make responsible use the easier path than concealment in assessment design; the eight interviewees mixed tools and rewrote output to evade detection while framing it as responsible, citing the absence of any clear institutional rule.
• Instructors. Address peer norms and trust calibration together, because social influence reached dependence only through perceived trust — conformity made AI feel credible rather than simply modeling use.

Limitations

• Cross-sectional, single-culture sample of Chinese college students in academic writing courses; causal claims and generalizability to other settings are limited.
• Self-reported measures of dependence, stress, and literacy; no objective usage data, so dependence is measured as perception rather than behavior.
• Small qualitative sample of eight interviewees, limiting the breadth of thematic claims.
• I-PACE was designed for hedonic apps such as gaming and social media; extending it to a study tool carries assumptions about what dependence means in academic work.

Citation

Liu, L., Zhuang, M., & Wang, J. (2026). Are Students Dependent on AI in Writing Courses? Analyzing Factors Influencing Dependence on Generative AI Through the I-PACE Model. Frontiers in Psychology, 17, 1905037.