AI evidence extraction

The Reliance Negotiation Framework: A Dynamic Process Model of Student LLM Engagement in Academic Writing

Shahin Hossain · ArXiv.org · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Mixed-methods design combining qualitative semi-structured interviews (n=14 interview participants), open-response qualitative coding (n=361 survey respondents), and quantitative hierarchical regression with moderation analysis.

Sample

N = 375, 9 groups

Primary method

Hierarchical regression analysis with moderation testing; multinomial logistic regression; intraclass correlation coefficient (ICC) analysis; qualitative thematic coding with frequency counting; cross-sectional analysis with predictor blocks. Parent study (Hossain, 2026) employed quantitative methods; current framework paper synthesizes qualitative interview data (n=14) and open-response survey responses (n=361) with reference to parent study's quantitative findings.

Main result

The study found that "observable LLM reliance behaviors are the outputs of an ongoing negotiation process" driven by four concurrent inputs (perceived benefits, perceived risks, ethical commitments, and situational demands). The research identified that "13.0% of respondents (n = 47) described a deliberate, values-grounded refusal to use generative AI for academic writing, not from low literacy or low exposure, but from categorical ethical commitment." Additionally, "greater prior LLM exposure predicts higher reliance intensity (β = .308) but also, counter-intuitively, membership in all non-Strategic reliance types in the multinomial model (ORs: 2.89-6.83)."

Reports effect sizes and confidence intervals.

Research paradigm

Mixed methods (qualitative-dominant with quantitative components)

Author conclusions

The authors conclude: "The era of generative AI in higher education has arrived without waiting for institutions to develop coherent responses. The students who most need those responses (first-generation students negotiating with amplified efficiency benefits and encounter-based literacy pathways, students at minority-serving institutions where the efficiency-atrophy tipping point operates asymmetrically by preparation level) cannot afford to wait for the field to generate the longitudinal evidence that would confirm or revise what this framework has proposed. The RNF does not ask institutions to wait. It asks them to act on the best available theoretical account of how student reliance decisions are made and how they develop, and to build the pedagogical infrastructure, policy environments, and equity-conscious interventions that treat those processes not as threats to be suppressed but as developmental capacities to be cultivated."

Risk of bias

Selection bias: Single-institution sample from MSI may not generalize; Social desirability bias: Retrospective self-report data vulnerable to bias; Attrition risk: Qualitative strand (n=361) outcome of larger sample not fully specified; Confounding: No random assignment; observational design; Measurement bias: Lack of validated abstention measures conflates three distinct populations; Selection bias: Single institution (MSI) sample; 33.5% first-generation students may not be representative; Retrospective recall bias: Interview participants recounted past reliance decisions from memory; Social desirability bias: Self-report data on academic integrity behaviors subject to misrepresentation; Lack of triangulation: No process-visible data (screen recordings, keystroke logs) to verify self-reported reliance behaviors; Attrition not reported: No information on dropout or non-response rates; Single-informant data: Only student reports; no instructor or institutional triangulation; Social desirability bias in self-report qualitative data; Retrospective recall bias in interview accounts; Single-institution design (limits generalizability); Selection bias in interview participant recruitment (n=14); Potential temporal order confounding in cross-sectional design; Single informant perspective (no triangulation with independent observers)

Limitations

The authors state: "The most theoretically consequential limitation is the absence of validated abstention measures
The Two-Model Architecture's most important practical claim (that abstention-mode students respond to institutional interventions differently from negotiation-mode students and low-exposure students) cannot be tested until these three near-zero-reliance populations can be quantitatively distinguished." Additionally, "the absence of longitudinal data
within-student variability rests on retrospective accounts" and "Self-report subject to social desirability bias
no triangulation with process data."

Open questions raised

Longitudinal data on recursive feedback mechanisms and within-student variability
Process-visible data (screen recordings, keystroke logs) to triangulate retrospective qualitative findings
Validated abstention measures distinguishing strategic restrainers from principled non-users from low-exposure students
Multi-site comparative designs testing MSI-specific equity dynamics
Prospective experience-sampling methodology across multiple assignments
Empirical tests of four falsifiable predictions constituting the RNF research agenda

Extracted from: pdfAgreement 55%

Explore related topics

Responsible AI Use Human-AI Collaboration Mixed Methods