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

This article introduces the Reliance Negotiation Framework (RNF), a dynamic process model derived from a mixed-methods study of 382 undergraduates that reconceptualizes student LLM engagement in academic writing as an ongoing negotiation among perceived benefits, risks, ethical commitments, and situational demands, thereby addressing the limitations of existing static typologies and offering new insights for pedagogy and policy at minority-serving institutions.

The Gist

The Big Idea: It's Not a Switch, It's a Conversation

Imagine you are trying to decide whether to use a GPS app to drive to a new destination.

Old theories about students and AI (like Large Language Models or LLMs) treated this decision like a light switch. They thought:

• "Is the student 'on' (uses AI) or 'off' (doesn't use AI)?"
• "Are they 'smart' (AI literate) or 'dumb' (AI illiterate)?"
• "Did they 'adopt' the technology or reject it?"

Dr. Shahin Hossain argues this is wrong.

He says student behavior isn't a light switch; it's more like negotiating a dinner menu with a picky roommate. Every time a student faces a writing assignment, they don't just flip a switch. They have an internal conversation, weighing different factors against each other. Sometimes they say "Yes, use AI," and sometimes "No, I'll do it myself," not because their personality changed, but because the situation changed.

This paper introduces the Reliance Negotiation Framework (RNF), a new way to understand that messy, daily conversation.

---

The Four Ingredients of the "Internal Negotiation"

According to the framework, every time a student looks at a writing task, they are juggling four invisible balls. They are constantly asking themselves:

1. The "What's in it for me?" Ball (Perceived Benefits):

   • Analogy: "If I use this AI, will I save an hour of my life? Will it help me brainstorm cool ideas? Will my grade go up?"
   • The Catch: If the deadline is tomorrow, this ball gets heavy. If the deadline is next week, it gets light.

2. The "What could go wrong?" Ball (Perceived Risks):

   • Analogy: "If I use AI, will the teacher catch me? Will the AI lie to me (hallucinate)? Will I forget how to write myself because I'm lazy?"
   • The Catch: Students often worry about getting caught (immediate risk) but forget about forgetting how to write (long-term risk).

3. The "What do I believe?" Ball (Ethical Commitments):

   • Analogy: "Is using AI cheating? Does it feel wrong to me personally, even if no one is watching?"
   • The Catch: Some students are like strict referees; they won't use AI no matter what. Others are like flexible players who will use it if the rules allow.

4. The "What's the situation?" Ball (Situational Demands):

   • Analogy: "Is this a huge final exam or a practice quiz? Is the teacher strict or chill? Is this a science class or a history class?"
   • The Catch: A student might be very careful in a History class but very casual in a Math lab. The context changes the rules of the game.

The Result: The student's final decision (to use AI, not use it, or use it a little) is the outcome of this negotiation. It's not a fixed trait; it's a moment-by-moment calculation.

---

The Two Types of Players

The paper discovered something surprising: Not everyone plays the negotiation game.

1. The Negotiators (87% of students):

   • These students weigh the four balls above. They might use AI for a low-stakes blog post but not for a final exam. They change their minds based on the situation.
   • Analogy: They are like a chef tasting the soup and adding salt or pepper depending on how it tastes right now.

2. The Abstainers (13% of students):

   • These students have a "Hard Stop" rule. They have a deep ethical belief that using AI is wrong, period.
   • Analogy: They are like a vegetarian who walks into a steakhouse. They don't negotiate with the menu. They don't weigh the "benefits" of the steak against the "risks" of eating meat. They just don't eat it.
   • Why it matters: Current surveys often mistake these principled students for "lazy" or "unskilled" students who just haven't tried AI yet. The RNF says: Stop confusing them. They are making a different kind of choice entirely.

---

The "Feedback Loop" (The Video Game Analogy)

The paper also explains that this isn't a one-time decision. It's a video game with a save feature.

• The Loop: If a student uses AI and gets a great grade, they might feel, "Hey, this works!" and use it more next time.
• The Twist: But if they use AI and realize they didn't actually learn the material, or if they get caught, they might feel, "Whoa, that was risky," and use it less next time.
• The Problem: Sometimes, students get stuck in a loop where they use AI so much they stop learning (skill atrophy), which makes them feel more dependent on AI next time. It's a downward spiral.

---

Why This Matters for Schools (The "MSI" Context)

The study was done at a Minority-Serving Institution (MSI), where many students are first-generation college students. The paper argues that these students face a unique "efficiency trap."

• The Analogy: Imagine two runners. One is a pro athlete (well-prepared student); the other is a beginner (student with a preparation gap).
  • The pro athlete uses a GPS (AI) to find a slightly faster route. They still run the race.
  • The beginner uses the GPS to carry them. If they rely on the GPS too much, they never build the leg muscles they need to run the race on their own.
• The Inequity: Schools often punish the beginner for using the GPS, but they don't give them the training to build those muscles. The RNF suggests schools need to teach students how to use the GPS wisely (literacy) and why they might want to run without it sometimes (ethics), rather than just banning the GPS.

---

The Takeaway: What Should Schools Do?

The paper suggests schools stop trying to be "AI Police" (just catching cheaters) and start being "AI Coaches."

1. Don't just ban it: Banning doesn't stop the negotiation; it just makes students sneakier.
2. Teach the negotiation: Help students understand the four balls (Benefits, Risks, Ethics, Situation). Teach them when it's okay to use AI and when it's a trap.
3. Respect the "Abstainers": Don't force students who ethically refuse to use AI to use it.
4. Fix the "Feedback Loop": Design assignments where using AI doesn't replace the learning, but actually helps the student learn more.

In short: Students aren't robots that either "use AI" or "don't." They are humans constantly making tough choices based on their goals, their fears, their values, and their deadlines. To help them, we need to understand the conversation they are having with themselves.

Technical Summary

1. Problem Statement

Current theoretical frameworks for student engagement with Large Language Models (LLMs) in academic writing are insufficient because they treat reliance as a static trait, a binary adoption decision, or a fixed competency level. The author identifies three critical gaps in existing literature:

• Within-Student Variability: Typological models (e.g., Strategic vs. Dependent users) cannot explain why the same student exhibits different reliance behaviors across different tasks or contexts.
• The Developmental Paradox: Technology Acceptance Models (TAM) explain adoption but fail to account for post-adoption quality. Empirical data shows that increased LLM exposure often leads to habituation and less regulated use rather than increased sophistication.
• Principled Non-Use: Existing frameworks treat the deliberate, values-based refusal to use AI as a failure of engagement or low literacy. They cannot theorize "principled abstention" as a sophisticated, ethical decision, often conflating these students with low-exposure users.

The paper argues that student LLM engagement is not a stable disposition but a continuously negotiated process shaped by competing beliefs, values, and situational pressures.

2. Methodology

The study employs a sequential explanatory mixed-methods design conducted at a public Minority-Serving Institution (MSI) in the United States.

• Participants: 382 undergraduate students.
• Data Collection:
  • Quantitative: A survey ($N=382$) measuring motivational beliefs, AI literacy, and reliance behaviors.
  • Qualitative: 14 semi-structured interviews and three qualitative survey strands ($n=361$ open-response items).
  • Coding: 1,435 coded instances of student reasoning and behavior.
• Analysis:
  • Hierarchical regression and multinomial logistic regression to identify predictors of reliance types.
  • Thematic analysis of qualitative data to map ethical reasoning tiers and developmental trajectories.
  • Integration of findings to construct a theoretical architecture.

3. Key Contributions: The Reliance Negotiation Framework (RNF)

The RNF reconceptualizes LLM reliance as an ongoing negotiation among four concurrent inputs, producing outputs that recursively modify future decisions.

A. The Four Concurrent Inputs

1. Perceived Benefits: Derived from Expectancy-Value Theory (EVT). Includes time savings, idea generation, and quality improvement.
2. Perceived Risks: Also from EVT (Cost dimension). Disaggregated into four sub-types:
   • Academic Integrity: Detection risk.
   • Epistemic: Hallucinations/factual errors.
   • Cognitive-Developmental: Skill atrophy (often underweighted due to delayed costs).
   • Social: Peer disclosure norms.
3. Ethical Commitments: Based on Kohlberg's moral development.
   • Tier 1: Compliance-based (fear of detection).
   • Tier 2: Norm-dependent (adherence to instructor policy).
   • Tier 3: Principled (self-authored values).
4. Situational Demands: Based on Situative Cognition (Greeno). Includes time pressure, task stakes, disciplinary norms (STEM vs. Humanities), and institutional policies.

B. The Two-Model Architecture

The framework introduces a critical bifurcation to account for the 13.0% of students who refuse AI use based on categorical ethical commitments:

• Negotiation-Mode (87.0%): Students weigh the four inputs dynamically. Their behavior is context-sensitive and changes based on the relative weight of benefits vs. risks.
• Abstention-Mode (13.0%): Students operate under a lexical constraint. Their ethical commitment acts as a non-negotiable side constraint that forecloses the negotiation entirely, regardless of situational demands or perceived benefits.

C. Three Defining Properties

1. Recursive: Outputs (reliance behaviors) feed back into inputs. For example, experiencing a hallucination (output) recalibrates the perception of epistemic risk (input) for future tasks.
2. Situated: The negotiation is constituted by the specific affordances and constraints of the activity system (e.g., a STEM lab report vs. a humanities essay).
3. Developmental: Trajectories are not linear. Experience can lead to sophistication or habituation depending on whether the recursive loop activates risk recalibration or simply normalizes use.

4. Results and Empirical Findings

• Motivational Dominance: Motivational beliefs (EVT) were the strongest predictor of engagement ($\beta = .630$), but their effect is mediated by situational demands.
• The Efficiency-Atrophy Paradox: Greater prior LLM exposure predicted higher reliance intensity ($\beta = .308$) but also membership in non-strategic reliance types (Odds Ratios: 2.89–6.83). Experience did not automatically produce sophistication; without structured feedback, it produced habituation.
• Principled Abstention: 13.0% of respondents ($n=47$) explicitly refused AI use based on ethical grounds (e.g., data provenance, environmental impact). Standard instruments misclassify these students as low-literacy or low-use users.
• Case Trajectories:
  • Tariq: Demonstrated how changing situational demands (policy on disclosure) shifted his negotiation output from heavy use to selective disclosure, even with unchanged beliefs.
  • Destiny: Demonstrated the "efficiency-atrophy" paradox. She initially used AI heavily for grades (rational calculation) until she realized she wasn't learning (developmental risk activation), leading to a shift toward strategic restraint.

5. Significance and Implications

The RNF shifts the field from typology (categorizing students) to mechanism (understanding the negotiation process).

• For AI Literacy Pedagogy: Interventions should target negotiation accuracy (calibrating benefit/risk assessments) rather than just engagement frequency. Instruction must be discipline-specific (e.g., verifying legal citations vs. general facts) to develop "adaptive AI literacy."
• For Academic Integrity Policy:
  • Detection-centered policies only affect Tier 1 (compliance) students and are surveillance-dependent.
  • Literacy-centered policies improve competency but may not address ethical tiers.
  • Ethics-reasoning policies are the most durable, fostering Tier 3 principled reasoning that is stable even without enforcement.
• For Equity (MSI Context): The framework highlights that students at Minority-Serving Institutions (MSIs) often face a steeper "efficiency-atrophy tipping point." Due to preparation gaps, they rely on AI for compensation rather than augmentation. Without institutional scaffolding, the recursive loop amplifies skill substitution rather than skill development. Policies that restrict AI without providing literacy support disproportionately harm these students.

6. Future Research Agenda

The paper proposes four falsifiable predictions to validate the RNF:

1. Within-Student Variability: Reliance decisions will show low intra-class correlation (ICC < .50) across tasks within the same student.
2. Ethical Tier Advancement: Structured ethics interventions will reduce surveillance-sensitivity in reliance behavior ($d \ge 0.40$).
3. Literacy Intervention Effects: AI literacy training will reduce self-efficacy inflation among dependent users.
4. Institutional Moderation: The relationship between reliance intensity and skill substitution will be stronger at MSIs than at selective institutions.

Conclusion: The RNF provides a dynamic, recursive model that explains how and why students engage with AI, offering a path toward equity-centered policies that cultivate developmental capacities rather than merely suppressing behaviors.