Human-in-the-loop GenAI advising for community college pathway decision-making: a design-based case study of requirement-group modeling and privacy-preserving student-support design

Technical Summary: Human-in-the-Loop GenAI Advising for Community College Pathway Decision-Making

Problem Statement
Generative Artificial Intelligence (GenAI) research in higher education has predominantly focused on tutoring, writing support, feedback generation, and assessment. A significant gap exists regarding the application of GenAI to academic advising and pathway planning, particularly within community college contexts. In these settings, learners navigate complex, flexible curricular structures involving fixed requirements, elective buckets, concentration choices, and transfer articulations. Current advising tools often reduce these structures to "flat checklists," which risks misrepresenting program logic, narrowing perceived student options, and failing to support learner agency. Furthermore, the integration of GenAI into student-success operations raises critical concerns regarding privacy, data governance, and the appropriate level of automation in high-stakes decision-making. There is a lack of established design principles for how GenAI can interpret flexible requirements, support "what-if" planning, and maintain privacy boundaries without over-automating consequential decisions.

Methodology
This study employs a design-based case study approach to investigate the development of a pathway-planning prototype and an associated advising interaction model at Ivy Tech Community College of Indiana. The research does not measure student outcomes, retention rates, or workload reductions; instead, it analyzes the design process itself to derive transferable principles.

• Data Sources: The analysis is based on a corpus of non-student materials collected over an eight-month design period, categorized into three source classes:
  1. Public Curricular Documents: Published program requirements, course descriptions, and catalog materials used to identify structural flexibility and ambiguity.
  2. Prototype Artifacts: Requirement schemas, pathway logic documents, interface mockups, architecture diagrams, and prompt templates used to operationalize curricular flexibility and GenAI roles.
  3. Practitioner Design Memos: Reflective documentation, architecture rationales, and workflow redesign notes tracing the justification for specific design and governance decisions.
• Analysis Process: The author utilized a deductive-inductive coding logic. Deductive constructs (curricular representation, learner agency, human oversight, privacy/governance) were applied alongside inductive coding to capture case-specific patterns (e.g., wildcard handling, ambiguity classes). Principles were retained only when supported across multiple source classes or when a single artifact documented a decision with clear governance consequences.
• Constraints: The study explicitly excludes individual student records and personally identifiable information (PII). GenAI tools were used only for manuscript preparation (literature organization, outlining, drafting), with all claims and citations verified by the author.

Key Contributions and Design Principles
The study proposes a Human-in-the-Loop GenAI Advising Framework organized around seven interacting layers: public catalog data, requirement-group modeling, the GenAI interaction layer, human advisor judgment, learner agency, privacy-preserving governance, and workforce alignment. From this framework, the study derives eight specific design principles:

1. Model Requirements as Flexible Groups: Academic requirements should be modeled as "requirement groups" (e.g., option sets, elective buckets, conditional alternatives) rather than flat, mandatory course checklists. This preserves the educational reality that multiple valid pathways exist within a single credential.
2. GenAI as Explanation, Not Record: GenAI should function as an "explanation and option-framing layer" that translates structured logic into plain language and compares scenarios. It must not serve as the institutional "system of record" or an autonomous decision-maker.
3. Architectural Data Separation: Public curricular data (catalogs, requirements) must be strictly separated from protected student-level records. A functional advising prototype can operate using public data alone, minimizing privacy risk and establishing clear governance boundaries before any extension to student-specific data.
4. Human Review for Consequential Recommendations: Human advisors must retain accountability for final guidance, particularly for recommendations affecting time-to-degree, cost, eligibility, or program direction. The system is designed for "reviewability" rather than automation closure.
5. Support for Learner Agency: Interactions should facilitate self-regulated planning by helping students understand trade-offs, sequence constraints, and recovery pathways, rather than merely issuing instructions.
6. Non-Deficit Language: Advising language should avoid deficit-centered or surveillance-heavy framing (e.g., "risk scores") in favor of constructive concepts like "momentum preservation" and "pathway recovery."
7. Visible Uncertainty: The system must explicitly expose ambiguity in curricular interpretation (e.g., underspecified catalog statements) rather than suppressing it with false certainty. Interfaces should distinguish between clear matches, likely matches requiring review, and ambiguous cases.
8. Workforce Alignment without Reductionism: Workforce relevance should inform pathway explanations but must not reduce advising to a narrow labor-market sorting mechanism that overrides student goals and exploration.

Results and Significance
The study concludes that effective GenAI integration in advising requires a shift from treating requirement interpretation as a data-parsing problem to treating it as a pedagogical design problem. The primary result is a conceptual and operational framework that allows institutions to deploy GenAI for pathway intelligibility while preserving learner agency and privacy.

The paper claims four specific contributions:

1. Extension of Research: It moves human–GenAI interaction research from classroom-centric tasks (writing, feedback) into the domain of advising and pathway planning.
2. Conceptual Innovation: It introduces "requirement-group modeling" as a more educationally faithful representation of community college pathways than flat logic.
3. Architectural Proposal: It demonstrates a privacy-preserving governance model grounded in the explicit separation of public curricular information from protected student data.
4. Practical Design Vocabulary: It offers a set of principles for institutions seeking to use GenAI in high-governance settings without replacing human judgment or making unsupported claims about efficacy.

Modesty of Claims
The authors explicitly state that this study does not demonstrate improved retention, completion rates, or advising efficiency. The contribution is foundational rather than empirical regarding student outcomes. The design principles are derived from artifact analysis and workflow modeling, not from a formal co-design trial or implementation with learner-use data. The paper argues that this privacy-conscious, human-in-the-loop framework is a necessary precondition for any future efficacy studies to be trustworthy, rather than a proven solution in itself.