Structuring Transparency: Developing Domain-Specific Generative AI Declaration Frameworks in Higher Education

Technical Summary: Structuring Transparency in GenAI Declarations

Problem Statement
The rapid integration of Generative AI (GenAI) tools into higher education has created a tension between upholding academic integrity and preparing students for professional environments where AI use is normalized. Current institutional responses often rely on generic, binary declarations (e.g., a single checkbox asking "Did you use GenAI?"). The authors argue that these coarse mechanisms fail to capture the nuanced, multi-stage application of GenAI in academic tasks. Consequently, binary declarations act as poor integrity safeguards, offer limited pedagogical value, and fail to distinguish between legitimate assistance (e.g., brainstorming or debugging) and academic misconduct (e.g., ghost-writing or full code generation). Furthermore, existing taxonomies of GenAI usage are primarily analytic instruments for researchers and are not directly integrated into student assessment workflows.

Methodology and Framework Design
This paper presents a design artefact: a domain-specific framework for structuring GenAI declarations within Computer Science (CS) assessments. The methodology involves translating an existing, systematic taxonomy of GenAI-mediated activities (proposed by Bouvier et al. [1]) into concrete, task-specific disclosure instruments.

The framework is guided by five design principles:

1. Granularity: Moving beyond binary yes/no to distinguish between multiple activity types.
2. Task Specificity: Providing distinct structures for writing-focused and coding-focused assessments.
3. Intensity Measurement: Requiring students to indicate the extent of use (Minor, Moderate, Extensive) for each activity.
4. Prompt-Level Transparency: Mandating brief explanations and example prompts to encourage self-reflection.
5. Reflective Framing: Positioning the declaration as a description of process rather than an admission of guilt.

The framework operationalizes these principles through two parallel structures:

• Writing-Focused Structure: Designed for reports and literature reviews, this structure categorizes usage into six areas: Planning & Structure, Textual Content Generation, Research & Analysis, Revision & Polish, Visual Content, and Evaluation & Feedback.
• Coding-Focused Structure: Designed for programming assessments, this structure categorizes usage into five areas: Planning & Design, Code Generation, Code Improvement, Understanding & Learning, and Documentation & Reporting.

In both structures, students must declare the presence and extent of GenAI use for each category, accompanied by a short explanation and example prompts. The "Extent" scale is defined as: Minor (limited use for a small part), Moderate (repeated use but substantive decisions remain the student's), and Extensive (integral to the final submission).

Key Contributions
The paper contributes a specific design artefact and a theoretical position:

1. A Domain-Specific Declaration Framework: A structured mechanism that adapts high-level GenAI taxonomies into practical assessment tools for CS, distinguishing between different cognitive and developmental stages of work (e.g., separating code generation from code improvement).
2. A Shift in Declaration Philosophy: The authors argue for shifting declarations from compliance artifacts to reflective instruments. By making the boundary between acceptable assistance and misconduct "legible" at the point of use, the framework aims to reduce the perception that declaring AI use is equivalent to admitting wrongdoing.
3. Complementarity to Existing Policies: The framework is designed to sit below policy-level instruments like the AI Assessment Scale (AIAS). While the AIAS defines what is permitted, this framework captures how it was used, providing the granular data educators need to adapt their practice.

Results and Claims
As a design contribution, the paper does not present empirical results from a deployed trial. Instead, it offers a theoretical and structural argument supported by the limitations of current practices documented in related literature (e.g., Gonsalves [7] on non-compliance due to fear of penalties; Maguire et al. [8] on the lack of granularity in free-text declarations).

The authors claim that this framework:

• Clarifies Boundaries: Helps students distinguish between legitimate learning aids (e.g., understanding error messages) and authoring tools (e.g., generating core solution code).
• Enhances Pedagogical Insight: Provides educators with actionable data (e.g., a cohort declaring "Extensive" use in Code Generation vs. "Minor" use in Understanding) to inform scaffolding and assessment redesign.
• Prepares for Professional Practice: Aligns academic assessment with professional environments where documenting GenAI workflows is becoming an essential job requirement.

Significance and Limitations
The paper positions this domain-specific approach as a necessary complement to broader institutional policies to foster honest assessment and AI literacy. It acknowledges significant limitations, primarily the reliance on accurate self-reporting. The authors note that novice programmers may suffer from an "illusion of competence," misattributing GenAI contributions to their own work, and that the definition of "GenAI use" can be ambiguous (e.g., IDE autocomplete vs. conversational agents).

The authors conclude that while the framework is developed for Computer Science, the underlying approach of task-specific disclosure is adaptable to other disciplines. Future work is proposed in three strands: empirical evaluation against observed workflow data, pedagogical evaluation of student reflection and psychological safety, and cross-disciplinary adaptation to test the generalizability of the categories.