- The paper establishes a dual-level analysis comparing institutional and course-level GenAI policies in US R1 institutions, revealing alignment and divergence.
- It employs a robust qualitative coding protocol across 116 institutional policies and 98 CS syllabi to identify trends in GenAI usage and restrictions.
- Findings indicate that comprehensive institutional guidelines correlate with more detailed course policies, highlighting ethical challenges and shifting pedagogical practices.
Comparative Analysis of Institutional and Course-Level GenAI Policies in Higher Education
Introduction
This paper presents a systematic analysis of generative AI (GenAI) guidelines promulgated by research-intensive (R1) U.S. higher education institutions and compares these with actual policy statements found in computer science (CS) course syllabi. As GenAI applications, notably ChatGPT, rapidly proliferate among students, institutions have sought to codify their stances on usage. The study distinguishes itself through its dual-level analysis, spotlighting the alignment and divergence between institutional-level policies and their translation into course-level implementation within computing education. The findings have particular salience for computing educators tasked with reconciling institutional mandates, contingent technological affordances, and the evolving ethical and pedagogical challenges inherent to GenAI.
Methodological Overview
Leveraging a dataset drawn from 116 R1 institutions' public GenAI policies and 98 CS course syllabi across 54 of these institutions, the research employs a rigorous qualitative coding protocol with established inter-rater reliability. The institutional guidelines were comprehensively catalogued according to thirteen primary codes, while syllabi were open-coded into sixteen categories capturing pedagogically salient course-level distinctions. The comparative analysis is framed by questions addressing (1) the implementation fidelity of institutional guidance at the course level and (2) patterns of convergence and divergence in GenAI policy across the two strata.
Key Findings
Alignment and Policy Coverage
The study finds that the breadth of institutional GenAI policy coverage (determined by the number of relevant codes present) is positively associated with the presence of course-level GenAI policies. Comprehensive institutions (with 8-11 policy codes) are more likely to exhibit course-level guidelines (48.3–71.4%), whereas less comprehensive institutions show more ad hoc, bottom-up development of course policies. This establishes a partial trickle-down effect: institutional comprehensiveness increases the probability, but not the certainty, of course-level implementation.
Range and Specificity of Policy
- Range of Consent: Although over half of institutions stipulate GenAI policies with granularity (such as "embrace," "limit," or "prohibit"), course syllabi are markedly more likely to issue explicit, categorical rules, with 50% outright banning GenAI use and 41% permitting partial, narrowly defined applications.
- Encouragement vs. Discouragement: At the institutional level, 63% encourage GenAI adoption in some capacity, whereas explicit encouragement at the course level is rare (fewer than 20% of syllabi). Notably, explicit prohibition remains prevalent at the course level, reflecting faculty risk aversion amid unresolved academic integrity concerns.
- Citation and Disclosure Requirements: Institutional guidelines commonly invoke citation (38%) and transparency mandates; course syllabi, however, operationalize these with greater stringency: 83% prescribe disclosure, and 65% classify undisclosed GenAI use as an academic integrity violation.
- Domain-Specific Guidance: Institutional guidelines often highlight STEM fields, especially CS, as primary loci of GenAI concern, yet practical course policy variations are largely a function of instructor discretion.
Idiosyncrasies and Gaps
- Anthropomorphism and Specification: Nearly 40% of syllabi anthropomorphize GenAI, describing tools as "assistants" or "classmates," a nuance largely absent from institutional guidance.
- Tool-Specificity: Syllabi frequently reference concrete tools (ChatGPT, Copilot, Bard), a level of operational specificity not present in institutional policies.
- Ethics and DEI: While institutional guidelines increasingly reference diversity, equity, and inclusion (DEI), this attention is rarely mirrored in course-level documents.
- Curricular Guidance: Institutions often advise faculty to consider pedagogical reconfiguration in light of GenAI, but such curricular reflection is rarely detailed in syllabi, likely due to audience and function constraints.
Theoretical and Practical Implications
The analysis elucidates a persistent gap between institutional intent and course-level enactment. The overwhelming institutional narrative is cautious optimism: GenAI is positioned as a tool to be harnessed, provided issues of privacy, accuracy, and academic integrity are addressed. Yet, course-level policies remain highly restrictive, reflecting instructor uncertainty, insufficient institutional support for policy translation, and heightened perceived risks in programming education.
There are salient implications for the trajectory of computing education. As GenAI becomes omnipresent, assignments reliant on traditional auto-grading, routine programming, or pattern-based assessment risk obsolescence—unless accompanied by assessments that are robust to (or benefit from) GenAI augmentation. The study underscores the urgency of upskilling CS faculty in AI literacy, encompassing not merely technical fluency but nuanced understanding of code generation limitations, security vulnerabilities, and reliability issues unique to AI-assisted workflows.
Furthermore, the anthropomorphic framing observed in syllabi points towards evolving conceptual models of GenAI for teaching. This could inform the design of educational interventions that treat GenAI not simply as a tool, but as a participant in the learning process, with implications for metacognition, peer learning, and formative assessment.
Limitations
The dataset's restriction to R1 U.S. institutions limits generalizability to smaller or international institutions, and the reliance on publicly available documents may miss non-public channels of guidance. Temporal mismatch between data collection windows (policy vs. syllabi) may also skew alignment assessments due to the rapid evolution of GenAI guidance during the collection period. Finally, the study does not address policy implementation or efficacy from the perspective of experiential outcomes or compliance.
Future Directions
Future investigations could expand the disciplinary scope beyond CS, explore the lived experiences of faculty and students in policy implementation, and quantify the impact of differing GenAI policies on academic integrity outcomes and student skill development longitudinally. There is also scope for institutional collaborations with AI developers to enable transparent, auditable use of GenAI in educational contexts (such as watermarking), ensuring both learning integrity and innovation.
Conclusion
This comparative study establishes that while institutional GenAI policies in higher education increasingly articulate the educational potential of these technologies and emphasize responsible use, course-level implementation—particularly in computer science—is far more conservative and operationally specific. Significant divergences persist, especially around the details of tool deployment, encouragement, and the operationalization of academic integrity concerns. Closing the alignment gap will require investment in faculty training, discipline-specific curricular redesign, and policy frameworks that are adaptable to both rapid technical evolution and pedagogical nuance.
Reference: "A Comparative Analysis of Institutional and Course Generative AI Policies within Higher Education: Implications for Instruction in Computing Education" (2607.12296).