Governing Generative AI in Higher Education

Patterns in University AI Policies Across R1 and R2 Institutions

Andrew Katz

Virginia Tech

2026-03-23

Motivation

The Landscape

• Generative AI tools are now pervasive in higher education
• Universities are responding with policies governing student, faculty, and staff use of these tools
• A growing body of work has analyzed these policies across institutions (e.g., Ghimire & Edwards, 2024; McDonald et al., 2025; Parker et al., 2025; Wang et al., 2024)
• But existing studies rely primarily on manual coding, top-down frameworks, and/or smaller corpora

Acknowledge prior work — several studies have now examined university AI policies at scale. Our contribution is methodological, not the first to look at this space.

Contributions of This Work

1. Structured extraction via ADICO — Using institutional grammar to extract operative rules as structured data, rather than manual content analysis

2. Scale — 367 policy documents across 226 R1 and R2 institutions, one of the largest corpora analyzed

3. Inductive codebook generation — LLM-assisted pipeline (GATOS) that builds a codebook bottom-up from the data, rather than applying predetermined categories

Research Questions

1. What actions are universities governing with respect to generative AI?

2. How do policies differ in regulatory posture (prohibitive vs. permissive)?

3. Do patterns vary by institutional characteristics (R1/R2, public/private, region)?

4. What themes emerge across the policy landscape?

The Corpus

• 367 policy documents from Carnegie-classified R1 and R2 institutions
• Sourced from publicly available university websites, syllabi portals, and policy repositories
• Collected throughout 2025 *

226 unique institutions represented

328 documents matched to Carnegie data (89%)

50 U.S. states covered

Methods

Analytical Pipeline Overview

A two-stage approach combining structured information extraction with inductive qualitative coding:

1. Stage 1 — ADICO Extraction Extract operative rules from policy text using LLM-assisted structured extraction

2. Stage 2 — Codebook Generation Inductively generate a qualitative codebook from the extracted rules using clustering and LLM-assisted coding

PDF Documents
     │
     ▼
Text Extraction
     │
     ▼
Chunking (250 tokens)
     │
     ▼
ADICO Clause Extraction
  (Attribute + Deontic + Aim)
     │
     ▼
Embedding + Clustering
     │
     ▼
Codebook Generation
     │
     ▼
Codebook Application
     │
     ▼
Analysis & Visualization

Stage 1: ADICO Clause Extraction

Each policy document is chunked and processed by an LLM to extract one operative rule per chunk in ADICO format:

Component	Description	Example
Attribute	Who the rule targets	students
Deontic	Regulatory modality	must_not
Aim	The governed action	use generative AI tools to write an essay

1,266 operative rules extracted from 367 documents (after filtering empty and ambiguous extractions)

Framework

ADICO: Crawford & Ostrom (1995). Institutional grammar.

Explain the ADICO framework briefly — Attribute (who), Deontic (modality), Aim (action). We only use these three components, not the full five.

Deontic Categories

Six regulatory modalities capture the spectrum from prohibition to encouragement:

Category	Meaning	Count
must_not	Prohibited	377
must	Mandatory requirement	388
should	Advisory recommendation	272
should_not	Advisory against	86
may	Explicitly permitted	78
encouraged	Actively promoted	65

Stage 2: Inductive Codebook Generation

The 1,266 extracted aims were processed through the GATOS pipeline:

1. Embed each aim using sentence-transformers
2. Cluster similar aims (agglomerative, target size = 5)
3. Generate codes — LLM suggests thematic labels for each cluster
4. Evaluate novelty — cosine similarity + LLM judgment deduplicates codes
5. Generate themes — group accepted codes into higher-level categories
6. Apply codebook — LLM assigns codes back to each aim

57 codes accepted

270 candidate codes evaluated

4 themes generated

70.3% of aims successfully coded

Pipeline

GATOS: Generative AI-enabled Theme Organization and Structuring (pygatos v0.1.0)

GATOS Pipeline Overview

Walk through the pipeline steps. Emphasize that this is an inductive process — codes emerge from data, not imposed a priori.

Codebook Saturation

Left: Cumulative accepted codes plateau around ~50. Center: Rolling acceptance rate declines over time. Right: The growing ratio of rejected-to-accepted codes indicates saturation.

Codebook Growth

Green: accepted codes (57). Red: rejected as duplicates (263). The widening gap indicates the codebook reached conceptual saturation.

Findings

What Are Universities Governing?

Top 25 codes by frequency. “AI-generated academic writing” dominates (146 aims), followed by documentation, data handling, and verification requirements.

Key takeaway: academic writing dominance is striking — it’s more than double the next most common code. This tells us where the policy conversation is focused.

The Prohibitive–Permissive Spectrum

Red (left) = prohibitive; Blue/green (right) = permissive. Most policies are prohibitive for academic writing but permissive for documentation and verification.

This is one of the most important slides. The diverging bar chart makes the regulatory posture visible at a glance — prohibition dominates writing, but other areas are more balanced or permissive.

Who Do Policies Target?

Row-normalized percentages. Students face the most prohibitions (41% must_not). Faculty receive more advisory and encouragement. Instructors are predominantly guided by “should” (49%).

R1 vs R2: Code Adoption Rates

Proportions normalized by group size (R1: 220 docs, R2: 98 docs). R2 schools are more likely to address AI-generated academic writing (37% vs 23%). R1 schools lead on data handling.

Public vs Private Institutions

Proportions normalized by group size (Public: 213 docs, Private: 105 docs). Adoption rates are broadly similar, with small differences in verification and tool-specific policies.

The Four-Way View

Code adoption % across R1/R2 x Public/Private. R2 Public institutions show the highest rates for academic writing (38%) and interaction documentation (23%).

Which Codes Co-occur?

Document-level co-occurrence. “AI-generated academic writing” frequently co-occurs with “Uploading Content to AI Tools” (22 docs) and “Confidential Data Handling” (23 docs).

Discussion & Future Work

Key Takeaways

1. Prohibition dominates — “AI-generated academic writing” is the most common policy topic, and it is overwhelmingly prohibitive

2. Students bear the regulatory burden — 41% of student-targeted rules are prohibitions, while faculty receive more encouragement and advisory guidance

3. R2 institutions tended to be more prescriptive — Higher rates of specific rules about academic writing and documentation requirements

4. R1 institutions tended to emphasize data governance — More attention to confidential data handling, uploading restrictions, and data classification

5. Public and private institutions are similar — Fewer differences than expected in code adoption rates

Summarize the five key findings. Emphasize the student burden finding — this is likely to resonate with audiences concerned about equity.

What the Codebook Misses

• There were 376 uncoded aims
• Common terms: “Cite,” “disclose,” “plagiarism,” and “assignment” are prominent
• Possible that citation/attribution requirements are a codebook gap (or information extraction gap)

Limitations

• Single rule per chunk — ADICO extraction captures one operative rule per text chunk; policies with multiple rules in close proximity may be under-represented
• Codebook coverage — 29.7% of aims went uncoded, particularly citation and attribution requirements
• Snapshot in time — Policies are evolving rapidly; this corpus reflects 2025 docuemnts
• LLM-assisted coding — While validated against codebook criteria, human inter-rater reliability has not been assessed

Future Work

• Full-document pipeline — Run information extraction directly on raw policy text (not constrained by ADICO schema) to capture the full breadth of policy content
• Codebook refinement — Address the citation/attribution gap; iterate on cluster size and novelty thresholds
• Longitudinal analysis — Track policy evolution as institutions update their guidance
• Cross-institutional comparison — Deep case studies of outlier institutions with highly comprehensive or highly restrictive policies
• Validation — Human coding of a subset to assess LLM coding reliability
• Online Database — Create an online database of AI policies for researchers and practitioners
• Github repository — Create a GitHub repository for the codebase for other policy researchers to use (needs to integrate docinfo https://github.com/andrewskatz/docinfo and pygatos https://github.com/ask-at-z/pygatos)

Thank You

Questions?

Andrew Katz | Virginia Tech | [email protected] | ideeaslab.com

References

References

Crawford, S. E., & Ostrom, E. (1995). A grammar of institutions. American political science review, 89(3), 582-600.

Ghimire, A., & Edwards, J. (2024, July). From guidelines to governance: A study of AI policies in education. In International Conference on Artificial Intelligence in Education (pp. 299-307). Cham: Springer Nature Switzerland.

Katz, A., Fleming, G. C., & Main, J. B. (2026). Thematic analysis with open-source generative AI and machine learning: A new method for inductive qualitative codebook development. Humanities and Social Sciences Communications.

McDonald, N., Johri, A., Ali, A., & Collier, A. H. (2025). Generative artificial intelligence in higher education: Evidence from an analysis of institutional policies and guidelines. Computers in Human Behavior: Artificial Humans, 3, 100121.

Parker, L., Loper, A. J., Hayes, J., Karakas, A., White, S., & Hallman, H. (2025). Comparative analysis of artificial intelligence policies in universities across five countries. Discover Computing, 28(1), 267.

Wang, H., Dang, A., Wu, Z., & Mac, S. (2024). Generative AI in higher education: Seeing ChatGPT through universities’ policies, resources, and guidelines. Computers and Education: Artificial Intelligence, 7, 100326.