Generative AI in Higher Education: Seeing ChatGPT Through Universities' Policies, Resources, and Guidelines

In plain words

Many universities in the United States have had to decide how students and teachers may use chatbots such as ChatGPT. This paper looks at what the top 100 American universities have published on that question: their official rules, teaching guides and support resources, as they stood in April 2024.

The authors read every document, sorted what they found into categories, and counted how often each category appeared. Most universities did not ban the tools. Many left the decision to each teacher, and most offered guides, sample syllabus wording or consultations. Concerns centred on ethics, accuracy and privacy. The authors suggest that teachers accept the tools, design assignments around learning goals, and rely on several kinds of evidence rather than on AI detectors alone.

TL;DR

• A qualitative study of official policies and resources from 104 U.S. universities in the top-100 list, collected up to April 2024 §4.1.
• Ban was chosen by none; most universities left the decision to instructors (57 of 104) §5.1 fig. 2.
• Resources are common, with shared articles most frequent (74 of 104) §5.2.2.
• No university treated AI-text detectors as fully reliable §5.2.3.
• Use it as a descriptive picture of the policy landscape, not as evidence of the effect of any policy.

Problem

Universities need to guide teachers and students on generative AI, but published research on how institutions themselves respond was limited at the time of the study §1 §2. The authors ask three questions: how U.S. universities manage and regulate ChatGPT and other GenAI tools; what resources and guidance they give teachers; and whether school ranking tier or academic specialization changes those stances and resources §3.

Method

This is a document study, not a model or an algorithm. Nothing is trained, and no computational procedure is specified beyond the counting and scoring below.

Step 1: Sample. The authors took the top 100 U.S. universities in the 2024 U.S. News Best National University Rankings, which gives n = 104 institutions §4.1. For each university they ran a search that combined keywords such as "GenAI policy" or "GenAI guidelines" with the university name §4.1.

Step 2: Inclusion. A document was kept if it came from an official university source (for example the provost's office, an academic senate, a teaching centre, or the library), and was either a university-wide policy on GenAI and academic integrity, or a guideline or resource on using ChatGPT or other GenAI in teaching, learning or research §4.1. Online news articles, blog posts, and sources from a single department or programme were excluded §4.1. The inclusion flow is shown in fig. 1.

Step 3: Coding. Thematic analysis §4.2 was used. One researcher read part of the data and proposed child codes and parent codes. The other researchers reviewed and finalized the schemes. The schemes were then applied to all the data, and the team met regularly to check the coding and resolve disagreements §4.2. Two schemes were used:

• The policy scheme has four parent codes: University Decision, Instructor Decision, Education Purpose, and Research Purpose tab. 1.
• The resource scheme has three parent codes: Target Audience (three child codes), Types of Resources (five child codes) and Content Analysis (seven child codes) tab. 2.

Step 4: Scoring. Each university receives a stance score from the policy it was coded with, and a resource breadth score from the resources it offers §4.3. The stance points are fixed in tab. 3. Undecided or unclear policy scores 0, and the score is meant to be neutral and open. The resource score adds one point per target audience (at most 3), one per resource type (at most 5) and one per content category (at most 7), for a maximum of 15 §4.3.

# Stance score for one university  [§4.3] [tab. 3]
# Input:  coded policy decision (a key of STANCE_POINTS)
# Output: stance score in [-5, 5]

STANCE_POINTS = {
    "ban": -5,
    "undecided": 0,
    "allow_with_conditions": 5,
    "instructor_prohibition_default": 1,
    "instructor_neutral": 2.5,
    "instructor_permissibility_default": 4,
}

def stance_score(policy):
    # Look up the fixed points for the coded decision  [§4.3] [tab. 3]
    return STANCE_POINTS[policy]

# Resource breadth score for one university  [§4.3] [tab. 2]
# Input:  sets of codes found in its resources
# Output: integer in [0, 15]

def resource_score(audiences, types, contents):
    # One point per distinct target audience, at most 3  [§4.3]
    score = len(audiences)
    # One point per resource type, at most 5  [§4.3]
    score += len(types)
    # One point per content category, at most 7  [§4.3]
    score += len(contents)
    return score

Step 5: Tiers and focus groups. Ranks 1 to 33 form Tier 1 (n = 34), 34 to 66 form Tier 2 (n = 32), and 67 to 100 form Tier 3 (n = 38) §5.3.1. Separately, the authors split the sample into "technology-oriented" universities (n = 24) and "comprehensive" universities (n = 80) §5.3.2. The paper does not state how a university was assigned to each group beyond these examples.

Hyperparameters

This study has no learned parameters. The table lists the numeric choices that affect the results.

Results

Author-reported. Setup: 104 official U.S. university sources, collected to April 2024, coded by hand. No baseline and no statistical test. Counts are as given in the text; percentages are as reported.

The paper's pedagogical suggestions are in §6.1 and tab. 4. Policy suggestions are in §6.2. The conclusion summarises both §7.

Paper vs code

No official code.

The arXiv abstract and comments contain no code link. The paper links a Google Sheet with the dataset §4.1. That is data, not code, and it is not used here. There is no project page. A GitHub repository search for "genai university policies" returned several candidates, but none could be tied to the authors, so none is used.

Pitfalls

• Counts don't add up in one place. Instructor-decides universities total 57 §5.1, but the two defaults add to 56 (27 and 29) §5.1. One university is unaccounted for.
• Percentages differ from counts over 104. For example, 19 of 104 is 18.3%, not the 17.8% in the text, and 15 of 104 is 14.4%, not 14.7% §5.2.1. Use the counts.
• Tier ranges disagree. Tier 1 is "top 1-33" but is counted as n = 34 §5.3.1.
• The stance scale is the authors' own. The point values in tab. 3 are not derived or validated in the paper. Ranking by these scores is only as reliable as the coding.
• Coding reliability is not reported. The paper describes team review but gives no agreement statistic §4.2.
• Correlation is asserted, not tested. The paper says there is no significant correlation between rank and depth of resources §5.3.1, but reports no test. The tier comparison is descriptive.
• Snapshot. Documents are as found up to April 2024 §4.1. Policies may have changed; the authors say so §4.1.
• Scope. Only the top 100 U.S. universities are covered. Department-level policies were excluded §4.1, so the findings do not describe what individual courses do.
• Dataset access. The dataset is linked as a Google Sheet §4.1. Its current availability was not checked.

Lineage

• Builds on: thematic analysis as the coding method §4.2, cited to Braun and Clarke (2006) in the paper.
• Compared against: none. The paper is descriptive.
• Superseded by: not known from content in this wiki.