How US CS departments talk about generative AI
RQ1 · What are students permitted to do?
Course-level stance (each syllabus / SE course page)
Ordinal codebook: 1 Prohibited → 5 Encouraged. "None" = page mentions AI but states no rule.
Disclosure is the common condition
Share of documents that ask students to cite, disclose or document AI use.
RQ2 · Tone of official publications
Mean sentiment by document type (−1 cautious … +1 positive)
Positive vs. cautious documents
Positive: score > 0.2; cautious: score < −0.2. Department pages and news promote AI; course syllabi focus on risks and integrity.
RQ3 · Is AI framed as a collaborator?
Framing found in AI sentences (documents)
Collaborator = "as a collaborator/peer/partner", "collaborate with AI", pair programming. Tutor = tutor, rubber duck, study aid. Assistant = tool framing (not counted as collaboration).
AI tools named in evidence sentences
RQ4 · Software-engineering courses
Stance in SE course documents
RQ5 · Differences between institution types
Mean course stance (1–5)
% of schools with disclosure / collaboration language
Schools
Evidence explorer
Every code links to the source page and the exact sentences that triggered it.
Method
Sample. 30 US institutions, stratified: 22 R1 (10 private, 12 public) and 8 non-R1, across four regions. For each school, one domain-restricted web search found public teaching-center guidance, CS department publications, CS course syllabi and software-engineering courses. These pages were downloaded (HTML/PDF → text).
Coding. A transparent, rule-based codebook (pipeline/code.py, specs/02-coding-analysis.md) codes only sentences that mention AI. It assigns stance (1–5), sentiment (−1…+1), disclosure, collaboration framing, and whether AI is the subject of SE instruction. School-level stance is the median of its course documents.
Validation. All tests are on the validation page.
Limitations
- Purposive sample: search ranking decided which pages were found, so this is not a census. Each school's few documents may not represent all its courses.
- Some pages could not be fetched: sites that block automated clients (HTTP 403) and pages rendered only by JavaScript. Failures are listed in
data/documents.json. - Keyword rules miss irony and nuance, and can pick up unrelated text (e.g., news sidebars). Teaching-center pages list every template option, so they code as "mixed".
- Sentiment is lexicon-based: integrity language ("violation") reads as negative even when a policy is permissive.
- Pages are a snapshot from the retrieval date shown in each record; policies change each term.