How US CS courses let students use generative AI

Coding. Each course policy was read at its primary source and coded on an ordinal 0–4 permissiveness scale, plus three tri-state features: disclosure required, AI banned on exams, course-provided AI tool. Fields a policy doesn't address are left null and excluded from rates, so they don't count as "no".

Measures. Level distribution, mean level by course level / institution type / year, and feature rates among policies that state them. This is computed by analysis/analyze.py and checked by tests against hand-computed fixtures.

Limitations. The sample is small (17) and non-random: it favors well-known R1 schools with public course pages. Courses behind logins (e.g. Princeton) or with broken certificates (Cornell, UChicago) were excluded. A single coder (Claude) did the coding; the student should spot-check a sample against the sources. Treat the numbers as descriptive, not as national estimates.

Next step for a stronger analysis. Draw a stratified random sample (R1 / R2 / liberal arts / community college × CS1 / CS2 / upper) and have two coders apply the codebook independently. Report Cohen's κ, then test the year trend with an ordinal regression.