AI guidance analysis

What AI can be used for

What has to be acknowledged

Academic misconduct

Tools and data

Risks and responsibilities

What kind of help institutions offer

What goes together

How the statements above line up with each other across institutions.

Noteworthy correspondences

Pick a statement

The bars repeat the sections above. Same bars as above. The faded part of each bar is every institution that says it; the solid part is the institutions that also say the statement you picked.

What the guidance says

Specific points from each institution's own document, quoted word for word. Each card links to the source. The one-line summary restates only what its quote says. Filter by theme, strength of wording, who it is addressed to, or a single institution.

How to read this

• Where the guidance comes from. UK and Irish institutions come from a frame of universities and higher education providers built from the ROR registry and drafted for inclusion by hand (not yet checked against the OfS and HEA registers), then a seeded stratified sample. Australia, New Zealand, Canada and the United States come from two seeded random draws from ROR of institutions whose names contain “university” (a pilot, then a second draw of new institutions). They are not a curated or representative sample, and non-universities in the draws (a system office, a research centre, a library consortium) were set aside. Some institutions in the second draw could not be searched because the search tool’s per-turn limit ran out; they are not coded. For each institution a web search looked for one student-facing page on generative AI, preferring an institution-wide student guide or policy, then a library or teaching-centre guide, then an academic integrity policy that names AI, then a student news page. Institutions in England, Wales, Scotland and Ireland also have a second page of a different kind where one could be found, usually the misconduct procedure. A second page counts under its institution, so counts are institutions, not documents. The page found is often a library guide or a landing page that points elsewhere, so what an institution says here is only what that page says. Institutions whose pages sit behind bot walls, JavaScript-only pages, robots.txt blocks or missing links could not be archived and are not coded; they are listed for manual fetching.
• Every value rests on a verbatim quote from an archived snapshot, checked by a validator. One model did all the coding passes, with a pilot of two independent coders before each full run. Agreement figures from the pilots are therefore optimistic.
• Counts are minimums. A statement counts only where a quote shows it, so a missing value means no quote was found, not that the institution disagrees. Some rows rest on a link label or on general misconduct wording that does not name AI; the misconduct section shows which.
• Tools and data: the pilot of two coders agreed on 75% of rows (tools only 54%, because the stance words overlap), so treat the stance of a tool row as the least certain value. The stance dots group provided and recommended together, and not endorsed, discouraged and warned together. Data stances have no “discouraged”, so “avoid sharing” is recorded as prohibited or warned. Many pages are library guides or landing pages with no tool or data statements, so most institutions have no row.
• Risks and responsibilities: the pilot of two coders agreed on 98% of risk values but only 83% of duties and 65% of who bears a duty, so read bearer and strength (must, should, will) as the least certain values. A risk counts only where the sentence says what can go wrong; a duty only where the quote states it. Many documents say nothing here, and long policy documents give far more rows than short guides.
• Documents with tiered rules (for example traffic-light levels) give several rows for one use, so their counts are lower bounds.
• “What goes together” is descriptive. The noteworthy pairs are corrected for the number of pairs compared, but an association in what documents say is not a reason, and three-way combinations are too thin to test with this many institutions.