Where does your institution stand?
Nearly every U.S. research university has an AI hub, a chatbot, and a policy page. Far fewer have a published record of anyone deliberating. This project catalogues the deliberation: reports with findings and recommendations, issued by named task forces, committees and working groups, at all 326 institutions classified Research 1 or Research 2 in the 2025 Carnegie update, and at 46 selective liberal arts colleges chosen by endowment and admit rate.
Two fields here appear in no existing collection: the document’s own date, and who convened the body that wrote it. The second is the variable that matters. A provost’s task force and a faculty senate committee produce the same genre with different politics.
The absence is the other finding. Forty-six institutions reached the crawler with twenty-five or more links that look like task force or committee documents, and no AI report among them: twenty-four at R1, twenty-two at R2. These are places that publish their committee work as a matter of course. Whatever they decided about AI, they did not decide it where they decide everything else. The list is on the Patterns tab, and the per-institution evidence is in the census.
The Inventory
One row per document. Institutions with several bodies have several rows. The headline set is deliberative reports at R1 and R2 campuses (the Carnegie class is shown under the institution); twenty-six supplementary rows are kept for completeness and excluded from the counts. They are documents that carry a named body or a public link but fail one part of the inclusion test: charge letters for bodies whose report is not yet posted, principles statements, survey reports, implementation plans, guidelines an office wrote after a body reported, a college-level committee, a systemwide senate report, an adviser report, a slide deck, and reports that have since moved behind a login. Each carries a note saying which.
Genre distinguishes reports from committee-authored guidelines that carry findings and recommendations, and from strategy plans. Access distinguishes full public text from summaries and one-page abstracts. Every headline document has been read and coded, 96 end to end and three (UNC Charlotte, UW-Milwaukee, Loyola Marymount) from the portion the public source exposes; ten of the 99 are the summary, abridged version or abstract published in place of the report.
The Codes column carries five of the ten coded dimensions: the default rule for student use, the position on detection tools, whether the report recommends permanent machinery, where students sat, and the specificity score, which counts how many of three things a report does (names an owner, sets a deadline, asks for money). Hover the posture, detector or standing-body code to see the sentence behind it. The recommendation count sits beside the score. The full coded table, the composition table and the coding script are under Data & Downloads.
One caution about the standing-body code. A value of none means no proposal appears in the report, which is not the same as a decision against one. standing_body_disposition carries the difference for all 22, from a reading of each full text rather than from a keyword rule.
| Institution | Convened by | Report | Date | Codes | Genre | Access | Link |
|---|
The Census: all 326 Carnegie R1 and R2 institutions, and 46 selective liberal arts colleges
Every institution designated Research 1 (187) or Research 2 (139) in the 2025 Carnegie Research Activity Designations file was crawled on 5 September 2026 by a search-engine-free crawler, version 0.4.2, and again on 8 September by version 0.5.0, that probes about sixty governance subdomains by name, seeds about sixty governance paths, follows links three deep with AI-tagged links and committee index pages first, and flags documents that look like AI deliberation. Every candidate carrying all three signals (an AI term, a body term, an output term) was opened and read, apart from a small number that no longer resolved.
Institutions and UnitIDs come from the Carnegie file itself, not from name-matching; that correction moved UA Huntsville from R1 to R2 and added Weill Cornell to R1, both of which v1.0 had wrong. Thirteen R2 institutions and one R1 (Howard) are HBCUs; Spelman, in the liberal arts segment, is the fifteenth.
The liberal arts segment (LAC) is a selection, not a Carnegie class. From the 216 institutions in the 2021 Baccalaureate: Arts & Sciences category, those with at least 300 undergraduates, endowment of at least $150,000 per undergraduate (College Scorecard) and an admit rate of 40% or lower: 46 colleges, from Amherst to Williams, including Barnard, Spelman and all five Claremont colleges. The cut is meant to isolate institutions that can afford instructor discretion; it is stated here so it can be argued with.
Three of the 46 refused the crawler (Colby, Holy Cross, Williams); I checked those three by hand, and their census notes say so. One of the 46 has a findable deliberative report, and I found it by hand after the site refused the crawler, so the protocol itself found none. Amherst’s 2023 task force report exists behind community login. Whether the rest deliberate in some form this inclusion rule does not reach, in faculty meetings, honor councils or teaching-center guidance, is not something this method can say.
Recall, measured. Run blind against the 55 R1 institutions whose reports were known before the re-run, v0.5.0 re-finds reports at 44 (80%) and 53 of 69 documents (77%). The v0.4.2 crawler used from v1.1 to v1.4.4 scored 42 (76%) and 50 of 69 (72%) on the same test, and the v0.2 crawler used for v1.0 scores 26 of 55 (47%); the “roughly 55–60%” stated in v1.0 was an estimate from agent notes and was too generous. The benchmark script is published, and the held-out set is frozen in bench_holdout.json, so the figure can be reproduced after later releases rewrite other columns.
Of the eleven institutions still missed, two also obstruct the crawler: Kentucky rate-limits it and Penn refuses most of its requests. All eleven host the report somewhere no link-following crawler reaches: a PDF orphaned from its own AI hub (UTEP), an academic-affairs page whose navigation is rendered by script (Buffalo), a chancellor’s site that no longer links its 2024 report (UNC Chapel Hill), an ombud office (Kentucky), a humanities lab (Penn), a teaching center (Boston College). Search engines found those because they indexed them when they were linked.
No R2 recall figure exists yet. One R2 report was known before the crawl (RIT, harvested on 1 September and held out of the headline set until v1.4.2 by a rule that no longer applied), which is one known positive, too few to estimate recall from. The R2 protocol was tuned after a first pass exhausted its queue at 52 of 139 sites, whose governance lives under paths on the main host rather than on subdomains, and the second pass read a median of 145 pages per site.
Read “no public AI report surfaced” as: no AI task force document was linked from the governance surfaces reached, at an instrument with roughly three-in-four recall on R1 sites and an unmeasured, probably lower, recall on R2 and liberal arts sites. Treat those rows as “likely no public report; worth asking,” not as a count.
About the numeric columns. Requests counts fetches attempted, including failures. Pages read counts HTML pages actually retrieved and parsed; this is the column that tells you how much of the site was seen. AI candidates counts links that scored as AI deliberation and were opened. Other doc links counts links that scored like task-force or committee documents without an AI term; those were not opened. Status is assigned mechanically: blocked when fewer than 30 pages were read and at least 40% of requests were refused; too shallow when fewer than 60 pages were read for any other reason.
| Institution | Class | Control | Status | Requests | Pages read | AI candidates | Other doc links | Since | Map quadrant | Note |
|---|
Who Sat On These Bodies
Seventy-one of the 99 reports print a membership roster, 1,484 seats in total. Six more print only their chairs; those are counted as leadership listings, not rosters. A person who sits on two of these bodies is counted twice, so the figure is seats rather than people. This page reports what those rosters are made of. It does not list the names.
The names are public in each report and are not republished here. A single index of them would serve no research purpose that composition does not, and would serve other purposes. Each report is linked from the Inventory tab and prints its own membership.
Roster seats by category, across every printed roster
Faculty 429, administrators 292, staff 207, students 27, external members 8, and 521 whose category the report does not make printable. Faculty outnumber administrators on the rosters, but not in the chairs: of the 178 chairs and co-chairs named across the 77 reports that print any, administrators hold 60, faculty 55, staff 23, and 40 carry no printed title. Six further bodies say a student served without naming which member did.
The categories come from the printed title: a dean is an administrator, a center director is staff, a graduate student is a student. Where a report prints names without titles the category is unknown, which is why that category is large. Twenty-two reports print no roster at all, and six print only their chairs.
The per-report composition counts, without names, are published as roster-composition_v14.csv. Researchers who need the named rosters for scholarly work can reach me through kylesaunders.com; the request is the point, because it puts a person rather than a crawler between the names and whoever wants them in bulk.
Patterns
Descriptive only. The denominator for discoverability is institutions where the crawl reached governance pages and either found a report or did not: 180 of 187 R1s, 127 of 139 R2s, 44 of 46 liberal arts colleges. Blocked and thin crawls are excluded. “Found” measures public discoverability under this protocol, not whether deliberation happened, and retrieval failures are not evenly distributed across institution types, so none of this is corrected for differential missingness. R2 recall is unmeasured. Differential retrieval may affect the observed gap, and neither the size nor the direction of that effect is established here.
Who has a findable report
Every bar in this group runs on the same scale, nought to one hundred per cent of the institutions in that row’s denominator.
By segment
The bifurcation in one chart. The third bar reads differently: one selective liberal arts college in 46 has a report findable under this protocol, and it was found by hand rather than by the crawler. Whether these colleges deliberate about AI in some other form, or in public at all, is not something this method can say. Of the fifteen HBCUs in the census, none has a findable report; four crawls were too thin to say, and the rest read normally and surfaced nothing.
By control
By when the institution entered its class
Institutions promoted in 2025 are less likely to have a findable report in both classes. Two readings fit: they have deliberated less in public, or their governance sites are thinner and the crawler reads them worse. The census columns let you judge which.
By quadrant on the Structural Divide map
R1s are overwhelmingly High Capacity, so the other three quadrants hold few institutions each. Read the differences as suggestive.
What the reports say, once you read them
The ten coded dimensions are defined in the codebook and computed by a published script from a structured reading of each document. Every bar in this group counts reports, on the same scale, nought to one hundred.
Campus or system
The printed title does not settle this. A chancellor runs the campus at UC Davis and the system in the Colorado State and California State systems, so the office was read from each document rather than from its name. Seventy-seven reports were chartered at the campus, ten above it, and twelve name no charging office at all. Of the ten, four name a system office outright and the rest were inferred from a remit that crosses campuses. Whether a report binds one campus or several is coded separately, in scope_level.
Who convened the body
What the reports cover
Most reports carry more than one tag, so these do not sum to 100. Teaching is near-universal and curriculum is not, which is the gap between advising instructors and changing what is taught. Fifty-one of the 99 reach administrative and operational use, so the genre is not only about classrooms.
The default rule for student use
Ten of the 99 reports state an institution-wide default at all, and exactly one of those defaults to permission. Whatever these bodies are doing, they are not banning AI; they are handing the decision to the instructor, or not making it. Of the 47 that state none, seven are summaries, abstracts or an abridged text, so their silence is silence in what the public source exposes, and a research or procurement charge need not settle student use at all.
Position on AI detection tools
Forty-two reports never raise detection, and nine raise it without taking a position. Of the forty-eight that take one, thirty-five are against it or cautionary and exactly one (Maryland) recommends piloting a detector.
Where students sat
Twenty-four bodies seated a student as a member, though six of those say so without naming which member is the student. Twenty-seven more print a roster, or say so in prose, with no student on it.
What the report says to buy
Seriousness: owners named, deadlines given, money asked for
Year of publication
The same question, asked by different offices
Shares again, on the same nought to one hundred per cent scale, among the reports in each convener group.
Share recommending a new or expanded standing body, by convener
This is the governance finding. Administrations institutionalize: provost-chartered bodies recommend a new, expanded or newly staffed standing body in 32 of 38 cases, jointly chartered ones in 14 of 16, senate-convened ones in 5 of 12. Counting only brand-new bodies, 22 of 38 and 13 of 16. Proposing none is almost never a refusal: of the 22 reports coded that way and read again in full, one decides against a continuing body in writing, nineteen route the work to structures that already exist, and two never raise the question. Not one senate report refuses. Each verdict and the passage behind it is in standing_body_adjudication.json.
Share taking a position on detectors, by convener
Share recommending central enterprise licensing, among reports with a procurement position, by year
The drift is away from central licensing and toward staying tool-agnostic. This is a claim about the procurement model, not about how many vendors a campus ends up with: the category counts reports asking the institution to license enterprise access, several of which name two or three products. Yearly counts are small, the 2023 bar rests on five reports, and 1 report with no publication date is not plotted.
Who cites whom
Ninety-two of the 99 reports cite an outside institution, organization or government document; 1,330 citations across 871 distinct sources. Entries that name several organizations at once are split before normalizing, the names are normalized by a published keyword map, and the printed string is kept in the data. This chart counts reports citing each source, so its scale is not the one above.
The fifteen most cited external sources
The two most cited sources are vendors, OpenAI and Microsoft, ahead of every university. Harvard (19) and Michigan (18) are the most cited individual universities, and one citation separates them, so read the order as a tie rather than a ranking. Harvard has no report in this inventory at all: what circulates is its July 2023 guidance, which disclaims being policy. The University of California sits above both only because its campuses and system office pool into one node.
The archive-without-AI pattern
The rule, applied to R1 and R2: an institution where the crawl found no AI report but reached twenty-five or more links that score like task force or committee documents.
Twenty-four R1 institutions meet it: New Mexico State (483), Southern Illinois (357), UC Irvine (357), Texas Tech (322), Binghamton (175), Connecticut (109), Montana (95), Oregon (88), UT Southwestern (77), Louisiana at Lafayette (73), Montana State (67), Baylor (62), Colorado State (59), Auburn (58), Saint Louis (49), West Virginia (42), NJIT (37), North Dakota State (37), Iowa State (36), Missouri S&T (36), Virginia Commonwealth (32), Ohio University (31), CUNY Graduate Center (30), Maine (28). So do twenty-two R2 institutions: CUNY Hunter (714), Portland State (469), Rowan (306), TCU (209), Southern Connecticut (147), UA Huntsville (141), San Jose State (134), Akron (127), CSU Long Beach (116), Kennesaw State (79), Middle Tennessee State (74), Stevens (55), Western Michigan (45), Tennessee Technological (44), Marshall (42), CSU Los Angeles (41), West Chester (37), Missouri-St. Louis (34), South Dakota (32), Eastern Michigan (27), Idaho State (27), North Carolina A&T (25).
New Mexico State is the limiting case: its faculty senate passed a resolution in December 2025 asking for an AI task force report by April 2026, and the report is not posted. Colorado State keeps its Provost’s AI Working Group drafts on SharePoint behind a login.
Candidates were identified by link text and not opened, so read this as a strong signal that these institutions publish deliberation routinely and a weaker signal about AI specifically; half of them do expose AI-tagged links, which on inspection were not deliberative documents. Whatever these institutions decided about AI, they did not decide it in the place they decide everything else.
Methods
Inclusion rule
A report is a deliberative document with findings and recommendations from a named institutional body: task force, committee, working group, commission or advisory group. Excluded from the headline count: policy pages, syllabus guidance, library guides, acceptable-use rules, meeting agendas, slide decks, news coverage, and single-author adviser reports. Two boundary genres are kept in the headline set and labelled: committee-authored guidelines that carry findings and recommendations (Kentucky, LSU and Rochester at the outset, eleven rows by v1.3), and a strategy plan built on working-group output (Texas A&M). Content decides genre, not the word in the title.
Access is recorded separately from genre: full text, abridged, public summary of a gated report, or one-page abstract. Abstract-only rows establish that a report exists, not what it says, and their recommendation field says so.
How the reports were found
Three passes. First, a search-engine harvest on 1 September 2026 across six segments (R1 public, R1 private, regional and community college, systems and associations, international, prior art). Second, a search-engine-free crawl of all 187 R1s, first on 4 September with v0.2 and again on 5 September with v0.4.2, and of all 139 R2s on 5 September with v0.4.2. Third, a re-crawl of the whole census, all 372 institutions, on 8 September with v0.5.0, after a reader-reported miss exposed a scoring bug.
The crawler (crawl2.py) starts from governance. It probes about sixty subdomains by name (provost, senate, council, president, academic affairs, ai, teaching, research, ombud), discovers more from the pages it reads, and seeds about sixty governance paths. It then follows links three deep against a priority queue, taking AI-tagged links first, then committee and report index pages, then everything else on a governance host, and it caps any single subdomain or section so that an AI department cannot absorb the budget.
Several fixes address how institutions actually publish. The crawler keeps documents hosted off-domain on Drive, Box, SharePoint or a repository when an in-scope page links them; resolves relative links against the URL the server actually served; falls back to a browser user-agent on refusal and to curl on a TLS failure; extends its request budget when nothing AI-tagged has surfaced; and follows homepage links whose path or anchor names a governance office, because at smaller institutions governance sits under a path on the main host rather than on a subdomain of its own. Every change was made after diagnosing a specific miss against a known report, and the version notes in the file say which. Every candidate scoring at or above threshold was opened and read before inclusion, apart from a small number that no longer resolved.
External audit, 5 September 2026
An independent audit of the live v1.0 site checked all rows and URLs. The errors it found are corrected in v1.0.1: a wrong date (UNC-Chapel Hill, August not September 2024); a recommendation attributed to a one-page abstract that does not contain it (Chicago); two rows that violated the stated inclusion rule (a single-author adviser report at Dartmouth and a slide deck at Ohio State), now moved to supplementary; missing reports at Michigan (2023 and 2026), UNMC (2023) and UNLV (2024), all hosted off the institution’s domain; a systemwide report attributed to one campus (Indiana); an ambiguous campus attribution (CU Denver / Anschutz); and crawler metrics labelled as more than they measured.
The audit also identified crawler design flaws, fixed in v0.3. The audit document is linked under Data & Downloads.
What the crawl can and cannot tell you
Recall was measured, not estimated. Run against the 55 R1 institutions whose reports were already known, v0.5.0 re-finds reports at 44 of them (80%, or 53 of 69 documents); the v0.4.2 crawler used through v1.4.4 scored 42 of 55 (76%), and the v0.2 crawler used for v1.0 scores 26 of 55 (47%) on the same test. The re-run surfaced 17 R1 headline reports that search had missed, at nine institutions previously recorded as having none. The Census tab gives the misses one by one and explains why no R2 figure exists; the benchmark script and its frozen held-out set are under Data & Downloads.
Fields
Institution, state, control, IPEDS UnitID, Carnegie 2025 class (R1 or R2), body name, convening authority, report title, publication date (from the document, unknown if absent), date accessed, URL, scope, document status, and one summarized key recommendation. Map quadrant and resilience score are joined from Mapping the Structural Divide by UnitID.
How the coding was done, and a limit stated plainly
All 99 headline documents were read in two separated stages, 96 end to end and three (UNC Charlotte, UW-Milwaukee, Loyola Marymount) from the portion the source exposes, which the read_coverage field records. Reading a document in full is not the same as reading a report in full: ten of these documents are the summary, abridged version or abstract the institution published in place of the report, which the access field records.
First, extraction: Claude agents recorded facts and verbatim evidence into a fixed schema, including every member of every printed roster, who the body reports to, and every recommendation with its owner, deadline and resource ask. Second, coding: ten dimensions chosen from what actually varied in the extraction, each computed from extraction fields by a published script, so that re-running the script over the published extraction reproduces the coded tables exactly. Where a document was ambiguous the extractor recorded both readings and the published rule decides, rather than the coder’s taste. Four candidate dimensions were dropped as near-constants: faculty training, student AI literacy, assessment redesign, and the problems a report names.
The scheme is in the codebook; the schema, script, extraction and coded tables are under Data & Downloads. Member names, titles and units are withheld from the published extraction and from the site; see Who Sat On These Bodies for the reason and for what is published in their place.
There is no independent second coder. This is a one-person project. Both the extraction and the coding were done by Claude under published instructions; I checked 14 inventory rows against their documents in September 2026 and concurred with all 14, and I reviewed the coded values against the documents they came from and concurred with those as well. No inter-rater statistic is reported, because none can honestly be computed from one coder plus a check. Readers who need coded variables at a reliability standard should re-code from the documents: the extraction and the script are published, so a second coder can start where this one did.
What already exists, and where it stops
- Eaton, Institutional AI Policies & Governance Structures: 19 rows, last updated November 2024, no date field.
- Eaton, Syllabi AI Policy Repository: 217 course-level entries, maintained, no URL column.
- eduaipolicy.org: 3,195 sources, 828 institutions; no source type for “report.”
- Teleki et al., FAccT ’26: 79 institutions; the task-force column holds 70 links, two of them reports.
- Illingworth, HEPI Policy Note 71: of 163 UK institutions, 96 had a findable AI policy. The policy-level version of the discoverability result here.
Data & Downloads
The data
How it was made
Everything underneath
The raw materials, for anyone reproducing or auditing the work.
Seed lists, crawl outputs, candidate triage, audit (14 files)
Data licensed CC BY 4.0. Cite as: Saunders, Kyle. 2026. University AI Task Force Reports: An Inventory and Census of R1, R2 and Selective Liberal Arts Institutions, v1.5.1. kylesaunders.com.
Add or correct a report
If your institution’s report is missing or a link has moved, send the URL and the name of the body that produced it. Reports that exist only behind institutional login are recorded as such. That category is a finding, not a gap.
About
This is a companion to Mapping the Structural Divide, which places 1,556 U.S. institutions on resilience and market position. That map asks where institutions stand. This inventory asks what they have said, in public, about the pressure that map tracks as market misalignment. The two are joined by IPEDS UnitID.
Why reports rather than policies: a policy is the output. The report is where the disagreement is. Whether to use AI detectors, whether every syllabus must carry a statement, whether to license one enterprise tool or none, whether to create a standing body: institutions split on all of these, and the split is only visible in the deliberative document.
Corrections and additions are welcome, and a row you can argue with is more useful than one nobody checks. Email Kyle.Saunders@colostate.edu with a public URL for the document. A report that is not posted publicly cannot go in the inventory, since a reader has to be able to open it, but a note that one exists can go in the census, and that is itself a finding. A submitted report will be recorded as submitted rather than folded into the crawler’s results, so the recall figures keep meaning what they say.
Kyle Saunders is a professor of political science at Colorado State University and writes Sacred Cow BBQ.
This project began in conversations with Michelle Dion, Amarda Shehu and Karen Estlund, and was sharpened by earlier work from Sam Illingworth and Ethan Mollick. None of them is responsible for what I did with it.