State of AI in Higher Education 2026

Governance won in the US. It hasn't won everywhere.

Two years after ChatGPT's release, the American research sector has effectively settled the governance question: 178 of 187 Carnegie R1 universities (95.2%) publish AI guidance on their own domains. The rest of the world has not caught up — 77% across our 186-university global audit, 46% among Global South flagships, 19% in Central Asia. And while institutions govern AI, they still will not price themselves transparently: 79% of online tuition listings required a computed estimate because no total price could be found on the official site. The tools, meanwhile, are not ready to be handed to students unsupervised: a small RAG tutor invented answers 30–70% of the time on questions its knowledge base could not answer.

95.2% — and one holdout

The R1 census is a complete count, not a sample: every one of the 187 institutions in the 2025 Carnegie R1 classification was checked against its official domain. 178 publish, 8 are unverifiable, and exactly 1 has none. The unverifiable eight are not a bloc — each failed differently: a policy gated behind a VPN, guidance hosted on a third-party domain, program-level rules with no university-wide page, a committee still drafting. The holdout is the University of Wyoming: three searches found no adopted policy, only a 2023 working-group report, library guides, and commission news. When 95% of your peers have published, absence stops being a neutral position and becomes the story.

Browse all 187 university pages → · Dataset + CSV →

92% at the top, 46% in the Global South

Our 186-university global audit found 144 publishers (77%), 24 without, 18 unverifiable. Split it and the gap opens: 46 of the top 50 (92%) publish, against 17 of 37 Global South flagships (46%). Central Asia sharpens the point further: of 31 universities, only 6 (19%) publish AI policies — yet 18 (58%) already teach AI programs. Teaching is running ahead of governance where governance capacity is thinnest. That ordering matters: students are being taught with tools their institutions have not yet written rules for.

Global audit → · Central Asia audit →

79% of prices had to be estimated

Across 740 online programs at 537 institutions, 587 listings (79%) required a computed estimate — the real total price was not findable on the official site, so the basis for each estimate is documented per row. Only 153 listings carried a price an applicant could actually read. An institution that publishes an AI policy but not a price is transparent about the easy thing and opaque about the thing applicants pay for. Price opacity is the sector's other governance gap, and it is larger than the policy gap.

Tuition dataset →

A small tutor invents answers 30–70% of the time

We ran a 0.5B RAG tutor (Qwen2.5-0.5B-Instruct, CPU-only) against 30 teaching questions — 10 answerable, 10 partial, 10 unanswerable — across 3 retrieval configurations, 90 generations, hand-graded. On unanswerable questions the model invented answers 70% of the time with top_k 3 and small chunks, 50% with larger chunks, and 30% with top_k 5. The cheapest fix was retrieving more context, not a bigger model — and retrieval itself was never the bottleneck; generation was. The implication is blunt: a small RAG stack cannot be handed directly to students. It needs retrieval depth and refusal behavior first.

Benchmark dataset →

Databases, not diplomas, decide who exists

A new institution does not become real by declaring itself. It climbs a ladder of databases — institutional self-description, then Wikidata, then ROR, then OpenAlex — and each rung makes it findable to the next system. Our first audit checked 35 institutions against that ladder; the second censused 16,822 universities worldwide. ACLAS itself sits on the first rung: its ROR record was requested in October 2026 and is under review. Recognition is infrastructure, and infrastructure is slow — which is exactly why it is worth documenting rung by rung instead of asserting.

Recognition dataset →

Three readings

For university leaders: in the US research sector, publishing AI guidance is now the norm and absence is the outlier. The remaining work is not writing a policy — it is closing the eight unverifiable cases and the program-level gaps behind them.

For students: check two things before enrolling anywhere: whether the institution has a published AI policy, and whether it states a total price. The data says the second check fails far more often than the first.

For builders: do not ship a small RAG tutor to students on retrieval defaults, and do not treat database recognition as a checkbox. Both are ladders: retrieval depth first, then refusal behavior; self-description first, then Wikidata, ROR, OpenAlex.

What these numbers can and cannot say

Policy audits counted only guidance published on the university's own domain — third-party hosts, intranet-gated pages, and program-level rules did not count as university-wide policy. That rule is why 8 R1 universities are “unverifiable” rather than “no”: the audit reports what could be confirmed publicly, not what might exist internally. The R1 snapshot is 2026-10-18; policies published after that date are not reflected. Tuition “estimates” are computed figures with a documented basis per row, not prices quoted by institutions. The RAG result covers one small model and one teaching knowledge base; it does not generalize to large models. The recognition census describes database presence, not institutional quality.

Every number above, downloadable

Study Intelligence is a research initiative by ACLAS College.