Five numbers,
one week.
If you read nothing else this week, read these. Each card answers one question a colleague will ask you on Monday.
Cornell & UC Berkeley in Science: 9% of 95,000+ undergraduates admit using AI to cheat — 26% among daily users. Authors call assessment reform “necessary and urgent” and push for discipline-specific redesign, not detection.
UC Berkeley CSHE working paper finds the share of A grades rose ~13 points (≈30%) in AI-exposed homework-heavy courses since ChatGPT's release. No comparable jump in in-person exams.
Cal State quietly renewed its OpenAI contract for three years at roughly $13M/year — keeping 470,000 students and 63,000 staff across 22 campuses on ChatGPT Edu despite an NPR investigation and a faculty petition to end the deal.
Anthropic shipped Claude Opus 4.8 with a 3× cheaper fast mode, a 1M-token context window default, and a Claude Code “dynamic workflows” preview that runs hundreds of parallel subagents.
Gallup & Walton Family Foundation: only 18% of U.S. K-12 teachers have received any formal guidance on AI use. The gap is widest in higher-needs schools — meaning your incoming first-years arrive without shared rules.
Assessment, redesigned
under pressure.
The two empirical papers below now form the strongest evidence base anywhere that undergraduate take-home writing and coding have lost their signal. Everything else this week is a reaction to that fact.
Science: a 95,000-student survey calls for urgent assessment reform
Cornell information scientist René Kizilcec and UC Berkeley senior researcher Igor Chirikov, with Ivan Smirnov of UTS Sydney, surveyed more than 95,000 undergraduates at 20 U.S. public research universities through Berkeley's SERU Consortium. They report that ~37% use generative AI at least monthly and 9% admit using it to cheat — a rate that rises to 26% among daily users and 62% usage among computer science students.
The authors push for discipline-specific assessment reform led by professional societies. They also warn that AI-integrated assessment could widen equity gaps if access and literacy remain uneven across student populations — a finding they say should temper any rush to plug AI directly into grading.
Cornell Chronicle, May 21 → UC summary, May 29 →Bring this paper to your next assessment committee. It is the most-cited evidence base anywhere for moving away from detection-first integrity enforcement and toward discipline-specific redesign — and it gives you the language to argue for either approach with senior administration.
Berkeley CSHE: AI is driving grade inflation in writing and coding
Chirikov's parallel working paper analyzed 500,000 grades in 319 courses across 84 departments at a large Texas research university (2018–2025). A-grade share rose ~13 points (~30%) in AI-exposed courses post-ChatGPT, with no comparable jump in in-person exams — and the rise comes from converting A− and B+ to A, not lifting struggling students. Employers have already responded: minimum-GPA-3.5 listings on Handshake jumped from 9% (2020) to ~25% (2026).
CSHE Paper 26-3 → Chronicle, May 14 →Cal State renews three-year OpenAI deal — without faculty governance
CSU's original $17M, 18-month ChatGPT Edu contract was quietly renewed for three more years at ~$13M/year. NPR's investigation found majorities of students and faculty skeptical of the educational value. SF State's Martha Kenney is circulating a petition; CFA and CSSA say neither was consulted before renewal.
NPR investigation → Sacramento Bee →Anthropic ships Claude Opus 4.8 — 3× cheaper fast mode, dynamic workflows
Default pricing unchanged at $5/M input and $25/M output tokens. Claude Code's research-preview “dynamic workflows” spawns hundreds of parallel subagents for codebase-scale tasks. Alignment team reports Opus 4.8 is ~4× less likely than 4.7 to let flawed code pass unremarked — and lands “similar to our best-aligned model, Claude Mythos Preview.” A more capable Mythos-class model is expected “in the coming weeks.”
Anthropic release → VentureBeat →Gallup & Walton: 60% of teachers use AI — and 18% have any formal guidance for doing so
A new Gallup/Walton Family Foundation survey of 2,069 U.S. K-12 teachers (Feb 9–Mar 2, 2026) finds 60% use AI at work but only 18% receive any formal guidance. 69% have no guidance for AI tutoring; 58% have none for AI grading or feedback. The gap is widest in higher-needs schools — a structural problem your incoming first-years will arrive carrying with them next August. Coverage in Washington Post, Axios, and U.S. News ties the guidance gap directly to burnout and retention.
Gallup release →If the Berkeley CSHE pattern shows up in our transcripts, what do we tell accreditors and employers — and which assessment do we rebuild first?
Cal State just locked in three more years of OpenAI without faculty governance buy-in. Vendor stack, SUNY-style policy framework, or hold one more year?
What scholars learned.
What faculty tried.
Two columns: a paper and a framework that should shape how you think; two real classrooms that should shape what you do.
From the Research
UF / TechTrends: AI-as-starting-draft produces better essays
Brian Harfe (University of Florida, Associate Provost) required students in his bioethics course to begin a final essay with a fully AI-generated draft and revise it. Across 310 essays, students who engaged with AI as a draft did more substantive revision and structural rethinking than students writing without AI. The grading signal moved to process artifacts.
What you could do: require an AI starting draft, prompt log, and a “what I changed and why” memo on one assignment. Grade the memo, not the essay.
Summary, May 29 →Crompton et al.: a 22-country Delphi framework for governing GenAI
A global Delphi study published in the International Journal of Educational Technology in Higher Education spans 22 countries and six continents. Its core argument: effective GenAI governance can't sit only at the classroom level — it has to span the classroom, the institution, and the wider higher-ed system.
What you could do: if your campus is drafting AI policy this summer, use the framework as a scope checklist. Does your draft cover all three layers, or only your classroom?
Lance Eaton's writeup, May 29 →How Others Are Doing It
Indiana University: free GenAI 101 now open globally
President Pamela Whitten, profiled in U.S. News, opened IU's eight-module, 16-lesson AI literacy course to anyone worldwide. 114,000+ enrollments since launch; 805,000+ alumni added in October. Includes a conversational learning agent (Crimson) and a LinkedIn completion badge.
What you could borrow: assign one module — fact-checking AI output — as next week's pre-class prep. Externally branded, free, and ships with a credential students actually want.
U.S. News interview →Colgate & SUNY Oswego: a philosopher and a business professor pick opposites
Maura Tumulty (Colgate, philosophy) limits early undergraduate AI use, citing student anxiety about “being the one idiot not using AI.” Mohammad Tajvarpour (SUNY Oswego, business) uses AI for visuals and image analysis that previously required hired RAs. Both keep explicit syllabus policies.
What you could borrow: write a 100-word AI paragraph for your fall syllabus that explains why your policy is what it is — not just what's banned. Both faculty credit clear reasoning, not strict rules, with their classroom buy-in.
Spectrum News, May 27 →Everything else worth
knowing this week.
Eight stories that didn't earn full commentary — and four dates that should be on your calendar.
-
1Stanford's Tutor CoPilot supports human tutors in RCT Largest mastery gains for students paired with less-experienced tutors · May 28
-
2Long Island University launches AI² Center Interdisciplinary hub spanning healthcare, business, design · GovTech, May 27
-
3The Atlantic: “America has a Pangram problem” Detectors keep improving, “still aren't good enough” for high-stakes calls · May 30
-
4$45K scholarship pulled over a 98% AI-detector flag Student says board refused to view Google Docs version history · Reddit / IBTimes, May 29
-
5Times Higher Ed: institutional trust is now the four-party model Warwick Business School authors call it “Pedagogic Paradigm 5.0” · May 28
-
6Ethan Mollick: “Choosing to Stay Human” Using AI to think with, not for; two paired Wharton studies · May 26
-
7UVic bans AI detection tools as sole evidence in integrity cases Canada — Martlet, May 28
-
8Phil Hill: two blind spots in SUNY's system-wide AI policy Labor-market shifts and course-level enforcement gaps · On EdTech, May 26