AI in Higher Education Newsletter
August 11, 2026 · Vol. 28
A weekly brief for the Management Department, Isenberg School of Management, UMass Amherst. By Matthew D. Langenkamp / 雷邁德, prepared in collaboration with Thea 🪻✨.
← Back to archive · ← Back to Dictionary
This Week’s Thesis: Adoption Is No Longer the Hard Part
Universities spent much of the past two years asking whether students and faculty would adopt generative AI. That question is becoming quaint. The harder question now is whether institutions can introduce AI without quietly redefining learning, professional judgment, privacy, and the relationship between students and their university.
This week makes the tension unusually visible. Colorado students are resisting university AI partnerships on intellectual, environmental, and privacy grounds. A University of Alberta computer scientist has built a local AI system that keeps course materials and intellectual property off the cloud. A useful course-design essay argues that AI literacy cannot be bolted onto the first year as a compliance module. And public universities in South Dakota are turning AI from a tool into a field of study, with new programs from bachelor’s through doctoral level.
The common thread is not whether AI belongs in higher education. It is what kind of institutional bargain accompanies it. Access without judgment is not literacy. Scale without consent is not trust. And a policy telling students what they may do is not the same thing as teaching them when an AI-assisted choice is educationally sound.
1. Colorado Students Are Not Waiting Quietly for the AI Future
The most consequential story this week comes from Colorado, where student resistance has collided directly with university-scale AI adoption.
The University of Colorado system announced a $2 million OpenAI partnership earlier this year. Colorado State University is also creating its own version of ChatGPT. CU presented the arrangement as a matter of equity, security, and workforce preparation: if students are already using AI, the institution should give everyone access and teach responsible use. By March, more than 28,000 people at CU Boulder had standard ChatGPT accounts linked to university email addresses, including more than 3,000 faculty and staff.
But nearly 800 CU researchers, students, and professors signed an open letter opposing the partnership. Students raised concerns about data privacy, environmental impact, uncompensated scraping of creative work, and what some described as cognitive decline from overuse. The backlash delayed the campus launch and led CU to require training, create working groups, and amend the contract to state explicitly that campus data could not be used to train OpenAI’s model.
Graduate student Mirakle Wright has begun studying students who identify as AI-avoidant. Her early finding is worth taking seriously: universities often assume that reluctant students merely lack experience and will eventually come around. Some will not. Their objection is not always fear of unfamiliar technology; it can be a considered view about what education is for. UMass very likely has a sizable cohort that is similar to what is being studied in Colorado.
One student put the disagreement plainly: “You don’t really learn something until you do it yourself.” That claim is not universally true, but it identifies the fault line. Faculty and administrators who learned their disciplines before generative AI may see it as an efficiency tool. Students asked to use it before they possess the underlying skill may experience it as removal of the very practice they came to university to acquire.
Why it matters for Isenberg: We should not confuse business students’ pragmatism with blanket consent. AI may be professionally relevant and still be educationally inappropriate at a particular stage of learning. A credible course policy should explain not only where AI is permitted, but what capability the student is supposed to develop before delegating part of the task.
2. Local AI Offers a Different Institutional Bargain
At the University of Alberta, professor emeritus Jonathan Schaeffer has built a local AI platform around a simple proposition: researchers and teachers should not have to choose between using AI and surrendering control of their intellectual property.
His company’s system, Kind, runs on a user’s own computer (and is similar to what the Management Department at Isenberg discussed in June). It indexes PDFs, presentations, recorded lectures, and other local files, then answers questions while pointing back to the relevant page or video timestamp. Schaeffer used it in an online course with his own lectures and materials. Students received answers grounded in the course they were actually taking rather than in whatever the public Internet happened to offer, and he reports receiving fewer clarification emails.
Local systems come with trade-offs. They may be slower and less capable than the largest cloud models. They also know only what the user supplies. In education, however, that limitation can be a feature: a deliberately bounded corpus can reduce hallucination, protect intellectual property, and keep the source of an answer visible.
Schaeffer’s phrase is the important one: “This is all information that I want a wall around.” Universities have treated local AI as a specialist’s hobby for too long. For research drafts, unpublished data, student work, recorded classes, and proprietary case material, the wall is not eccentricity. It may be the correct design requirement.
Why it matters for Isenberg: Before sending sensitive teaching or research material to a commercial cloud service, we should ask whether the task can be done with institution-controlled or local infrastructure (some of it can). The relevant comparison is not simply which model scores highest. It is capability, provenance, privacy, cost, and control together.
3. AI Literacy Cannot Be a Compliance Module
Nicole Brownlie of the University of Southern Queensland offers the week’s most useful teaching argument: AI literacy should be embedded in disciplinary learning, not attached to a first-year course as a list of permitted and prohibited uses.
Her distinction is between compliance and judgment. Policies can tell students what is allowed, what must be disclosed, and what may trigger an academic-integrity violation. They do not necessarily teach students when AI is useful, what it is likely to flatten or omit, what requires verification, and what responsibility remains with the human learner.
Brownlie’s examples are modest and therefore useful. A student might use AI to map the sections of a difficult article, clarify an unfamiliar term, or generate an example. The student must then return to the source to recover the author’s actual argument, check accuracy, and decide whether the example fits the concept. In another exercise, students compare an AI-generated explanation of a learning theory against course readings and identify what it oversimplifies.
The governing question changes from “Did the student write this?” to “How was the student thinking?” That does not solve every assessment problem, but it is a much better design question.
Why it matters for Isenberg: Management education is full of judgment-intensive tasks. A future manager can ask AI to draft a competitive analysis, valuation narrative, market-entry recommendation, or stakeholder map. The educational work lies in deciding what evidence matters, what assumptions are hidden, what the model missed, and whether the recommendation survives contact with reality. Those judgments need repeated practice inside courses, not one orientation video at the start of college.
4. Public Systems Are Turning AI into Degree Infrastructure
The South Dakota Board of Regents announced four new AI-focused programs at Dakota State University and South Dakota State University, ranging from bachelor’s to doctoral degrees. The programs build on a state plan adopted earlier this year to integrate AI into teaching, research, and workforce development.
This is a different level of institutional commitment from purchasing chatbot access. AI is becoming curriculum, credential, research infrastructure, and economic-development strategy at the same time. The Board of Regents’ language joins technical experience with an “ethical foundation,” which is exactly the right pairing, though the proof will be in course and assessment design.
There is also a strategic warning here for business schools. If AI remains housed exclusively in computer science, students may learn how systems are built without learning how they alter incentives, organizations, governance, markets, and work. If business schools teach only tool use, students may learn the interface without understanding the system.
Why it matters for Isenberg: The durable curricular opportunity is neither “AI for everyone” nor a narrow technical concentration. It is domain-specific AI judgment: how managers evaluate models, redesign processes, allocate accountability, protect data, and recognize when automation has made a decision less legible rather than more intelligent.
5. The Product Layer Is Moving from Chatting to Doing
OpenAI’s August 4 education announcement introduced new ways to use ChatGPT Work and Codex for teaching, learning, research, and building. The strategic signal matters more than any individual feature: the educational AI interface is moving from a student asking a chatbot for text toward systems that can carry out longer, multi-step work across files, connected applications, and the web.
That shift will make familiar AI policies age badly. A syllabus rule written around “generating text” does not tell a student whether an agent may gather sources, restructure a dataset, test code, prepare slides, or operate across an approved collection of course files. Nor does a disclosure statement by itself tell a faculty member which parts of that workflow still demonstrate student learning.
This is where assignment-level AI zones become useful. Faculty can identify parts of a task that must remain human, parts where AI assistance is permitted with disclosure, and parts where using an agent is itself the professional skill being assessed.
Why it matters for Isenberg: Agentic tools make process evidence more important. If the final artifact can be produced through a chain of delegated steps, students should sometimes be asked to preserve decision logs, intermediate judgments, rejected alternatives, source checks, and a short defense of the final recommendation.
6. The Governance Gap Is Becoming the Story
Several higher-education pieces this week converge on the same operational problem: model capability and campus use are moving faster than the committees responsible for policy, risk, curriculum, procurement, and academic integrity.
The answer is not necessarily a larger central committee. It is a clearer division of labor. Institution-wide governance should set the non-negotiables: privacy, procurement, accessibility, security, data retention, transparency, and appeals. Programs and departments should define the professional judgments graduates need. Faculty should retain responsibility for assignment design and the evidence of learning required in a particular course.
Without that separation, institutions produce either vague principles nobody can apply or overly detailed rules that expire before the next semester begins.
Why it matters for Isenberg: We need a small number of stable rules and a larger number of revisable teaching practices. Governance should establish the floor. Disciplines should decide what competence looks like. Faculty should decide how students demonstrate it.
One Practical Move for Fall
Before the semester begins, take one recurring assignment and add three short statements:
- The capability being learned: what students must be able to do themselves.
- The AI boundary: where AI is prohibited, permitted, or expected.
- The evidence required: what students must retain or explain so that their judgment remains visible.
That is not a complete AI policy. It is better: it is a usable piece of course design.
The Capstone
The universities moving fastest on AI are beginning to discover that access is the easy part. A contract can be signed, accounts provisioned, and a training module assigned. None of that settles the educational question.
The deeper task is to decide what students should still learn to do when a machine can do part of it for them. That decision cannot be outsourced to the vendor, the central administration, or even the policy committee. It belongs inside the discipline.
The useful position is neither resistance nor boosterism. It is conditional adoption: use AI where it expands inquiry, protects privacy where the material warrants it, preserve the productive struggle by which novices become competent, and require students to show the judgment that the final answer alone can no longer prove.
That is less exciting than announcing an AI partnership. It is also the work we should embrace at this point in time.
Sources cited: Elizabeth Hernandez, “Some Colorado Students at Odds With Universities Shifting Toward AI,” The Denver Post / GovTech, Aug. 10, 2026; Abby Sourwine, “University of Alberta Researcher’s Local AI Keeps Data Off the Cloud,” GovTech / Center for Digital Education, Aug. 7, 2026; Nicole Brownlie, “AI Literacy Cannot Be Bolted On to First-Year Courses,” Times Higher Education, Aug. 10, 2026; C.J. Keene, “State Higher Education System Further Embracing AI With New Programs,” South Dakota Public Broadcasting, Aug. 10, 2026; OpenAI, “New Ways to Learn and Teach With ChatGPT Work and Codex,” Aug. 4, 2026; Times Higher Education, “With AI Models and Use Outpacing Governance, Committees Need Help,” Aug. 6, 2026.
Questions or topics for next week? Reply to mlangenkamp@umass.edu. Prepared in collaboration with Thea.