AI is much more expensive than the price we see when we pay per token. Its costs are reflected in carbon footprint and water usage, in community health and interpersonal trust. And its advantages—especially in higher education—can seem dubious. It is hard to look past the technology’s vast potential to facilitate academic dishonesty.
Yet AI remains a fixture in higher education. Despite skepticism from students and faculty alike, adoption has continued to increase among both groups since 2023, suggesting that the question is not whether we want AI tools in the classroom, but how we can use them well.
The current situation—high cost and low quality—is the inverse of value. Too much attention to policing misuse and too little to using it effectively encourages students to keep using AI badly. It leads to waste.
But it does not have to be this way. Generative AI, well deployed, can extend student learning. Institutions and instructors can make decisions that help students think critically about how they approach their coursework, and that help them see disciplinary knowledge from perspectives that might otherwise not be apparent.
The key to high-value AI is to ask ourselves, with every pedagogical decision, whether learning would be meaningfully better with AI in the loop, or meaningfully worse in its absence. Put a different way, we should consider:
- Where AI’s particular strengths might help us extend student learning.
- Where its drawbacks might get in the way.
- How we can head off the kind of clandestine, low-value AI use that currently drives our academic integrity crisis.
As August 2026’s MIT AI and Education report points out, this type of redesign requires significant effort. Instructors across disciplines must educate themselves on AI’s capabilities to make informed decisions about how to proceed—and where that work falls to contingent faculty, institutions must compensate them for their effort. The result may be that no changes are needed at all, or that syllabus and assignment changes are a light lift. Not every use case requires a full teardown.
And the upshot is that some approaches to AI may in fact save instructors time and effort. Diverting students toward more productive engagement with the technology means less time spent policing AI during grading.
Regardless, the first step is investigation. To make the juice worth the squeeze, we must know how to match the technology’s strengths to our goals. And we must know how to create assignments that allow us to steer the tool, rather than the tool steering us.
LLMs as Pattern-Matching Machines
The large language models that drive generative AI are, at their core, advanced engines for spotting patterns and extending them through statistical processes. Models are trained on vast quantities of information, which they use to predict plausible responses to a user’s queries. This makes them exceptionally good at processing large datasets—an essential function for learning across disciplines.
AI’s use case in quantitative fields may seem obvious. Feed an LLM 500 pages of observations, and it can spot patterns, offer directions for analysis, and write analytical functions in Python and R that streamline students’ workflow. This is convenient. But in this configuration, its learning value is questionable.
Designed with intention, though, that pattern-matching ability can help students see what they might otherwise have missed without doing the work on their behalf. Imagine a public health course in which students are asked to assemble and analyze a dataset from public sources. Once they have a draft, the instructor asks them to feed it to an AI tool for feedback. The tool’s pattern-matching prowess can easily spot biases students inherit from their sources and identify gaps that could distort conclusions. Students verify that feedback by tracing each flaw back to the data. Then they draft a mitigation plan, and the instructor critiques it before the analytical work begins.
In this scenario, AI is an extender, not a replacer. It serves as an extra set of expert eyes, easily completing a task that is distinctly difficult for humans, and making students’ work better while still requiring them to practice their skills.
This approach is not confined to quantitative fields. In a first-year composition course, for example, the instructor might ask students to upload a draft essay to an AI chatbot and have it outline the argument’s logic. This is pattern matching, too, and a particular strength of AI. The output helps students put their prose aside and see the weaknesses in their structure. Rather than doing the rewriting, AI becomes a tool for noticing that can help students close gaps in their reasoning and write a more complete final draft.
Finally, we might use this ability to help students work at scale. Upper-level courses frequently ask students to look deeply at a limited corpus of texts. This is invaluable but can make it hard to see the big picture. An AI-assisted companion assignment might provide disciplinary context. For example, teams might use AI to map the intellectual genealogy of a course reading—a tree of the authors and publications that influenced it. They then validate the AI’s connections, add context from their own research, and present it to the class.
In all of these cases, the same pattern-matching algorithm that could allow AI to replace student cognition extends their learning in ways that would be difficult or impossible without it.
LLMs as Palimpsest
AI’s other great pedagogical advantage is its malleability. Models come with a personality preinstalled. But with some planning, users can overwrite that default. We cannot entirely overcome their predisposition toward certain characteristics like sycophancy. But give them a detailed character sheet—background, motivations, limits of their knowledge, etc.—and they will play that role.
In a language class, for example, an AI chatbot might make an ideal conversational partner. Most models are natively multilingual. They can be instructed to be patient and give minimal correction. And users can converse in writing or aloud, with real-time voice tools. Sustained conversation is resource intensive and should be used sparingly. However, for many students, it has the advantage that because chatbots are inanimate, they can provoke less anxiety than performing for a more proficient human partner.
After each chat, the AI can write a report on areas where the student did well and areas where speaking, writing, and comprehension need work. And those can become a blueprint for personalized learning.
Similarly, courses teaching community interaction, AI provides an excellent practice venue. Chatbot characters cannot replace humans, but they can improve human interactions when they happen. For example, if business students are asked to talk with executives about management problems, talking first with a virtual executive can instill confidence, as well as a preliminary sense of what questions will land. Or in nursing classes, roleplaying with virtual patients first can make live interactions more fluid, allowing instructors to concentrate their critiques on issues that can only be addressed in a face-to-face encounter.
We can use the fact that AI models are palimpsests—surfaces we can write a new role onto, even though the original layer is never fully erased underneath—to help students get better practice at lower cost, even as we are cognizant that the underlying tendencies of the model never entirely go away.
Where to Go Next
The point here is not to dismiss AI’s harmful influence on higher education. Because AI use is so prevalent, and because misuse presents such a challenge, instructors and administrators must be prepared with both a carrot and a stick.
The stick is the academic integrity policy—give students guidance about unacceptable behaviors, and reinforce those points with clear adjudication procedures and information about the hidden cost of AI. Tell students about the datacenters and erosion of trust. Enumerate the consequences of cognitive offloading and cognitive surrender.
But temper that with the carrot of innovative pedagogy. Coursework that models good practices and shies away from wasted tokens teaches students both disciplinary knowledge and essential digital literacy. It teaches how to balance the cost of the technology against the benefit they stand to gain, a calculation instructors must make in their own designs, too.
Situations where a lack of guidance leads students to use AI to replace their own cognition are not high value. They are waste. Value emerges when AI extends human capabilities, letting students and instructors see their specific work and their fields of study in new ways, and helping them practice skills more effectively.
This requires institutional support. At its most basic, that means money in the form of licenses for AI services that contractually guarantee student access and privacy. Without this, we shift the cost of AI onto students and price low-income students out of their own coursework. It also means a comprehensive strategy, matched to institutional priorities. Policies must provide clear guidance for what is in and out of bounds, support for instructors working on innovative ideas, and training options for instructors who are not there yet.
This is an opportunity we can make worth the price. AI, used well, can help students meet learning objectives through novel insights and low-stakes practice while teaching them about the tool itself. By modeling responsible use of a costly item, we can move away from asking whether AI is good or bad, and start asking what it is actually good for.
References
Cardon, Peter W., and Anthony W. Coman. “Professionalism and Trustworthiness in AI-Assisted Workplace Writing: The Benefits and Drawbacks of Writing With AI.” International Journal of Business Communication, published online July 16, 2025. https://doi.org/10.1177/23294884251350599.
Carter, Jacob. “The AI Data Center Boom Is an Environmental Justice Crisis.” National Wildlife Federation Blog, May 7, 2026. https://blog.nwf.org/2026/05/the-ai-data-center-boom-is-an-environmental-justice-crisis/.
Digital Education Council. AI in Higher Education Global Survey 2026. Barcelona: Digital Education Council, 2026. https://www.digitaleducationcouncil.com/resource-library-items/ai-in-higher-education-global-survey-2026.
Hemmer, Patrick, Max Schemmer, Niklas Kühl, Michael Vössing, and Gerhard Satzger. “Complementarity in Human-AI Collaboration: Concept, Sources, and Evidence.” European Journal of Information Systems 34, no. 6 (2025): 979–1002. https://doi.org/10.1080/0960085X.2025.2475962.
Kapoor, Maya L. “The Largest Study of AI Use by Undergrads Is In, Revealing Disparities in Access — and in Cheating.” Berkeley News, May 21, 2026. https://news.berkeley.edu/2026/05/21/the-largest-study-of-ai-use-by-undergrads-is-in-revealing-disparities-in-access-and-in-cheating/.
Noorbehbahani, Fakhroddin, and Kiemute Oyibo. “AI-Overdependence and Human Cognitive Decline: Hazards, Evidence, and Mitigation Strategies.” Computers in Human Behavior Reports 22 (2026): Article 101102. https://doi.org/10.1016/j.chbr.2026.101102.
Sallustio, Joe. “AI in Higher Education Is Moving from Experimentation to Strategic Integration. Here’s What the 2025 Data Shows.” Ellucian (blog), April 2, 2026. https://www.ellucian.com/blog/ai-higher-education-2025-survey-findings-move-strategic-integration.
Seghier, Mohamed L. “An Active Human Role Is Essential in Big Data-Led Decisions and Data-Intensive Science.” F1000Research 10 (2021): Article 1127. https://doi.org/10.12688/f1000research.73876.1.
Shaw, Steven D., and Gideon Nave. “Thinking—Fast, Slow, and Artificial: How AI Is Reshaping Human Reasoning and the Rise of Cognitive Surrender.” The Wharton School Research Paper, University of Pennsylvania, January 11, 2026. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646.
Adam D. Zolkover is Director of Instructional Design and Technology Services at Haverford College. Prior to that position, was Associate Director for Curriculum Design and Online Education in the University of Pennsylvania’s Department of Medical Ethics and Health Policy, where he helped launch the Master of Health Care Innovation. In his teaching and research, Adam is interested in developing pro-social, pro-education uses of generative AI.
