Last year Dan Myers and Anne Murdaugh introduced generative AI in several courses at Rollins College. But they did more than tell students it was OK to use those tools in assignments. They required them to.
Students completed semester-long research projects using Claude and Copilot to brainstorm paper topics, conduct literature reviews, develop a thesis, and outline, draft, and revise their papers. At each step, students used logbooks to write down the prompts they used, the responses they received, and how the experience shaped their thinking.
The two professors were impressed with how engaged the students were in the process, while the students described where they thought the AI had helped, such as in brainstorming and outlining, and where it wasn’t as useful. Meaningful literature reviews still required lots of independent work, for example, and many found AI writing underwhelming, preferring their own words.
The experiment illustrated to Myers, an associate professor of computer science, and Murdaugh, an associate professor of physics, something they had been thinking about for a while. What if, instead of outsourcing work to AI, students learned how to learn more effectively by working alongside it?
“The conversation for us has been less about how do we radically reimagine things” says Myers, “and more about what are the right places in the curriculum to put touch points so that every student is going to get the chance to build these fundamental AI skills?”
This shift toward collaborating with AI doesn’t unsettle Myers or Murdaugh in the way that it has many professors. The reason, they say, is that the skills that students use to engage thoughtfully with AI are the same ones that colleges are good at teaching. Namely: knowing how to obtain and use information, thinking critically and analytically, and understanding what and how you’re trying to communicate.
This year, more professors are moving in the same direction. Two years after ChatGPT, the first user-friendly large language model, exploded onto the scene, colleges themselves have begun wrestling with how to incorporate generative AI strategically across the curriculum. Given that it may eventually take its place in the pantheon of game-changing technologies used every day in education — alongside calculators, search engines, and Excel — the questions posed by its existence are foundational to teaching and learning.
Colleges are debating what AI literacy means for their students and whether that requires new courses and revamped majors. They are considering how to develop AI fluency among faculty members, many of whom feel overwhelmed by the technology. Some institutions have created state-of-the-art tools and systems: so-called “walled gardens” in which people can experiment without fear that their work will be used to train future iterations of the technology.
If 2023 was the year in which professors began adding AI-use policies to their syllabi and AI-detection tools to their practices, this academic year may become known for campuswide advances.
Recent surveys highlight one of the tensions that colleges face as they dive into these deeper waters: namely, how to encourage students to use AI thoughtfully when the vast majority of professors say they themselves don’t know how.
A survey by the Digital Education Council, which includes universities and corporations from around the world, found that a large majority of undergraduate and graduate students are using artificial-intelligence tools. Nearly 70 percent use AI as a search engine; 33 percent use it to summarize documents; and 24 percent use it to create first drafts of work. Another survey, commissioned by the Walton Family Foundation, reported that 49 percent of college undergraduates said they were using ChatGPT every week. More than 60 percent of them used it to study for tests and quizzes.
Meanwhile, a recent survey of instructors by Ithaka S+R found that while two-thirds were “somewhat familiar” with AI tools only 18 percent said they understood how to apply it in their teaching. And there remains deep skepticism among faculty members about generative AI’s value. Only about one in five thought it would benefit teaching in their field. Most — 56 percent — were simply unsure.
Omid Fotuhi, director of learning innovation at WGU Labs, an arm of Western Governors University, is not surprised to see so little organized movement around AI thus far within higher education. Faculty and administrators see its potential benefits, he says. “But unless they have confidence, or examples, of how it can be implemented effectively, then there’s this reluctance. There’s hesitance around how we implement, what scale, what supports. And so what we’re seeing is a lot of inaction.”
In his lab, Fotuhi and his team are testing and prototyping AI tools to share with a group of institutions called the College Innovation Network, helping provide those concrete examples.
Another organization hoping to remove some of the uncertainty around AI is the American Association of Colleges and Universities. This fall it began a seven-month-long institute on AI, pedagogy, and the curriculum, led by C. Edward Watson, an expert on teaching and learning who is now the group’s vice president for digital innovation. Interest has been “immense,” Watson says, and the institute is working with teams from 123 campuses to help achieve their goals. The most common goal, Watson says, is to determine how to integrate AI into the curriculum. The second most common is figuring out a faculty-development strategy. Ethical considerations come in third.
Watson attributes this shift toward strategic thinking in part to faculty members looking beyond what’s happening in their classrooms and into the world their students will enter. “If you think about what students require post-graduation, it’s hard not to embrace AI as a learning outcome.”
If you think about what students require post-graduation, it’s hard not to embrace AI as a learning outcome.
That’s what brought Berry College to the institute. “We’re not saying every major needs to do this,” says David Slade, Berry’s provost. “We’re starting at a little broader point to say, ‘We recognize we need to teach our students how to use AI ethically and effectively. So what are the learning outcomes around that?’” Ethics and safety, in fact, thread through all of Berry’s goals. It wants to offer faculty members reliable tools to explore AI, it wants to keep people’s data private, and it wants to develop “more sophisticated and supportive” ways to talk about academic integrity, says Slade.
That includes making some tough decisions about whether to use AI detection tools. “I don’t know if it’s the best use of our time to put together an internal policing unit,” he says. A small college like Berry is never going to be able to stay on the cutting edge of technology but that’s OK. “I see it as our obligation to teach students to be thoughtful users, thoughtful citizens. To be critical about how they ask questions and handle responses,” Slade says. “Access to LLMs doesn’t necessarily help shape that in a student.”
Indiana State University is one of a handful of institutions sending several teams to the institute. Their various starting points and interests illustrate how complex AI challenges are.
A team representing the College of Business has several members who have experimented with AI, notes Aruna Chandra, a professor of entrepreneurship. Her dean wants to encourage all professors in the college to adopt AI in their classes in whatever way possible. But over in the College of Arts and Sciences, “we have a lot of departments that are in the scared mode,” says Christopher Fischer, the associate dean, who is leading his team. They’re thinking, he says, “How do we work around AI, in history, English, languages?”
A third group, representing health and human services, is working with the knowledge that AI is already being used in the healthcare industry, says Catherine Paterson, a professor of applied medicine and rehabilitation in the College of Health and Human Services. Her team’s big question: “What type of environment are our graduates going to enter into, and how best can they meet the demands of that industry?”
All three say a central challenge is developing AI literacy among faculty members in order to build coherence toward AI in the curriculum. “We’re quickly approaching a moment where this volunteerist, incrementalist approach may not be sufficient,” says Fischer. “How do we make professional development around an issue this wide sweeping and this pervasive something that we do as an institution or as a faculty?”
A number of large, well-resourced institutions are taking the lead on several AI fronts. Arizona State University, the University of Michigan at Ann Arbor, and Yale University have built their own platforms to allow faculty members and others to create customized chatbots, among other tools, to be used in research, teaching, and other activities on campus. One of the benefits of this approach, in addition to providing a layer of privacy, campus leaders note, is that it helps put students on equal footing. Those who cannot afford, say, the latest version of ChatGPT aren’t stuck with weaker tools than their wealthier classmates.
Yale released a report earlier this year from an AI task force outlining changes already happening on campus and those yet to come, including new courses in engineering and computer science, and anticipation of AI’s impact on language and physics instruction. But the university as a whole is at the beginning stages of considering what AI’s impact on the curriculum should be, notes Jennifer Frederick, associate provost for academic initiatives at Yale, and executive director of the Poorvu Center for Teaching and Learning.
That reflects both the decentralized nature of higher education — schools and departments are where a lot of these conversations occur — as well as differences of opinion within disciplines. “I’ve heard computer-science instructors talking about, ‘Do we really even need to teach coding anymore?’” Frederick says. “And for some people, the answer is probably no. And for some people, the answer is probably yes, because of the conceptual understanding that it brings. But these are the kinds of questions that instructors need to be asking because we have tools at our disposal.”
While smaller colleges lack the resources of places like Yale and Michigan, it can be easier for them to wrestle with some aspects of curricular reform. Rollins’s Myers, for example, notes that he is one of two computer-science professors there who teaches an introductory course in the discipline. The challenge of changing the curriculum “is not a huge concern for us,” he says. “But if you have a dozen sections of programming or physics, many taught by TAs or lecturers, it’s hard to get everyone on the same page.”
He is revamping his intro course this fall to think in terms of online and offline skills. The latter are those skills he wants students to master without the aid of technology. So he will do all of that skill building in class, to include reading and understanding program code. And students will do most of their coding practice during in-person labs. Assuming students will use AI outside of class, he is planning more expansive projects in which he will ask them to document the design process in the same way he did with the research and log books in his prior courses.
To accomplish all of that, he says, he has to cut content. Traditionally the last quarter of the semester would be an introduction to web programming, which he will no longer do. Again, that’s not a major challenge, he says, because his department is small enough that the faculty can adjust subsequent courses accordingly.
Aside from intradisciplinary challenges, AI’s relevance is so dependent on what’s being studied that it’s tough for colleges to define AI literacy. Are students supposed to get under the hood to study how LLMs work, consider the technology’s impact on the economy and the environment, or something else entirely?
To have AI do a kind of intelligent search and summary would put more power in the hands of the public-health professional more quickly.
This fall the department of writing and rhetoric at the University of Central Florida, in collaboration with the philosophy department, unveiled an interdisciplinary certificate in AI. Stephanie Wheeler, the undergraduate director who oversees the curriculum and supports students in the certificate program, says the purpose is to help professors and students develop conceptual knowledge about and critical inquiry into AI, not to build technical expertise you might find in computer science or engineering. Writing and rhetoric have long considered how technology shapes these disciplines, she notes, so this is a natural progression.
In the School of Public Health at the University of Michigan at Ann Arbor, curricular reform comes with different considerations.
Sharon L.R. Kardia, senior associate dean of education, says AI holds enormous potential to benefit public health through its ability to aid in data analysis, research review, and the development of public-health campaigns. Students still have to master the foundational skills without the use of AI, she says. But once that is accomplished, AI tools can save time and deliver insights. “Some health policies are thousands of pages long,” she notes. “To have AI do a kind of intelligent search and summary would put more power in the hands of the public-health professional more quickly.”
At the same time, Kardia says, large language models absorb and reflect the social biases that lead to public-health inequities, which is one reason her school is “stepping lightly” into AI skill development. To that end, students in Michigan’s masters-level public-health program are being introduced to AI tools such as Perplexity, which provides links to its source material, to help conduct research and analyze the results. Eventually, though, faculty members will have to consider AI applications at all levels of public-health education at Michigan.
“It’s kind of huge,” Kardia says. “And, to be honest, I can have this conversation with only a handful of faculty without their eyes completely glazing over and them saying, ‘Oh my Lord, I have way too much on my plate to be able to handle this.’”
That such feelings are common even on campuses like Michigan, which has a relatively deep bench of AI talent and infrastructure, says a lot about faculty uncertainty about AI’s effect on teaching. Workshops on how to use these tools go only so far, say teaching experts. Professors crave concrete, discipline-specific, tested examples of how to use AI in the classroom. “Intuition is great,” says Chad Hershock, executive director of the Eberly Center for Teaching Excellence and Educational Innovation at Carnegie Mellon University. “But we aspire to have more than that to inform our next decision about course design and delivery.”
Carnegie Mellon began funding a series of teaching experiments by faculty members in the spring of 2024. These studies attempt to answer some common questions. Does using AI while brainstorming generate more or fewer distinct ideas? Can a generative AI tool give less-experienced students a better chance to be successful in technical courses? To what extent does using AI help or hinder writing skills? Carnegie Mellon is also supporting a series of experiments considering single questions across courses, to factor in different classroom contexts. One question is: Does having generative AI as a thought partner enhance students’ ability to make a claim and support it with evidence?
The findings for all of these studies will be analyzed and published, says Hershock, so other professors can learn from them. Still, he notes, the results of any teaching experiment are hard to extrapolate beyond the environment in which it was conducted. He encourages faculty to start by reviewing their teaching goals and then consider how AI tools might help. Can an AI tool, for example, help students with review and feedback of the material they are learning?
Some faculty members thinking about how best to leverage AI for learning are gravitating toward the development of AI tutors. The idea is appealing because, under the right circumstances, the content, as well as the tutor’s “personality” — that is, how it responds to students’ prompts — can be controlled by the instructor.
Greg Kestin, a physics lecturer at Harvard University, released a study earlier this year showing that an AI tutor he created for a physics class increased students’ engagement and improved their test scores. He trained the tutor exclusively on course materials to cut down on the likelihood of hallucinations, which are common in chatbots trained on internet content. Kestin and his co-authors also used best practices from pedagogical and psychological research to structure how the tutor behaved, so students would receive guided responses and feedback, and couldn’t skip over the hard work of learning a concept or solving a problem.
The study, conducted in the fall of 2023, was small. Kestin and his collaborators designed it for about 200 students in a physics course for life-science majors. In one section, students participated in standard active-learning classroom activities, taking pre and post tests on the concepts they were studying that day. The other group studied the same material from home using the AI tutor. Then those two groups switched during another class session.
The students using the AI tutor learned more quickly and scored higher. “A lot of the comments,” Kestin says of the responses they received from students, “were about how the students could go at their own pace and could have this unfettered access to a nonjudgmental instructor who would not get frustrated or annoyed no matter how many times they asked.”
Of course, that fuels a dystopian vision for academics when it comes to artificial intelligence. If a robot can perform better than a professor, are human teachers even needed?
But Kestin says their experiment does not show that AI can replace teachers. In fact, he drew a very different conclusion: That a well-designed AI tutor can help students build basic, foundational knowledge on their own time. That means that professors could spend time in class with students wrestling with more complex problems.
Harvard was pleased enough with the results of the experiment that it is piloting similar AI chatbots in a few large introductory courses this fall. And Kestin says he has been talking to professors at MIT, Yale, Stanford, and elsewhere to help pilot AI tutors there as well.
Other universities are also experimenting with AI tutors, including some at scale. Michigan’s generative AI infrastructure — also available on the Flint and Dearborn campuses — allows faculty members to create chatbots that are trained exclusively on their course material. A service called Maizey allows people to develop such AI tools inside walled gardens across campus which protects their privacy, says Ravi Pendse, vice president for information technology. About 2,500 Maizeys, he says, are in production in many different departments and divisions, in and outside of the academic side of the house.
Kardia, the senior associate dean in public health, believes AI tutors can be “the bridge for faculty to be able to trust AI and get creative in its use in teaching.” To that end, her school is piloting virtual teaching assistants in three courses this fall.
I would not for the moment let a chatbot loose in my class.
There are other ways in which faculty members have adopted AI more quickly. Since ChatGPT first showed up on the scene, professors have experimented with AI tools that aid in course design and assessment, such as turning lecture notes into PowerPoint slides, generating multiple-choice exams and designing learning outcomes for a course. AI’s ability to streamline these necessary but time-consuming tasks has created AI fans in academia.
Melanie M. Cooper, a chemistry professor at Michigan State University, and Mike Klymkowsky, a biology professor at the University of Colorado at Boulder, are hoping to do one better. They want to design an AI tool that supports more complex assessments, so that STEM professors who teach large courses can move away from multiple-choice exams, which can’t reveal how well students grasp foundational concepts or reason their way through problems. The challenge, the professors note, is that designing, grading, and giving feedback on more complicated assessments is time-consuming. Yet students learn best through such explanatory work.
A well-designed chatbot could potentially reduce this time, Cooper and Klymkowsky note, if it could generate activities and assessments, aggregate students’ responses to highlight patterns in their understanding, and serve as a tutor by providing feedback and support. Michigan State was recently awarded a grant from the National Science Foundation, through which the two professors and others will collaborate on the design and testing of some of these ideas in general chemistry courses.
Yet, Cooper notes, this is an extraordinarily complicated project to pull off. How do you train a chatbot to provide the right kind of interactions? “There’s a lot of ebullience in the AI field,” she notes, but “it’s important to be wary.” It’s quite easy, for example, to misuse the AI and override the system to get a quick answer or otherwise use it as a crutch. “I would not for the moment let a chatbot loose in my class.”
Such concerns are at the root of a lot of the push-pull in higher education right now over AI. Professors are worried about employing AI tools that may spit out wrong answers or allow students to avoid the hard work of learning. Academics are also quick to reject the rhetoric of AI evangelists who promise that their tools will make learning easier, faster, and more fun.
But they are also readier now to talk about how understanding AI is a necessary part of higher education’s responsibility, whether that means knowing how to code with AI, debating the ethical implications of its existence, or exploring its role as a workplace technology.
Of course, in many places such inroads have yet to be made, for lack of time and resources, or are in very early stages, because of the staggering challenges ahead.
Meanwhile colleges will continue to explore how to leverage their limited resources to provide the most useful and safest technologies, along with relevant and useful professional development and training to faculty and students. It’s a big lift, especially considering that generative-AI tools are advancing at a rapid clip. Just recently OpenAI came out with a version of ChatGPT that can perform at the Ph.D. level, opening a debate about how close we are to artificial general intelligence — one that’s not just geared to specific tasks. But that makes it all the more important, academic leaders say, for colleges to dig in.
“Universities really need to be a counterpoint to the big tech companies and their development of AI,” says Frederick, of Yale. “We need to be the ones who slow down and really think through all the implications for society and humanity and ethics.”
