The University of the Future
17 min read
What would a university be doing if it were fully embracing AI? Not just teaching students to use AI effectively, which is tremendously important, but leading by example: educators integrating AI into their own classrooms and pushing the envelope of what these tools can do.
We are just starting to understand how to use them.
These things are brand new.
The number of people using them seriously is very small, and it is still far too hard to get access to them.
Andrej Karpathy recently described the gap in understanding AI as a funnel: billions of people who have barely touched a language model, a billion or two casual users of free chatbots, maybe twenty million people working with frontier models every day, and a few thousand who see what is coming next.
Here are eight things I would look for at a university that was serious about this, in rough order of importance:
- Requiring and supporting experimental, evidence-based redesign of existing courses and degree programs, expecting that nearly all subjects fall into that category but prioritizing ones most directly affected—such as computing
- Addressing institutional barriers that impede adaptation, such as long delays for approving new courses or degrees
- Training faculty to use AI to improve their courses and pedagogy
- Celebrating and rewarding the ways that educators integrate AI into their classes to support student learning, motivation, and curiosity
- Making frontier models safely available to faculty, students, and staff, with as few usage limits as possible
- Creating facilities where student learning can be assessed either without or with controlled access to AI and other online resources
- Adjusting IT policies and support procedures to accommodate the growing number of educators who can build and deploy pedagogical technology
- Banning AI detection software and eventually requiring approval for course policies prohibiting AI usage
Redesign Courses and DegreesRedesign Courses and Degrees
This is first because it is the big one: the one we will be working on the longest, and the one that will reshape the entire institution. Before computers there were no computer science departments and no colleges of computing, and computing went on to reshape how we teach many other subjects. AI will have the same impact. There will be entire schools, departments, and curricula devoted to it.
This is the first thing I look for when I look at what other universities are doing. I usually don’t find much. The lowest-effort move is to highlight existing courses that mention AI. Nothing to see here. A step up are the minor rebrands: certificates that bundle courses you already offer under a new label, or AI degrees that are a computing degree with one required AI course. This is why outside rankings of “AI-forward” universities are mostly nonsense: they reward research and buzzwords, not what is happening in classrooms. Hollis Robbins pointed to the AI Campus Index, which puts Carnegie Mellon, MIT, and Stanford on top “with Berkeley and UIUC close behind.” Faculty here don’t even have access to AI! More on that below.
Meaningful changes are actual work: designing and deploying courses that never existed before, and comprehensive redesigns that leave a course transformed. Not “we’ll allow you to use AI,” which in many cases just lets students skip the work. Fully pulling AI into the classroom and reckoning with what it is capable of, which poses very different challenges in different disciplines.
Next I look for courses designed for non-computing majors. A lot of universities seem to be assuming that computer scientists need to learn about AI but nobody else does. Everybody is going to have to learn this, so you need courses for everybody. Many of the ones I see are AI “literacy” courses. I don’t love the term. Literacy sounds passive, and the courses tend to be passive, a lot of talking about AI: ethics, societal impact, and so on. But we are all implicated in this technology, and we can all use it. It’s the most accessible computing technology we’ve ever created. So I look for courses where students use the most powerful AI models to solve real problems. That’s the only way to learn what these tools are capable of.
I suspect two things hold us back from creating these courses.
One is that faculty aren’t using AI much themselves, so they don’t know what it can do and find it hard to teach.
I’ll come back to that.
The other is fear of what students might do with these skills.
Further up the ladder are new degrees and significant redesigns of existing ones. I think new degrees are often easier, but either is great. These things build on each other: you can’t have a new degree without new courses, much less a department responsible for that degree, built on courses we don’t have and faculty we haven’t hired. And using AI well is a skill. We don’t teach students to compute, or to philosophize, in one semester, and we won’t teach them to use AI well quickly either. Getting good at it will require continued practice across sequences of courses.
Remove Institutional BarriersRemove Institutional Barriers
AI is evolving rapidly. University processes are slow and sclerotic. Before a generational technology showed up, maybe it was fine if it took years to approve new courses, degrees, or units. That’s not acceptable today.
This varies from place to place. I suspect some universities are better at it. Illinois is not. I taught a new course, Using and Understanding AI, last spring and have been trying to push it through the process of getting a permanent number and then general education credit, which is the only way it will reach a large number of students. It has taken forever. It has been delayed and roadblocked multiple times, and it makes no forward progress unless I prod it through layer after layer of review. We should be pulling these courses forward, not requiring faculty to push them.
We are not going to learn how to teach these courses by waiting. We are going to learn by doing. Yes, the courses will have to keep adapting, because the material is changing rapidly. But they need to be out there so that students are taking them, so that we are getting feedback, and so that we can start to build an understanding of how to teach this material. Long, slow review processes make that nearly impossible. This is a process problem, and it can be fixed.
Train FacultyTrain Faculty
You can’t teach what you don’t do.
I’ve written before about how many people teach computer science without doing it themselves.
I’ve always been a software developer, and I bring twenty years of writing code into my classroom.
But some colleagues not only don’t do that, they’ve been trained to be suspicious of firsthand experience: you’re an expert, so you can’t understand how students approach the material.
Follow that reasoning and the best person to teach a course knows nothing about it.
Being an expert doesn’t qualify you to teach a course, but it amplifies the pedagogical ability you pick up as you teach.
What those colleagues are working from is pedagogical content knowledge: the knowledge of how to teach a subject, distilled by somebody else who did have the experience of doing it and wrote down what they learned.
That knowledge doesn’t exist yet for AI. Nobody knows how to do this. We are all figuring it out together, at the same time, and the models have gotten really good, particularly in the past year. To the degree that anyone is skilled at this, it is a very young skill, and it will take a while to systematize. In the meantime, if you want to teach it, you have to be using it. Frontier models, not browser chat. Agents. The latest tools, given hard tasks. That is the only way to know what they are capable of. Otherwise you are teaching students from a very limited understanding.
You can tell when an educator has crossed over. They have a look in their eye, because they just got their hands on an incredible new tool to support student learning. I feel that way every day. A lot of people aren’t there yet, but I don’t want to put that entirely on them. We aren’t putting the tools in their hands. They also need training: guidance, a program that brings them up to speed and shows them what is possible. I was part of one last summer, Illinois’s Summer AI Studio, and everyone I observed in it moved in a positive direction in terms of using AI to teach their courses better. That is what we are here to do.
There is a cultural benefit too. As faculty learn to use these tools to teach better, students see that you can do good things with AI. AI isn’t only here to take your job or do your homework. There are responsible adults using it to do remarkable things. That’s a message we need to get out. And there are entirely new pedagogies opening up: things you can do in a class with an AI model that were never possible. We need educators exploring them together.
Celebrate and Reward IntegrationCelebrate and Reward Integration
We need to be pointing at courses that are doing this well and saying: That’s amazing! Elevating the people doing it, rewarding them, raising their visibility. I get an email from the provost every couple of months. He has said very little about AI, and he has never pointed to an educator on campus doing exciting things with it in their courses and said, look, this is a model. Maybe we need awards for this. Maybe we need to give these people the coin of the realm in academia, a little funding. We joke that students will do anything for one point of extra credit. The faculty equivalent is $1,000.
Observable examples also tell you who to talk to. If you want to know what the future looks like, don’t convene a committee of faculty who aren’t using these tools. Find the educators using them in exciting ways, and put those people on the committee that recommends campus policy. I’ve looked at the membership of some campus-level AI committees and can’t find anything these people are doing in this space. Maybe it’s invisible to me. But the way to find the leaders is to recognize what they are doing in their classes.
Make Frontier Models AvailableMake Frontier Models Available
This is not hard to say, and I don’t know why it is so difficult to do. People need access to these tools. Everybody. Cost has to stop being a barrier. I spend $200 of my own money every month on AI, and another $200 of my school’s money, because I’m using it to do incredible things in my courses. I also have university-provided access for integration into my courseware. I probably have a hundred times more AI access than most educators.
The concerns are familiar: nobody wants to pay for it, and there are worries about FERPA and the other things we fret about on campus. Fine. Make them go away. Sign the deals you need to sign and give people the tools. Don’t put limits in place. If somebody is really burning through tokens, talk to them about it. That might be how you identify the people doing the most interesting work. And these things don’t cost that much. The thing we worry about someone doing in their course costs cents, or maybe a dollar, against the hours a human would spend doing the same thing.
Think about it from an educator’s perspective. If you aren’t willing to spend your own money right now, the students you teach know more about AI than you do. They have better access and they are using better models. How are you supposed to teach them to use something you don’t know how to use? Universities should be desperate for educators to use AI. Heavy usage should be good news: you’re learning, you’re trying things, you’re figuring it out. We cannot learn how to teach this together if we aren’t using it. So: access for students, faculty, and staff, across the entire university. Expensive at that scale? Maybe. We need it anyway.
And these have to be the latest, greatest models.
There is an emerging fascination with running local models on campus because some lab has a few GPUs.
That is probably not going to work.
The compute required to support serious use by everybody on campus is massive.
If you aren’t breaking ground on a data center, you aren’t serious about it, and nobody is.
Build Assessment FacilitiesBuild Assessment Facilities
There are two parts to this. First, there are absolutely things we want students to be able to do without AI, and that is the easiest argument for these facilities right now. As long as I want to teach students to write code by hand, which won’t be for much longer, I need somewhere to put them where they are forced to do that. Computer-based testing facilities are amazing. Craig Zilles and his colleagues have done tremendous work advocating for them, and they are a huge part of this story.
The next phase is more interesting. Even when we are teaching students to use AI, we will want to make sure that it’s the student using the AI and not their agent. I have a new tool that delivers something like an oral exam over chat. Outside a testing center, students might look up answers, but that’s different from telling an agent to complete the assessment. And if a project says “use AI to do this,” a student can still just tell their agent to do the project. We need a human in some of these loops. Today these facilities mostly let students work without AI. In the future I suspect we will also want controlled environments where students are forced to be the ones manipulating the AI: sitting down, typing, communicating with the model, rather than offloading that work to another AI.
Illinois has one.
Other institutions are following: UBC, Berkeley, and UC San Diego have them, and Berkeley has humanities courses using theirs, which is also exciting.
Adjust IT PoliciesAdjust IT Policies
Here is a shift universities aren’t ready for: faculty can now build technology to support learning. For a long time we forced faculty into silos. Thou shalt use Canvas, a remarkably limited piece of educational technology that universities pay millions of dollars for.
But now a faculty member can build their own tools through a conversation with an AI agent. I watched it happen last summer. Ben Clegg had never used AI before, and a couple of weeks later he was building Socratic tutors to help his students learn biology concepts. John Clark has built a whole suite of freely available design tools for architecture students, and is now chaining them together with AI. This is happening.
The problem is that these educators can build things but have nowhere to put them. Their tools don’t fit into Canvas. For years faculty couldn’t be bothered to create their own websites, so everyone was told to use the LMS. Those days are over. A university leading in this space would recognize that we are on the cusp of educators building their own tools, and that campus lacks the substrate to deploy them. They need somewhere to run the code they created, somewhere secure, on campus, and protected in all the ways it needs to be. Right now campus IT mostly doesn’t know what to make of this. I’m the outlier at Illinois. I probably have access to more computing resources for my courses than any other instructor on campus, plus the AI models, because for a long time I had the skills to use them and nobody else did. Now everybody has the skills. They just need the resources. Nothing excites me more than imagining all of the incredible educational technology that my colleagues are going to start building.
Ban AI DetectionBan AI Detection
This is a smaller point than the others, but it sends the right message. AI detection tools don’t work. They are prejudiced against students who use weaker models or use AI poorly. Eventually, I’d argue, course approval should include some oversight of policies that ban students from using AI in every part of a course. There are real arguments for having students do certain things without AI. But I would always start by asking an educator: Can you teach a student to do this with AI, effectively? Can we restructure things a little so that’s possible? Detection tools tell students that using AI is bad. In reality it can be a force for good, if we teach it well. That is what this list is designed to do.
A ScorecardA Scorecard
You can use these as a scorecard. At Illinois we have the testing facility, so that’s number six. We’ve done a little of number three: the program I was part of last summer, for a couple of dozen educators, nowhere near scale. Course redesign? I’ve advocated for it in my own unit and it has gone nowhere, and the new courses I’ve created are stuck in approval. Celebrating and rewarding? Haven’t seen it. Frontier models for everyone? No. IT policies? No. Banning AI detection? No. One out of eight. Maybe 1.5 with a bit of grade inflation. I know some places doing better, with more forward-looking AI policies. But to me these are the signs that an institution is doing the right thing, and the things that educators should be advocating for.
One closing note. None of this is sufficient. This is how a university starts bringing AI into its programs. There is an equally important conversation about how we restructure the university to rebuild intellectual community among students, faculty, and staff. If we don’t, we will be replaced by these tools: teaching assistants first, then faculty. Well, maybe we’ll let them take the administrators first! In many places there is no intellectual community of human connection supporting students’ relationships with each other and with the people who teach them. We have built barriers between faculty and students, and research-obsessed institutions pay far more attention to their graduate students than to the undergraduates whose tuition pays the bills. That more holistic rethinking of higher education, one that brings human values and human meaning back into these spaces, may be the most important conversation to have, and I want to be part of it.
But as far as leadership in teaching and learning with AI, these are the things I’m looking for.