The University in the AI Era

The University In The AI Era

Carson Gross June 11, 2026

As I mentioned in “Yes, And”, I teach computer science at Montana State University.

In that earlier essay, I say that computer science is probably still a reasonably good area to study, but that you should also expand your skills beyond “just” computer science to help make yourself more employable in the future.

In this essay I want to think more about what AI means for universities in general and computer science programs in particular.

Note: I apologize that this is a longer essay. I have provided a Table of Contents to help you navigate it.

Table of Contents

First: Is The University Still Relevant?

Writing Code

Signaling Competence in an AI World

Towards An AI-accepting CS Curriculum

Current Changes

Homework Is No Longer A Strong Signal

Homework Can Be More Ambitious & Realistic

AI is a Great TA

The Return of Butt-in-chair, Handwritten Tests

Demos & Visualizations Are Cheap

Class Content Should Be In Markdown

Class Analysis & Improvements

Automate Everything

Upcoming Changes

Stronger Pseudocode Standards

AI & Non-AI Tracks

Open Source Work

Clearly, Honestly Communicating The Dangers of AI

Speculative Changes

The “CS+” Concept

Network Isolated Computers

Interview-Based Grading

Conclusion

First: Is The University Still Relevant?

An initial question that many people are asking is: in the era of AI, is the University still relevant?

This is not a new question. Many people have pointed to famous software industry figures who dropped out of college as proof that a university education isn’t useful in technology. And most people who have worked in Silicon Valley know at least one excellent engineer who either dropped out or simply never went to college.

So a college degree has never been a hard requirement for a successful career in technology. But, in reality, most software engineers have some sort of college under their belt and many of the best developers have studied computer science in their undergraduate education.

That being said, there is clearly an emerging crisis in Computer Science education that needs to be addressed in order to keep the university relevant in the post-AI world.

Writing Code

Historically, many computer science departments have looked at writing code as a secondary skill, to be picked up by students on their own, while the department focuses more on the theoretical foundations of computer science.

Since I was mature enough to have an opinion on the matter, I have viewed this as wrongheaded: I think you need to learn how to write code in order to appreciate those deeper theoretical foundations of computer science. If you can’t code up a linked list or use a hash table effectively, learning about the big-O behavior of them is much more abstract and difficult to grasp.

Ironically, in the era of AI, many professional environments are also starting to look at raw coding somewhat skeptically, sometimes insisting that their own engineers not write code at all, but rather use agents to generate it.

This approach may work for more experienced seniors, who have already written a lot of code and know what reasonable code looks like, but it puts junior developers in a bind: they don’t have pre-AI experience writing code, and now they are going into environments where no one is writing code.

As I said in “Yes, And”, you must write the code if you want to develop the ability to read code.

How is that supposed to happen at companies where nobody is writing the code?

I think this presents an opportunity for Computer Science departments: we can be the places where young software engineers write the code. By refocusing our curriculum on practical, code heavy assignments we can give students a safe environment, free of the time pressures and demands of corporate work, to write the code.

This experience can then put them in position to go into environments that use AI more heavily with the confidence that they know how to code and, because of that, that they are in a position to read and understand the code necessary for their career.

Signaling Competence in an AI World

Now, of course, students are famously lazy and famously clever in figuring out how to be lazy. So, many students will use AI to complete many of these code-heavy assignments. They will learn very little or nothing, but will get a good grade because, let’s be honest, AI can perform at or above the level required for most reasonable undergraduate projects.

Here another irony of the AI era becomes evident: Universities are now in a position to signal competence in a way that nearly no other institution can. AI has made online testing pointless. I know this because the last semester I offered online tests (which I like to do because it is convenient for my working students) the testing scores were through the roof.

While I feel I am a pretty good teacher, this was clearly a case of AI being used by my students, despite my pleas.

When thinking about what I could do about this I realized that we had all the infrastructure for the perfect answer: in person, on paper testing. Universities have large lecture halls, expensive printers, testing centers for people who need additional help, etc.

Previously, I would have scoffed at this infrastructure as antiquated. But now I see that it puts me in a position to more accurately establish the competence of my students in a way that is difficult to game: I offer in-person quizzes, with one page of handwritten notes and no digital equipment, roughly every three weeks of my courses.

Of course students can still cheat, but the quizzes are proctored and now at least they have to work for it.

This in-person, on-paper testing infrastructure puts universities in a unique position to provide a high signal-to-noise indication of the competence of their students to the outside world.

Towards An AI-accepting CS Curriculum

While I believe that the University CS degree is not only still relevant, but perhaps of more value than it was in the pre-AI era, I do think that significant changes need to be made to adapt to the new state of affairs. In this section I will describe what I have done over the last year with my courses, what I plan to do in the near future & then finish with some more speculative changes that I believe would help increase the usefulness of undergraduate CS degrees.

Current Changes

First, let’s start with the changes that I have already made to my courses over the last year to deal with the new AI reality.

Homework Is No Longer A Strong Signal

As with take home quizzes, due to the use of AI, homeworks & projects are no longer a strong signal of a students understanding of material.

Homeworks must become for the student’s benefit, opportunities for them to learn the art of writing code, rather than for evaluating their competency.

This is actually a good thing: homeworks can be more ambitious and the students that want to learn will have more opportunities to write more advanced code.

Yes, some (many?) will cheat on assignments, but the good students will have an opportunity to write code in a supportive environment.

To address this fact, I have reduced the weight of assignments in my classes from 60-80% (I have always had code-heavy classes) down to 50%, and I expect most students will get A’s on most assignments.

Homework Can Be More Ambitious & Realistic

Another homework related change that I have made is that my assignments are now more ambitious and realistic. I don’t mean they are much harder.

Insead, what I mean is that I can, with the help of AI, present much larger software systems to my students with better sample data, as a basis for their projects.

This allows students to see software systems that go beyond “Hello World” levels of complexity and to develop the ability to navigate, read and write code in a larger, more complicated and realistic context.

AI is a Great TA

Another thing I have found in the post-AI era is that my office hours traffic has dropped precipitously. I have always done my office hours in the computer lab on campus and, particularly for my compilers class, expected large crowds of students to come in asking for help on projects.

I think, unfortunately, this is most likely due to many students using AI to solve their programming problems.

However, there is a more optimistic read here: the students are using AI to better understand the projects and therefore do not need as much one on one help.

While I am ambivalent in many ways towards AI, this is an area where AI can significantly improve the university experience for students: with proper use, AI can be a fantastic TA. It is infinitely patient, has no other students waiting in line or it’s own classes to attend to and it is usually very competent at undergraduate level concepts in computer science.

The danger, of course, is that students simply use AI as a code generator to complete assignments and head off to the bars.

To address this danger, I ship a CLAUDE.md/AGENTS.md file in my class repos that directs AI agents to act like a good TA rather than a code generator.

Of course students can modify or delete this file, but there is no system so perfect that no one needs to be good.

Stanford University has recently modified this file for one of their own classes, and I encourage other people and departments to do the same: it is public domain.

The Return of Butt-in-chair, Handwritten Tests

As I discussed above, Universities have infrastructure for in-person testing that make them uniquely qualified to assess expertise and competence in the post AI world.

I have switched to all in-person quizzes, roughly every three weeks. The three-week cadence gives enough time to cover a significant amount of material, even if holidays are interspersed in those weeks, while de-escalating each quiz when compared to a traditional midterm/final setup.

I also allow one page of handwritten notes. I do not allow printed notes. The idea here is to force the knowledge through the student’s eye-brain-hand pathway multiple times in order to help reinforce it.

My students have grumbled about this process, but also admit that it works in helping them learn the material.

My questions are all written response, never multiple choice. Sometimes I ask for prose, sometimes I ask for pseudocode, sometimes I will provide code and ask students to annotate/explain it, etc. This makes it harder to grade the tests, but also makes it much harder to cheat.

I have found that AI is very good at suggesting questions based on class material for quizzes. I will work with an AI agent based on my class slides (see below) to create appropriate quiz questions and then create a quiz review sheet to help students study for the quiz based on it.

Students love the review sheet because it helps them focus their studing efforts.

I think that, from a learning perspective, the butt-in-chair quizzes have been the single most positive change I have made to my classes. I now make quizzes 50% of a student’s grade, and my class grading curve has returned to a reasonable shape.

Demos & Visualizations Are Cheap

Another adjustment I have had to make is that demos & visualizations are now very cheap to create with AI.

For a long time I was unhappy with the computer emulators that were available to me to teach my computer systems class. I wanted a 16-bit computer that struck a balance between the simplicity of something like The Scott CPU and the full complexity of something like SPIM.

Two summers ago we spent an entire summer building such a computer, called The Montana Mini Computer that provided strong visualizations of how low level computing works.

Unfortunately, when I got into a class using it, I realized that the architecture I had picked was too exotic (mixing concepts from MIPS & the JVM) and that students would be better off learning an assembly closer to x86.

[truncated for AI cost control]