NDSU AI Degree: Teaching AI From the Inside Out

The NDSU AI Degree, Explained

Q: How long has this program been at NDSU? Simone: We are just rolling out the AI program now for the fall. We have had a minor in AI for three years, I believe. We have also had a data science track as part of our bachelor's degree in computer science for maybe six years, and AI is part of that track as well.

We also have a certificate in AI that we launched a year or two ago. Students throughout the university can take it, so it is a nice, entry-level AI certificate.

Q: Did you say computer vision? Sean: Yes. For example, when you log into your phone using Face ID, that is a computer vision task. It literally is one of the examples I use during the course I am creating. How does that work? What is it using? How do we do that?

Self-driving cars are another great example of a computer vision task. Facial recognition on your phone or camera is computer vision. A Ring doorbell identifying that a person was there, as opposed to simply giving you eight hours of uninterrupted recording, is another example.

Computer vision is more ubiquitous and operates more at the infrastructure level, but it has not gained the notoriety that LLMs have. Those are the two technologies that serve as the expansion courses for the AI degree.

Q: Through this degree, what jobs do you see becoming available for students? Simone: Job titles are changing all the time, and I think more of them are now geared toward AI and actually spell it out. Employers want an AI developer or someone who can do work involving AI or LLMs. There is definitely demand for a specialized degree.

Students who are taking the computer science degree and have a minor in AI will also be very well equipped after taking these courses. However, this is a very specialized degree that really focuses on AI.

It is more theoretical. AI degrees are popping up all over the place, but many of them focus only on the application side of things. We still have the foundations, and then, of course, we have the applications. I think that is a major distinguishing factor of our degree.

Q: What are students going to gain from having a major as opposed to a minor and completing that longer stretch of coursework? Will it be more specialized? Simone: It is a specialized degree. We have our bachelor's degree in computer science, which is very general and is still our flagship degree. In recent years, we have added degrees in software engineering, data science, information technology, and cybersecurity. Now, we are rolling out the AI degree.

We are hoping we can attract more students, especially because AI is being used everywhere nowadays.

Sean — It will be interesting to see because there is obviously some pushback from the younger generation as well. The hope is that now, if you have a degree specifically focused on AI and go into it in depth, you will have a particular set of courses that gives you an understanding of how AI is created and how it works.

You could take many of these courses as electives while pursuing a computer science degree. This degree simply means you have chosen a specific set of courses.

The required courses go into large language models, which have become the main thing everybody talks about when they talk about AI, and computer vision, which is not quite as popular but is frankly more commonly deployed right now. Those are the two main technologies we are focusing on.

Three Terms to Know

1. Hallucination — An answer generated by AI that sounds plausible but contains false, invented, or unsupported information.
2. Human in the Loop — A process in which a person reviews, approves, or corrects an AI system's output before it is used.
3. Vibe Coding — A development approach in which a person describes what they want software to do and relies heavily on an AI coding assistant to generate the code. The approach can make development faster, but the resulting code still requires knowledgeable review and testing.

NDSU's Growing AI Education Options

• Artificial Intelligence Major — Launching in fall 2026, the specialized degree will explore the theory, development, and application of AI systems.
• Artificial Intelligence Minor — The university has offered an AI minor for approximately three years, allowing students in other majors to supplement their primary field of study.
• Data Science Track — AI has also been included within the data science track of NDSU's computer science bachelor's degree.
• AI for All Careers Certificate — Designed as an accessible introduction, the certificate does not assume students have a computer science background and is available to working professionals as well as traditional students.

How Large Language Models Actually Fail

Q: Would you say the main difference is between having a program grounded in theory and one grounded only in application? Sean: I would place much more emphasis on understanding the systems. There is some discussion of how you use them, and we are incorporating that into some of our lower-division courses as well. However, the courses I am creating are more about understanding exactly how these systems are created so that you know when they should and should not be used. We have this tool, and various people are saying everybody should be using an LLM, but you need to be very careful about when you should use it and when it can actually benefit you.

I am pushing the courses in that direction so students can understand exactly how the systems are created and exactly where they fail and succeed. Then, if they have a choice, they can determine whether an LLM would be beneficial.

I love the term "automagic." When you see a lot of computer programming from the outside, it is automagic. It just kind of works, and you hope it continues to work.

The point of the upper-division courses in this degree is to move beyond that. You are not going to understand every implication of every decision within a 200-layer system. The AI companies do not understand all of that either. However, you can get an idea of how a system was designed, what its likely properties are, how you go about building it, what issues developers are encountering, and why it takes a couple of billion dollars to train one of these models.

Until maybe the middle or end of last year, LLMs were always trying to be a general solution to everything. The main application now is coming of other specific tasks. With the frontier models, it really comes down to which company had which training data and what it focused on.

In terms of more general failures, yes, they all fail. One thing I have been telling anybody who will listen is that LLMs think very syntactically. They love patterns in speech, but they do not think semantically. They have no concept of whether something is right. They have a concept of whether something looks right.

There is a famous research paper called "Stochastic Parrots" that argues for this idea. Another way I have heard it phrased is that they are "sentence calculators." They create something that looks like a valid, grammatically correct sentence, but that sentence may or may not actually have meaning.

The more knowledgeable that person is, the better. LLMs create very eloquent sentences that may make absolutely no sense. There are human biases built into that. We tend to like and skip over things that sound very fluent. Because of that, we have a bias we are unaware of.

It is legitimately difficult to read random, non-fluent sentences out loud. Because of that, when we interpret data, an LLM can create something that looks fluent to us, and our bias skips over things when it says, "That's good. I don't have to look at it too closely."

As humans, we generally learn the substance of a topic and then learn how to say it fluently. An LLM does not necessarily work that way.

If businesses are going to deploy an AI system, and everybody seems to be deploying one in some manner, they need somebody on the other end who can identify where the problems are.

There are really two aspects to that. One is knowledge of the field. You can get what you want, but you really need someone with knowledge of the field to determine whether that output is correct.

The other is knowledge of how AI systems are likely to fail. Ideally, you would have at least two people involved: one is a subject-matter expert, and one has knowledge of LLMs.

Large Language Models vs. Computer Vision: Large language models may receive most of the attention, but they represent only one major branch of artificial intelligence.

Large Language Models: LLMs identify patterns in language and generate text that resembles human writing. They can assist with drafting, coding, summarizing, and organizing information, but they do not inherently know whether their output is true.

Computer Vision: Computer vision allows systems to interpret images and video. Common examples include Face ID, facial recognition, self-driving vehicle systems, and doorbell cameras that distinguish people from hours of ordinary footage.

Q: What advice would you have for companies or individuals using LLMs? Simone: You have to be careful. You always have to verify whatever you ask ChatGPT or another application to do. Do not trust it all without verification.

In our computer science world, a computer can produce code, but if nobody understands and can verify the code, many things can happen, including dangerous things. It is really important for people to be careful and ask whether the output is correct. They may need to go back and check the sources as well because a lot of hallucination is happening.

Sean — I think of them as accidental evil genies. You can ask them for what you want, but they may totally misinterpret it. They are not doing it maliciously.

You have to realize that they essentially boil down a whole bunch of different decision trees. When you ask a question, they are hoping you get one of the trees that leads to the right answer, but you have no guarantee that will happen.

That is how you get hallucinations. The way you ask the question or the extra information you provide might light up a portion of the network that does not contain what you are looking for.

When AI Goes Wrong

Q: You mentioned that dangerous things can happen when some of this goes unchecked. Do you have any specific case studies you discuss with students that readers might be interested in learning about? Simone: We have two dedicated ethics classes. Usually, our degrees require only one, but for AI, we doubled down and have a course specifically on AI and ethics.

I do not teach that course, but I know we use case studies.

Sean — It depends on the degree of damage you are looking for. I can definitely give you some examples. I did my Ph.D. in privacy and machine learning, so there is a lot of overlap. I have a slideshow of horrible situations involving AI.

For LLM-specific examples, there have been several that are almost darkly humorous. One of the big evangelists for AI, LLMs, and vibe coding recently had an AI system delete his entire home directory, essentially all the documents on his computer.

There was another example in which an LLM deleted an entire production server from an actual working company. The company lost its customer information from the last several years.

Those examples should be available through a quick Google search, and they occurred within the last few months. There are older examples involving AI more generally rather than LLMs.

Q: I assume LLMs are what most of our readers would be familiar with. Sean: If you are dealing with a business audience, they probably will not care as much about the distinction. I would focus on examples from the last six months.

There was the production server deletion. Meta's chief AI officer had reportedly had an AI delete her entire email inbox. That was, I believe, in March.

I have been putting off preparing the lecture about where we are headed in the future because what that lecture would have said when I started in January is very different from what it would say now, even compared with three weeks ago.

The newest systems have become notorious for deleting production systems, to the point that one model's white paper reportedly says it has a tendency to delete things.

The Two-Person AI Check: Before an organization uses AI-generated work, Sean Whalen recommends having at least two knowledgeable people review it.

1. The Subject-Matter Expert — This person understands the industry, assignment, or technical field well enough to determine whether the output is factually and professionally correct.
2. The AI-Savvy Reviewer — This person understands how large language models behave, including their tendency to hallucinate, misinterpret instructions, and produce confident-sounding errors.

Q: Is it difficult to structure courses that focus more on applications when the main programs people are using keep changing and releasing updates? Sean: I am not as involved in that portion. The biggest application emphasis is data engineering.

From what I have seen discussing the courses with others, they are mostly using AI through chatbot integration. Because we are computer scientists, that often means AI coding assistants.

Those systems change, but the changes are almost invisible to users. It may simply be a drop-down menu that determines which model is responding to you.

In my courses, even when I go into particular systems, I focus on foundational concepts. We run local LLMs rather than worrying about the differences among the latest ChatGPT settings. That is easier to do using slightly older models that we do not have to pay for.

There can be some problems, but we mostly deal with the invisible aspect of the changes.

Students may be using VS Code, for example. If VS Code changes which chatbot is technically answering when they ask it to write a small snippet of code, that is largely invisible to them, unless it suddenly starts producing worse output.

So far, AI has not reached the point where the entire paradigm changes every three years. It follows more of the traditional software-as-a-service model. We do not have enough money to maintain racks of GPUs to run the equivalent of ChatGPT or Claude ourselves. Those racks can cost a couple of million dollars. In that regard, we are still essentially dealing with a chatbot and its integration into other systems.

Using AI Responsibly

Q: Do you use LLMs? What do you mostly use them for? Simone: I am the chair of the department, so I have to do a lot of letter writing and recommendations. English is not my first language, so I use LLMs to polish my sentences.

For research, I use them a little to pull together information about a particular topic area, but that is mostly it. I mostly use them for writing because they can write much better than I can.

Sean — I am a native English speaker, and they use better sentences than I can in certain regards.

Honestly, I have been using them to help create my courses. If you think about it, the systems have ingested a large number of university courses. If you ask one to produce something that looks like a university course, it does a pretty good job as a first draft.

Again, I would never trust anything automatically. Nothing goes straight from the LLM to the students. You have to verify it, tweak it, and incorporate your own ideas.

It is very good at creating text once you know what you want it to say and can verify that it is saying it correctly. I can have an idea and describe it in six sentences, and a chatbot can turn those six sentences into something resembling 20 minutes of PowerPoint slides or rephrase it in a more fluent manner.

The issue with chatbots has never been whether they can say something fluently or even eloquently. You just have to make sure what they are saying is accurate.

My brother is also a professor at another school, and one of his co-authors used an LLM to help produce a mathematical proof. They obviously checked the proof, but the system was very good at producing that first draft, and the proof was correct.

There are many applications for LLMs for anybody who has to produce text.

Q: What else do you think the business community should hear? Also, is the certificate available only to current students, or could a working professional enroll? Simone: A working professional can sign up for the certificate or a graduate program. We have a number of certificates available at both the undergraduate and graduate levels.

If someone has schooling in another area but wants to strengthen their AI knowledge for the business they run, they can absolutely take it.

Q: In that scenario, what could someone who runs a business or is already working as a professional gain from the certificate program? Simone: It is an introduction to what AI is. The first course covers the entire AI spectrum, not just the LLMs everybody is talking about, which I think is nice.

We also cover the ethical aspect and have a communications course through the communication department. We have a one-credit self-study course in which we ask students to write a report about how AI is used in their field, whatever that field may be.

It will be beneficial because it exposes students to different AI programs. They also have to write a little code, or ask AI to write some code, and then run it.

We do not assume students have any computer science background. It really is for everyone.

We originally wanted to call the certificate "AI for Everyone." The first course is called "AI for Everyone," but the certificate itself is called "AI for All Careers." That is what the provost wanted, so that is what we ended up with.