Faculty Feature: Professor Patrick Watson on AI, Learning, and the Human Brain

To begin, tell us a little about yourself. You earned both a B.S. in Molecular Biology and a Ph.D. in Neuroscience from the University of Illinois. What drew you to these fields, and how has that background shaped the work you do today?

‍

I came into university to pursue bioengineering but the program got shut down the year after I arrived so I switched to MCB. I've generally been interested in the mechanical workings of biological systems, and neuroscience was a natural continuation of that. I'd really enjoyed my undergraduate courses on neural networks and just sort of kept working in that direction. 

‍

The major way it's shaped my work and understanding is that I come to AI and machine learning from biology. I think it's more common to learn the quantitative tools first and apply them to the biological, but I start with the biological and then try to understand the best way to model it. It means I have a lot more detailed knowledge about brain physiology, cell biology, and anatomy than the typical AI professor.

‍

Studying both biological and artificial intelligence gives you an unusual vantage point. What does the human brain do that today's AI still cannot, and what can AI do that the brain is surprisingly bad at?

‍

One way I often start is by pointing out that when people say AI in 2026 they actually almost exclusively mean "digital media generators." The last ten years have seen relatively incremental progress in AI broadly, it's really just been the rapid commercialization of transformer-based generative models trained on internet content. If you don't use the term "AI" and instead say "internet content generator" this question begins to look a lot different. The brain can do all kinds of things besides generate internet content. 

‍

There's lots of things AI can do besides that as well, but we mostly call those by specific names. Internet search is an AI technology. As is computer vision, speech-to-text, recommendation algorithms, reinforcement algorithms that play games and so on. But all of those are built differently than internet content generators. The brain also works like this--it has a number of different systems that carry out different tasks. Even within a single function like "memory" there are several different dissociable memory systems doing different things. And this is necessary to get things done.

‍

Public conversations about AI often swing between utopian promises and predictions of human extinction. Which risks deserve serious attention now, and which fears distract us from problems already unfolding?

‍

When we want to work against a risk like climate change, we're usually pretty specific about what needs to happen (less carbon in the atmosphere). AI risks are always presented nebulously. If a risk is nebulous, it's best to ignore it and focus on things where you can do some concrete good.

Most of the ways in which digital media generators are a problem are fairly obvious--they're strong vectors for fraud, are straightforward cybersecurity threats, and can have negative consequences in social media and other internet algorithms that rely on human curation. 

These are all areas where there are a combination of straightforward fixes (like requiring social media companies to index the address a posted photo was retrieved from), and a number of less-straightforward-but-ammenable-to-research areas like better encryption. 

I think that the idea that AI should be scary is itself a bit of a con. Tech companies developing AI need to be able to sell the product as "powerful" and they need to spread that idea on the internet where fear and anxiety are more rapidly shared.  There are dangers from AI but they're mostly ordinary. The existential threat ideas mostly come from armchair internet philosophers who don't know anything about either AI or how difficult it is to actually engineer a physical change in the world. 

‍

What does intellectually ambitious and responsible AI use look like in your classroom? How do Minerva students learn to challenge these tools without either rejecting them or trusting them uncritically?

‍

As the machine learning professor I tell students on day 1 that A) they can use AI for whatever they like and B) that I have generally found that using AI results in worse learning and grades than completely avoiding it.  Coming from a neuroscience background, I think of learning as building a little model of something in your brain. If you want to learn linear algebra or something, you're going to move neural connections around until you have a little machine in your brain that can (somewhat) reliably perform linear algebra operations. My job as a teacher is to set up the conditions where students can construct those little machines. 

Now sometimes as a teacher, I find it useful to run a diagnostic on the student's brain machines by asking a question or having them perform some knowledge about the thing I'm teaching. This helps give me an idea of how well the machine is working so far and what might productively be done to fine-tune it. I fundamentally don't think that anyone ever "knows" something. If someone gives me the "right" answer that's my cue to just construct a harder question. 

I think that this is where students and teachers get into trouble with AI. They mistake "answering questions correctly" for "building a working brain machine." The point is the working brain machine. If you're using an AI to generate answers you're not improving the brain machine. I'll just keep asking you harder and harder stuff until you fail. Most of the work I have students do in my class is to produce original machine learning research about data from their lives (things like "can I build a classifier to tell me and my brother apart?") If they're successful there, I can just keep upgrading the difficulty until they're publishing original research (which isn't common, but it's cool when it happens). 

Generative AI can help set up some of the code, but it doesn't know if you have enough images to tell you apart from your brother. It doesn't know what features of the two faces are the most distinguishing, or the most similar. There are no general answers to those questions. They depend on you and your brother. That's ultimately where I'm trying to get students. I want them to actually be able to use the topic I teach to do something meaningful within the scope of their own lived experience. 

‍

Can you share an example of a Minerva student using AI in a way that genuinely surprised you? What made the work intellectually substantive rather than simply faster or more polished?

This is a funny question and makes good use of something I mentioned in a previous one. If by AI you mean "digital media generation" I have never had a student use it to do anything surprising or interesting. 

If by AI you mean anything within the scope of machine learning, then all kinds of interesting projects. I'll give a few quick ones. I had a student use satellite images and unsupervised learning to generate permanent addresses for houses in slum settlements. I've had a student create a chat bot for preparing documents for tenants in landlord disputes. I've had a student use audio recorded on microphones in the jungles in Congo to identify bird songs and wind. I've had a student create a phone app you can hold up to graffiti and it will animate it. There's no shortage of cool stuff you can do with AI if you have a rich inner life and cool ideas.

‍

In a world where AI can write, code, and analyze information, what should students come to a university to learn that a machine cannot teach them?

I think I mostly covered this in my "little brain machines" answer above, but I'd add one thing: ethics and a set of internal values. The original purpose of education was to give folks the breadth of experience to understand and construct what they considered a "good life." One of the tragedies of higher education is that we tend to have really smart students who are quick on the uptake.

They've figured out that "a good life" is a thing you're supposed to get from education and so they come to university super ready to articulate what a good life is. But most of what they articulate are things like "help other people", "save the environment", and "combat inequality." These are nice sentiments, but they're too abstract to actually apply to anyone's life. They're AI answers to the question "what makes a good life?" Reasonable, but functionally impossible to put into practice. Real life meanings look a lot more like "do I want to keep dating this person?" "Is it ok to work at job where I don't like this boss?"

Those things are much harder to answer because they're specific. That person might be right for someone else to date, but not you. Other people might enjoy working with that boss. University is a good time to learn about your own character and values.  

‍

Minerva asks students to approach complex problems across disciplines, cultures, and contexts. How does that model prepare them to shape the future of AI versus simply adapt to it?

I think Minerva students run the risk of conflating the generality of expertise for depth. It's good to have an overview of the basics of a lot of different fields. Real world problems are complex and involve a lot of different disciplines. You'll always need broad knowledge to understand what's really going on and communicate with everyone on the project. 

But It's equally important to remember that you're still actually pretty bad at everything! Specialists will have knowledge about a topic that is simply more sophisticated than what you know. In some disciplines, it's not a huge advantage, but in others they can be many orders of magnitude better at what you're trying to do than you are. And you won't know whether this is a small difference or a big difference. 

So I think most Minerva students will need to keep learning. Being an expert at something takes a long time--about ten years of full time effort. That seems like a long time when you're starting out, but over your lifetime, you'll become an expert in about a half-dozen things. One of the things that's nice about Minerva is that it tends to set you up well for that continued program of study. It's just important to know it's a first step and not a last step.

--

How should universities prepare students for an AI-shaped future? Explore how Minerva’s active learning model helps students think critically and apply their knowledge to meaningful problems.