Professor of Comparative Literature and Humanities Katherine Elkins is the director of the integrated program in humane studies (IPHS). She was recently interviewed by Forbes about her innovative use of AI in a liberal arts curriculum. Kenyon also received an award from Schmidt Sciences, which will fund Elkins and Visiting Instructor of Humanities Jon Chun’s project digitizing archives using AI. The Collegian sat down with Elkins to talk about her, the job market and the future of AI in the liberal arts.

The following interview has been edited for length and clarity.

Tadhg Sahutske (TS): Tell us a bit about yourself. Your background, your work, your education and something that people might not know.

Elkins: Well, I guess one detail about me is that I originally thought that I would join the Foreign Service, work for the State Department abroad, and I’ve lived abroad after I graduated from college, in Paris and Prague and then in Asia. And it’s just a little bit of chance that when I came back to the United States, I decided to go to graduate school and get a Ph.D. I had interviewed for the Foreign Service and passed the exam, but they didn’t have enough openings for that year, and so I also applied for Ph.D. programs. I ended up in Berkeley, California, to get my Ph.D. So that’s how I ended up here.

TS: You have a background in comparative literature, right?

Elkins: I do. So comparative literature is very interdisciplinary. I like to say that I’ve never really had a discipline, because, even as an undergraduate, I was a literature major, which, at Yale, was extremely interdisciplinary, even more interdisciplinary than comparative literature, English or modern languages. So in comparative literature and literature, we like to work in everybody else’s backyard. Interdisciplinary can be very surface. I like to think we do it very well, and that requires learning everybody else’s language. So you know, one thing that I teach here in IPHS, and that I’ve always taken with me, is the importance of learning how many different fields can be looking roughly at the same question, but be asking different things about it and using different kinds of language. And that is definitely the case when we look at computer science and computational humanities, that in order to work on AI, as I do now, it’s very important to understand the kinds of language and concerns that people are bringing from their different fields and speak those different languages.

TS: So, when you’re talking about AI, do you mean large language models (LLMs) or something else that I’m not aware of?

Elkins: That is a great question, because AI has been around for a long time. There are many different kinds of artificial intelligence. There are a lot of people who call themselves AI experts these days, and many times their training is actually in very different kinds of AI from LLMs. One of the really exciting things that we did here at Kenyon is we started working with ChatGPT before it was ChatGPT. So in 2019, we did some of the very first experiments with GPT-2, which was an early beta version only released to researchers, a select number, and we had access here because we had already been working with AI. So early on, we were actually experimenting with GPT-2. It was much harder to use than ChatGPT, and so we trained it to write like different kinds of writers. We trained it to write Sex and the City episodes. We trained it to write like Chekov. We benchmarked and wanted to ask, how well can it write like these different writers and in these different genres? So we started working with generative AI in the very first days and have been there all along. But in the AI curriculum we have here, we also teach the much older forms of artificial intelligence, and some of those are still better at solving certain kinds of problems. There’s also this larger field of machine learning. It’s been around for much longer, and that is called machine learning, because instead of programming this kind of earlier, older form of AI with rules. We just give it lots of data, and it infers the rules from it. Some of that is based on just traditional statistics. So it looks a lot more like what you would find in a math department. So that’s a much older form of AI. And then there are also rules-based AI systems that manipulate symbols or logic or follow rules. So you know, AI has been around for a very long time, but GPT-2 and this really new linguistic approach of large language models that we have now seen in Gen AI, that was a huge breakthrough. And it was really exciting here at Kenyon to be part of that early history.

TS: There’s been a lot of advancement in LLMs specifically. How do you keep up with all of that from day to day?

Elkins: Saying in the field that most of us say, I’ve never worked so hard to feel so behind. And I have talked with fellow researchers who describe sleeping at work, under their desk and still being scooped on their research. And you can work on something for three months or six months and then have somebody else get there first. So it is a crazy pace. It is very exciting, but it is very hard, even for those of us in the field to keep up with.

TS: In your IPHS classes, the course descriptions mention using AI productively and critically. What does that look like in an assignment in an IPHS class? Can you talk us through something like that?

Elkins: Absolutely. There are two separate approaches. One is leveraging AI for humanities and social science research. So, in programming humanity, we start with the intellectual framework, beginning with the basic question of ‘What is data?’ and moving all the way to AI. So part of it is providing the entire intellectual framework to understand it, but we also leverage AI for asking questions that are important to us in the humanities and social sciences. So instead of using AI for humanities and social science research, we’re actually researching AI. So early on, we were researching GPT-2, we were looking at bias in these models. Now we look at emotional intelligence, persuasion, safety and things like that. So that is bringing a critical perspective to these models and to AI, but also a deep understanding of how they work, because a lot of times, critique is not that engaged. It’s a critique from afar, but it’s not a critique that is truly based on understanding how the models work. So it’s AI leveraged for humanities and social science research. It’s also bringing that humanistic critique to an investigation of how AI is working.

TS: I’m sure, working with AI and the humanities, the question kind of naturally comes up, about the distinct humanity of human work over AI-generated work. And do you feel like your research has brought you to a better understanding of that separation, the distinctions?

Elkins: People often ask me what will be left for humans, and I like to encourage us to think about what should be left for humans, because if we remain confident that there are things that an AI will never be able to do, if we’re wrong, we have not prepared for our future well. So I am very concerned with actually testing AI for these things that we often think of as distinctively human: creativity, emotional intelligence, persuasion, not because I’m excited about taking over, but because I think we need to be ready for if and when it actually can do these things that we have thought that are traditionally human tasks, so we are essentially automating intelligence and creativity. What does that mean for humans? Well, I would like us to start thinking about that now. I’m thinking about our future, what we want to be left for humans, and not just what will be left. What happens if an AI system is actually better than a human at some things? What do we think about that? What will we decide? I was giving a talk at Weill Cornell Medical School in Qatar in October, and with me was Adam Rodman from Harvard Medical School. In his research, he has found that AIs sometimes diagnose better without a human than with one. And if that is the case, if that becomes the case in many things, will we opt to have the AI do it if it does it better without the human in the loop? Traditionally, we’ve always said we need a human in the loop everywhere. What happens if that human in the loop makes it worse? Do we still insist on having the human in the loop? That’s a very important question that our humanists need to answer. So I think that humanists and social scientists have a lot to say, but in my opinion, we need to stay current with the questions that we’re just beginning to think about that will be here tomorrow.

TS: On that same line of thinking. Kenyon calls itself the “writers’ college,” and many of the students here want to work in creative or research fields. They might feel that new, globally accessible technologies like LLMs are going to put them out of work and hurt their future. How would you respond to somebody saying that IPHS as a whole is harmful to the students here?

Elkins: I would say to all of those writers that they are right to be concerned and that I am with them on their concerns. You know, we have had writer strikes in Hollywood. I believe those strikes are entirely justified. I think that writers should be concerned, and it’s not just the creative writing; it’s the kinds of writing that writers have traditionally done to support their creative writing, the more technical writing, the everyday copy editing, those kinds of jobs are disappearing. I regularly hear from writers who have lost work because of AI. The problem is, we can’t stop AI development. That’s not likely to happen, even if we all resist. And so what I have been doing is trying to be part of the conversation about where we go next. And it is actually one reason I’ve participated in conversations with [the United Nations Educational, Scientific and Cultural Organization] about the future of culture. On the one hand, AI can really benefit us in preserving our cultural heritage, and that’s some of the work that I’m doing with human science. But on the other hand, the worst outcome would be if we were the last generation to create new cultural heritage. And this is not just writers, it’s artists of all kinds. This is something we really need to think about. But I would also like us to think about how much of this is an economic issue, because if we could actually support our artists economically, they could still write, right? We don’t have a good social safety net, or our jobs are typically our sole access to healthcare. So this is a much larger question about how we support artists going forward, with a very strong economic component. And it is not just artists. There are many, many other kinds of jobs that I think we need to think about and be concerned about. There’s a whole new field of work that concerns AI’s impact on jobs, and now, actually, computer scientists are at risk. So we are all in this together, and writers are right to be concerned, and artists are right to be concerned. But I also would like us to think about this kind of larger human endeavor. How important is work? Do we need to work to be happy? How will we support people if we no longer have work for everyone, and if each group is just advocating for themselves, whether it’s the writers or the doctors or the coders, then we’re not all together, collaborating on solutions.

TS: Kenyon recently received a $330K gift to support work supporting small archives. Can you tell us about that? What is your involvement?

Elkins: Jon Chun here at IPHS and I are principal investigators on this project, and this is the kind of exciting work that AI can help us with, and that is allowing these small archives that don’t have the money or the human labor to actually digitize everything that they have to use AI to augment that digitization process So that will help save and preserve these crumbling archives that are disappearing. It will help communities have access to their cultural history. We are also trying to build in intelligence into these systems so that people can easily make new discoveries and connections in these archives. We’re focused on New Orleans, because the New Orleans community is multilingual, so that is Cajun and Creole, which are languages that are not very prominent in AI training. So when we can bring these kinds of resources into our AI systems, they have a deeper understanding of history. They have more community understanding of these communities that speak other languages, but ideally, we’re trying to create something that archives all over the world can use to preserve their crumbling resources and our cultural heritage. So I’m very excited about it, very grateful, because there are very few given in the world. And we could not be more excited to get started on that. 

TS: You said that AI is not a tool; if not, then what is it?

Elkins: Well, a tool. You know, if we think of a tool like a hammer, we can predict what it’s going to do. And therefore we can always know if I use it in this way, it’s going to do this. If I hold it at this angle, if I use this amount of force, traditional back, that older kind of AI that I was talking about, some of those systems we programmed with rules. If somebody asks you this, this is what you answer. They’re programmed. Those are tools. We know exactly how they’re going to perform. Problem with the current AI systems is that, A, we don’t fully understand how they work. B, we can’t fully predict how they will behave, and so, C, we’re not fully in control of using them as a tool.

TS: We don’t understand how AI systems work?

Elkins: We understand elements of them, but we don’t fully understand. They’re so large that we understand pieces and algorithms, but we don’t understand the whole system, so they still exhibit unpredictable behavior. It’s why I spend a lot of my time working on AI safety, and as a member of the U.S. AI Safety Institute (AISI). 

TS: What does AI safety exactly mean?

Elkins: So we’ve just changed the name because there’s been a new executive order, but until recently, it was called the AISI, looking at questions of AI safety. The current administration has shifted its focus to innovation and is moving full steam ahead with AI development, mainly because many countries are working on this. And so there is a global situation in terms of who will develop the most powerful AI first, but we still are working together with the U.S. government to make sure that we’re thinking through safe deployment, safety evaluation and all of these kinds of things. So a lot of our time is actually spent working on AI safety evaluations.

TS: So that would mean determining whether it should be used in a certain industry or something, right?

Elkins: Well, determining rules, some guidelines for where we need more rules. I don’t know that we will have regulation there. We’ve had students work on projects on regulation, and a great project shows that across the aisle, there tends to be a fair amount of agreement on regulation, but I’m not sure it will actually happen. So it’s all the more imperative that many of us investigate these models to see how they behave in different kinds of situations. You could imagine putting one in a drone, a weapon, which people are already doing. And so we’re really trying to test the behavior of it. So you can see with these kinds of questions, do we want humanities and social science students trained and sitting at the table to ask and answer questions about AI, safety, ethics, future of work and I hope we would agree that Kenyon students should be involved in these conversations.