Episode 51: Using Interdisciplinarity to Tackle Audio Deepfakes

2:24PM Dec 20, 2023

Speakers:

Dr. Ian Anson

Jean Kim

Kiffy Nwosu

Chloe Evered

Keywords:

umbc

audio

ai

work

chloe

linguistics

voice

social sciences

replicate

listening

project

algorithm

interdisciplinary team

human

research

researcher

dialects

team

technology

give

Hello and welcome to Retrieving the Social Sciences, a production of the Center for Social Science Scholarship. I'm your host, Ian Anson, Associate Professor of Political Science here at UMBC. On today's show, as always, we'll be hearing from UMBC faculty, students, visiting speakers, and community partners about the social science research they've been performing in recent times. Qualitative, quantitative, applied, empirical, normative. On Retrieving the Social Sciences, we bring the best of UMBC's social science community to you.

Like most Americans, I commute to work by driving a car. In fact, on average, Americans spend almost 30 minutes each way getting to their place of work, mostly using cars, but sometimes propelled through their cities and towns by buses, Metro trains, ebikes, scooters and everything else in between. Regardless of the method, the amount of time that we spend getting to work has fluctuated significantly across the past few years because of COVID. From 2019 to 2021, the American Community Survey showed average commutes dropping in duration by almost two minutes per trip, due in large part to clearer roads from a growing work-from-home workforce. But by 2022, average commute times had already started to creep back up by about 48 seconds per trip year over year, reflecting the longer term trend of longer and longer commutes for American workers. And given this trend, I ask you what better thing to do with all this extra commute time than to listen to a podcast. You know, Americans whose time is often occupied by commuting, household chores, and low-intensity tasks at work are increasingly living in a world stimulated by audio. According to the Pew Research Center, while terrestrial radio has seen marginal declines in listenership over the past decade, now almost 1/3 of all Americans report listening to podcasts at least weekly. We consume a ton of audio-only information about everything from true crime to politics to celebrities, even to the social sciences. But what if our ears deceive us?

That's a question that has been taken up recently by a couple of enterprising undergraduate researchers as part of a National Science Foundation-funded project on audio deepfakes. Deepfakes are audio recordings that sound very convincingly like they were recorded by someone else. Their implications for our society are enormous, especially as we continue to rely on audio-only media just like the one you're listening to right now, for information for entertainment and even for the transfer of personal information. And so as audio deepfakes improve, we're finding ourselves in a really big pickle. But thankfully, interdisciplinary science is here to help assuage our worst fears. A National Science Foundation-funded project that spans several campuses, including UMBC, has recently brought together sociolinguists and computer scientists to help understand audio deepfakes and potentially safeguard against their effects. Today, I'm delighted to bring you a conversation featuring two highly impressive researchers who worked on this project. Kiffy Nwosu is an undergraduate computer science student from Maryland, who's worked as a researcher at UMBC all the way since high school, and is now a student at the Rochester Institute of Technology. Chloe Evered, originally of Houston, Texas, is a recent graduate of the Georgetown University Department of Linguistics with a minor in Chinese. Chloe is now pursuing a master's degree in linguistics also at Georgetown, and has recently published content related to deep fakes in our very own UMBC Review. I'm so excited to bring you this fascinating, worrying, and hopeful conversation about audio deepfakes and the promise of interdisciplinarity right now.

All right. Today I have the distinct pleasure of welcoming two wonderful guests to Retrieving the Social Sciences. I want to say, I'm really excited about this conversation. I think that this is very topical, and all of our listeners are going to find this to be of great importance to their ongoing daily lives. So first want to introduce Kiffy Nwosu. Thank you so much for being on the podcast, Kiffy.

Thank you, Ian, for having me. Hello, everyone.

Absolutely. And then, of course, we also have Chloe Evered who's here to tell us a bit about the subject.

Yeah, it's good to be here. Thank you for having us.

Absolutely, and again, thank you so much for taking the time. So the subject of today's podcast is one that I have certainly had some questions about. I'm sure that many of our listeners have had questions about. I will say as a disclaimer, as an introduction to this subject, that to the best of my knowledge, and perhaps to the best of your all's knowledge, we are not currently the subject of audio deepfakes on this call, are we not? I don't think so because I can see your faces, so I think we're probably in the clear that neither you nor I are being audio deepfaked in the present moment. But I can imagine that our listeners are probably at least somewhat aware of this topic and probably at least a little bit concerned about this. So I want to start the conversation by asking you all, is this something that we should be worried about? And why right? What are some of the scenarios maybe in which an audio deepfake could be used to sort of mess up my ability to understand the world around me?

Yeah, I can take this. So audio deepfakes in particular, and I'm sure we've all seen deep fakes on social media, online. It's content that's AI generated, and usually intended to manipulate or deceive. But audio deepfakes in particular, is just the ability to reproduce human speech using AI. So this can be done through voice cloning, or through text to speech. It's a wide range of stuff that's out there. And one sort of scary way that this kind of technology can be used is in fraud. So often, banks could use automatic speaker recognition, as like a biometric authentication thing. So your voice kind of becomes your password. If your bank has your voice, you can use that over the phone to get into your account. But if someone's able to clone your voice, and if they know information, like your birthday, then it could be really easy to hack into it that way. So that's a potential area that we've seen in recent years has become more popular. And these, these scammers can become a lot more sophisticated in these attacks. And also, as I'm sure like we're all familiar with these deepfakes can reproduce the image of public figures and their voices and make it appear like they're doing or saying things that they haven't actually said. So yeah, like you said, this technology may not necessarily be targeting the everyday person right now. But it does become more and more sophisticated and more and more easy to access. So these are big concerns, a lot of unknowns at the present moment. And we can only expect this to become more and more prevalent in our lives.

Kiffy, I want to get your take as well. Are we in trouble here, or?

We are in pretty big trouble for so many reasons. The first would be that there is no robust, or there is no set in stone method of being able to detect these audio deepfakes. There is no software you're going to actually put in an audio and it tells you it's real or fake like. Now we have like an AI detectors, text and AI detectors are cool, but we don't have AI detectors for images and for audios or video. So with that being said, as of right now, if I know Chloe has talked about like banks and frauds and misinformation, like from public figures, right. But then if I want to bring it down to a much smaller scale, where it's like between, let's say, now your voice, somebody picks up your voice. It's a little call. It could be as little as you saying, "Hello, my name is Ian." or "Hi, this is Ian speaking, how can I help you?" There's robust technology out there right now to replicate whatever anybody wants you to say, with the tiny portion of your voice that they have. And the thing is, the technology for this is getting way better. The technology to detect it is getting worse. And we're in big trouble as the time goes on. I think that's the best way to say it, in terms of the definition of audio deepfakes and what they are. Chloe has explained it in the best possible way that they can be. And there is different types of audio deepfakes like she said. There is text to speech or speech synthesis, where the impute is text and the output is audio. So it's a very famous one is the Google text to speech, where you're typing something and then Google says it back to you. Or there's something called speechify, where you're typing something, and you can select voices of like different things, celebrities, and then that thing is being said to you. That's one. And then voice cloning is the other one I talked about earlier. This is the one that is more likely to be useful for fraud. It's where the impute is a voice and the output is a different voice. So I'm speaking, but there's some type of transmitter between me and the person listening, that changes what is, that changes the voice being said. And the other one is mimicry. Mimicry, where phase will work, so we're still trying to figure out is mimicry AI or is mimicry not, because mimicry is basically like I'm just flat out, say, I just flat out sound like someone else. I just flat out sound like a celebrity. That's an example of mimicry. But like the main two types of audio deepfakes or AI generated audios, per se, are voice cloning and speech synthesis.

Wow, this is a lot more sophisticated, a lot more complex than even I realized. And I want to also react a little bit to something that you said, which makes me a little bit alarmed as a host of a prominent social science podcast, which you're mentioning that even with a little bit of my voice, somebody might be able to create a pretty faithful reconstruction of the way that I talk. I have a lot of hours of my voice out there on the internet. So they might even be able to do a better job of recreating my voice than maybe somebody who's just got a small batch of this kind of test audio to work with. But I'm also really interested in this from like a social science perspective, not just as a subject of worry for my own personal life. Because it seems like, you know, a lot of people probably out there thinking about this from the perspective of like, you know, this, this idea that we're able to create just the impression of somebody's voice that it might sound sort of like their vocal cords are producing the content. But language is a lot more complicated than just like whether somebody has like a deep voice, or a high voice, or a gravelly voice, right? I mean, this is something that is kind of pushing the frontier here, because it's not just about the ability to modulate the register of the voice or something. It's also about somebody's linguistic patterns, too, is it not? So somebody's dialect perhaps might have a pretty strong impact on how this works. Is that Is that something that this AI is able to handle, or is it really just, you know, the text that you're writing that's kind of just replicating the words on the page, so to speak? Or is it actually able to handle and replicate our dialects as well?

Yeah, absolutely. That's a great question.. thought about a lot and, and a lot of research on. And like, if you mentioned, there's different methods, so voice cloning, if you no matter what dialect, you speak, a clone voice is gonna replicate exactly what you sound like. So there might be some things that it misses, like maybe there's some vocabulary words that you use, because of where you're from, that the AI might not know, to replicate. But yeah, there's this potential to create deepfakes in all kinds of different dialects with all kinds of different accents. And so I think we're familiar with voices that sound like Siri or Alexa that are very standard. And we can tell that they're, you know, machine generated, I think a lot of people might, right. They might not know that these AI generated methods can also replicate voices that are not what we would consider a standard. So yeah, there's a ton of stuff out there.

Wow. Yeah, Kiffy, do you have other, other insights?

I have a CS on machine learning data science background, more or less than the social sciences in the past couple months working with kuleana commands, and I've been able to learn a lot about like the accidents that come with different voices. And I can tell you that AI has it finds it very hard to not whatever you give it, it's what it's going to give you back, like the whole point of machine learning specifically, is that it's learning. So if you're feeding data with like, let's say, a South Indian accent, right, and you try to replicate someone who has like an American one English accent, it's going to keep the accent, give it back to it well, making while trying to modify it to make new ones. However, with the dataset that I've been working with, I've seen a lot of accents that are not like, AI is so good that if even with text to speech, you're able to modify and change it and play with it however you want to make an accent. So it's not just Siri, you can make it a little voice with like, an accent from like the middle of the South. Or you can make you can make it say whatever you want. It's just a matter of how well versed are you with the accent that you're trying to replicate? Because again, it's not human. So whatever you tell it is what it's going to do. As someone who's not well versed in linguistics, I will find it extremely difficult to make audios because you know, I don't know what this person sounds like. I don't know how it works. But that doesn't mean that I won't be able to do it. So an all the accents, they all, no, every accent is different. And they all play into how the audio is perceived. But it's not impossible to replicate it. It's not one of those things that can happen, it can happen. And it is not necessarily easy, but it's pretty doable.

So Chloe and Kiffy, I mean, these kinds of insights together are both giving me a lot more unease, but also maybe potentially a silver lining for the future, which is to say that interdisciplinary research, it seems to me is pretty darn powerful in this realm, and collaborating like this across the social sciences and in sort of STEM/computer science fields to understand this problem, it almost feels a little bit like the old, you know, Spider Man quote, like "with great power comes great responsibility," you know. By your powers combined, then we could really be onto something that could be effective at replicating language in a way that would be virtually undetectable even to sort of native speakers of dialects, right? You'd be able to replicate that stuff, and have people who maybe, let's say, from a specific region of Appalachia say, oh, yeah, this voice sounds just like somebody who actually lives in that specific region. If we had right the context knowledge of what those dialects sound like to be able to feed it algorithms to do this, do this replication. At the same time, and this is what I'm hoping you might be able to tell me about maybe by your powers combined, there's also some potential solutions to this problem. Is that the case or, or not? I especially am interested in your work on this NSF funded project that you've been working on. Has this collaboration led to any insights into how to potentially spot deepfakes, how to potentially overcome some of these challenges?

Yeah, so I have said earlier I do the computer science or the data science and machine learning expert, and then Chloe and Dr. Melanson and the rest of her team, they do the linguistics part. And I just want to give like just a very dumbed down version or a very simple version of how both of our powers come together. So What COVID and her team does is that they annotate audio. So they spend hours. That's where I give them that is something that most of us on the data science and will probably not do. They spend hours listening to clips, we had to listen to over 1000 clips, where they listen to those clips, and they annotated it. There's these things called elf expert, the fact that you see features, and they are basically features that are in spoken English language. And they're what we're using as the baseline of detecting audio defects. And the fight ELS we have pitch, pause, breadth, first and sound quality. There are five of them, I want you to remember that. And while they're listening to these audios, you're listening and you're annotating to see if any of these factors are present. And not only are they checking if they're present, they're checking, is it anomalous? Is it normal, how long it is, we have presence or absence of births, we have a novelist pauses, normal pauses, and these are all things I've learned in like the past three months from the linguistic experts. And then what we do on the computer science, and is that we now we have this data, we have audio clips, I'm like, Okay, in this audio, there is pitch, pause this and this and this. Now we work to find so and so ELS on the audio. So we write the code, we do all the machine learning. And we are currently following unsupervised approach, because we are still trying to narrow it down to what exactly the EDL is, and the best possible algorithms. So when I write my code online, when I'm working on my congregants, I get timestamps of when there's possible CDLs. And then I map that to what Chloe has given me to see if, oh, I found something here, that Chloe find something there, or Oh, I didn't find something here. Why is why did Chloe see something? And why can I found it's how can I fix my code now to or fix my algorithm to be able to pick up the things that we have found. So with that, being together an interdisciplinary team, it makes it way easier, because you're getting the best of both worlds per se, because I am, like I said, we have no very minimum knowledge on linguistics. So if I didn't know that there ADLs, if I didn't know that there, there is a pitch or burst, or I learned a new word called timbre, which is also a way voice could sound if I didn't know that all these things existed, there was no way I'm going to be able to find them. And now that I know that they are there, I'm going to be able to find them. And in terms of our solutions, like for the long term, like I said earlier, we are looking at elf in real audio, we're looking at your lips, and fake audio. And there's a specific presence or absence of ADLs, that helped us determine if an audio would be possibly real or fake. So as of right now, we're still selling like the preliminary phases of like trying to figure out what types of DDoS we're seeing and how to detect them. And then the next step from here, and this from my part of the project is to see if there's a correlation with the types of elf being found. And the type of audio defect there is that for example, if you listen to an audio, and it had like, no pauses, and it's just talk, talk, talk, talk talk, you get a little concerned or like why there isn't any pause, tells me this audio is a little fishy and stuff like that. So the end goal would be to have a very robust algorithm that is able to sift through all this audience that are amazing experts have annotated for and then with these audios you were able to find the right either left or the right lack of elf, were able to find out this audience possibly went through the fake. Well, all this is from a CS or a missionary perspective. I think I'm gonna hand it over to Chloe to give us like from her linguistic background, what it feels like to be in an interdisciplinary team.

Absolutely. Yeah. Thank you so much for all that detail. That's super fascinating. Yeah, Chloe, tell us about your perspective on the projects, please.

Yeah, absolutely. Well, our approach to this problem is kind of novel in a couple of ways. And one is that it's interdisciplinary with the social sciences and computer science, data science. And another way in which it's different from existing approaches is that we're centering human listeners and human perceptual ability. So like, if he kind of explained the other linguists on the team, like me, have backgrounds in sociolinguistics. And specifically variation is socio linguistics, which I know that's a lot of big words. But it's basically just the study of, in what ways human language can vary, because it does vary a lot. But it varies in pattern systematic ways, and in ways that are impacted by social contexts and social factors. So what we're doing to approach this problem is taking his knowledge of language variation, specifically in spoken English, at least for now. And then we're listening to AI generated speech finding linguistic features that are easily discernible to just the average human listener. And that also distinguishes fake speech from real speech. So another component of this project outside of the algorithmic, robust algorithmic detection we're trying to achieve is a training program that we're developing so we're working on a curriculum, and actually we have given some trainings to some UMBC classes. Teach to people how to listen to these EDL F's are expert to find linguistic features that we've, we've created or developed, we've sort of created the right words that we've heard in the clips. So our training program is to equip just the everyday person with tools to listen for cues that an audio might be fake, because I know we've all seen, like image effects, I'm sure. And you get little cues, like, I've heard one that's like, look at the hands. generators, they're really bad. Like a fingers. Exactly. Yeah. But there's not resources like that as much for audio like I don't really before this project, I didn't know like, what should I listen for? If I think an audio is fake? So that's the kind of thing we're also trying to develop from a social scientist perspective to fight this problem?

Yeah, that's fantastic. I mean, I think it speaks also to just the power of socio linguistics in helping to address this problem, because language is socially embedded. And we as humans, who use language to communicate, but we aren't even doing right now, there's so much about tone, we know so much about all these different sort of ADLs, as you're talking about, that we use as a subtext to understand the intention of speech. And so to think about the human ear, basically, at our brain power, as this incredibly powerful tool, is a really interesting lottery. Because, you know, we often think about computers as being at this stage and in the world far more potent, essentially, as processors that we are, I mean, we can just look at the chess chess, examples of AI versus humans, you know, AI, now beats humans in chess, you know, all the time and go other kinds of board games, that kind of thing. Certainly that that, that's been an ongoing sort of realization is that for a lot of complex tasks, computers are much better than we are at this stuff. That language is really an interesting frontier in this in this world, just because language is so socially determined, it's so socially coded that humans actually are much better trained, perhaps today than computers at this kind of detection. So that's super interesting. And I'm so glad to hear that this collaboration has been so fruitful, and especially excited to hear that you've been able to apply some of this back to UMBC. Right, to be able to bring this back to the to the campus. Because this project is, as we've seen, right has kind of spilled past the disciplinary and institutional bounds of UMBC as you all continue through your your work. Yeah, Kiffy. I think Kiffy wanted to chime in.

I just wanted to also bring up the fact that no matter how hard we try, like, and this is from a machine learning data science background, AI is never, never say never, but I don't think it's gonna be as perfect in detecting audience because everybody with between just the three of us, we all sound very different, very different accents, and put that on a larger scale where there's 7 billion people. So that is why our training specifically comes in very handy. Because, yes, we have AI. But again, you're not going to hear audio difficulty, like hold on, let me pull up that algorithm real quick and see if it's real or fake. That's where the listening comes in. And we call it listening to learn. That's one of our papers. It's I don't, it's one of my favorite titles. But the point being that, in terms of it being an interdisciplinary team, I really don't think we'll be as successful as we are without the other, at least from my end, I don't think I'd have been able to do all of this without the help of the rest of the team. And I also think that from a socio linguistics, and I'd like to say that it also be the same. And we're also targeting different audiences. Like even though we're together, we also target different audiences in the sense that my algorithms are still targeting people that actually know how to use these articles, and actually, could try to implement the algorithms I'm working on. But with the social linguistics, they're targeting, like the everyday human like, we you're sitting and you're listening, and they're teaching, you're teaching you Oh, this is what you look out for. If it sounds like this, it's probably best if it's this is probably this. And when these two words come together, were reaching twice the people I would probably reach and in the long term, maybe outside of UMBC, outside of maybe within our different colleges audience keeps getting bigger and bigger, because we're interdisciplinary teams, like I mean, social linguistics, and computer science. Those are very two different things coming together, submit a very killer project, and I'm really good on where we're at now.

Yeah, I'm excited too. I really hope that we can follow up with you in the in the years and months to come and see how this project has continued to develop and grow. And also, I want to think a little bit about your own research trajectories as well, because, you know, you're obviously students that have started somewhere in your research journey, and then gone on to this really awesome project and beyond. And, you know, I wanted to kind of combine two questions, if you wouldn't mind just giving a little bit of background about sort of how you came into this project. Tell me a bit about it. sort of your trajectory from your first foray into research how you got involved in this in the first place to where you are now. And in doing so maybe if you could put an eye towards telling us a bit about how you might give advice to students who are hoping to replicate this pathway.

It's an interesting way that I came to this project. So actually, last spring, Dr. Mallinson, who's one of the principal investigators on this research team, reached out to one of her colleagues at Georgetown, in linguistics, looking for a student to join the project who's interested in all these intersections of AI and society and language and also, you know, phonetic variation. And it just so happened that I had taken a class with that professor the same year, and she recommended me for the RU position, and the rest is history. So it was just felt very lucky that I came into this project. And I've really enjoyed it, like both Kippy. And I have stayed on the project since last summer. And I at least will stay on as long as they'll have me. But in terms of advice I would give to students looking to do something like this, especially if you're an undergrad, maybe just starting out in your research journey. Just from my own experience, it's really valuable to get to know your professors and going to office hours. And being involved with things in your department. Like here at Georgetown, the linguistics department holds a little speaker series almost every week. So they have scholars from all over come over give a talk. And a lot of the times the research was way above my head, especially like as a first year, a sophomore, but I still showed up. And I think professors like recognized me and it was just a small way to start building those relationships. So I know this kind of thing is a little scary. If you don't feel comfortable going to office hours, you don't feel like you have anything to say, I would really encourage you to try and do what you can to get over that hump and go build relationships with people, because at least in my experience, my research trajectory would not have been the same at all if I didn't have those connections. So yeah, getting to grad students as well, I would definitely recommend that too, if you have grad students in your department, because they're great resources.

That's fantastic advice. In part, I would say it's fantastic, because it's the advice that I like to give to students as well, to just show up, right. And I will say as part of my response to that, that if there's any student listeners out there who are listening to this episode, take a look at the social sciences at UMBC. And the lecture series that we have going on, you can look at the calendar of events at the CS3 website, and make sure that you come out and attend some of these and you never know what might happen, you might make a connection as Chloe was mentioning. So it's true that some of the stuff might go over your head, but there might be a lot actually that really galvanizes your interest. So definitely great advice. Thanks so much for that. And it's great to hear a bit about how you got connected to the project, Chloe, and how you continue to work on it in the future. Kiffy,what kind of advice, and tell us a bit about your journey?

I started interning at UMBC since I was a sophomore in high school (Dr. Anson: Oh, wow). So I started in a different department. I actually started with Dr. Childs, I started in the College of Engineering as an intern. Then by my senior year of high school, I joined this class it was called independent research and where you have to intern and you know, write a paper about your research. So I had already been working with Dr. John Asia and Sarah and the rest of the grad team for like maybe almost a year and a half by my senior year. So I was just like, Oh, hey, I might as well just keep working. Like I just I went in as a high schooler it was like my first ever experience and I just never. I have I had never I just continued right. So by and I'm also want to preface it by saying I'm a first year I'm a my first year of college. So by my time I graduated of the teenagers like oh, hey, you can't you're done, though you're not a high school intern. You cannot be like, are you so an undergraduate researcher. So I just continued the program as an undergraduate researcher. And honestly, I like folks that I'm probably going to be here for as long as they ask me, I, I've learned so much when I joined this program, I barely knew how to code like I, I was, I was struggling to see the police. And not to say that I'm still not struggling right now. But like, I had like, bare minimum knowledge. And they believed in me and they said, Oh, hey, we see you. We see where you're going. I think you're gonna be a great part of our team. And I just, I've been working with them for advice, I'll get good. Honestly, I'd say reach out to people. Because I sent an email to the oxygen agent, I, I might have spammed her inbox. I sent maybe 123 emails, I always followed up with her. I was like, Oh, hey, I really, really appreciate I really like it. I want to work with you and your team. So yes, always reach out. And you because as a first year student, most of us don't have experience and most people are held back by the fact that Oh, I don't have experience I wouldn't be able to do it. I honestly think that they don't. They don't expect you to have experience. So if I'm advising someone at work, don't be afraid. Don't let the fear of not having experience be the VC holding you back. Because at the end of the day, we're all learning Even as a team, as an eager team, we all learn from each other every single day. So, reach out to people get involved, but while getting involved, do it because it's important. And then my other advice is like, don't let anything hold you back per se, like, You're not expected to have users like the work I'm doing. They didn't expect me to learn, I had to learn it over the summer, and I'm gonna keep learning. But by the end of next summer, as I'm working with a team, I'm gonna have learned so much more. And it's, it's a learning curve at the end of the day, like, that's the whole point of researching every researcher that here is a learner. And if you want to go into research, get ready to learn.

"Every researcher is a learner." I love that line. Another thing I love about your response, Kiffy, is that it really is sort of the backstory of this is to try to defeat your imposter syndrome as much as possible, right? Because I think every researcher, and Chloe, maybe this resonates with you as well, every researcher comes to that journey with a lot of trepidation because they don't feel like they know enough to be able to do research. But in a project like this one, I think it's really clear that you're going to eventually kind of get over that learning curve. And you're going to be able to start making really impactful contributions. And how better to do that than to collaborate across disciplines. Collaborative even across institutions. This is a really great project and Chloe Evered, Kiffy Nwosu, I want to thank you again, so much for coming on the podcast talking us, talking to us today about this project about audio deepfakes and the dangers therein, but also potentially the pathways to better understanding them through a collaboration across the social sciences and computer science. Thank you both so much for your time. And yeah, thanks again for teaching us so much about the topic.

Thank you so much.

Yeah, thank you also so much for having us. Again, I just learned so much from this team. And I think my big takeaway was doing interdisciplinary work is that it's scary because you have to trust other people so much. Like I probably only understand 20% of what Kiffy's doing at any given time. She has this technical expertise I just don't have. But that ability to trust each other and see value in each other's work, even when we don't fully understand it, I think has just been so exciting to me. We're able to expand what we can accomplish by just one researcher, just one discipline, when we work together. So thank you for having us. It's been so great to talk about this with you.

Now it's time for Campus Connections, the part of the podcast where we connect today's featured content to other work happening at UMBC. Our production assistant, Jean, has been learning a lot about audio editing and podcast production over the last few months. But Jean, what else can you tell us about this subject?

Hi, Dr. Anson. For this Campus Connection, we'll be looking at the work of Dr. Jennifer Maher, an associate professor of English here at UMBC. Dr. Maher teaches in the department's Communication and Technology track and as an affiliate faculty member in the Language, Literacy, and Culture PhD program. Her research focuses on rhetoric and technology and her published article entitled "Good AI Computing Well" discusses the topic of AI and rhetoric, further emphasizing the need for a more nuanced understanding of AI's role in society, including ethical considerations and accountability. The text explores the rhetorical education of AI, emphasizing proficiency and natural language as AI'squote, "achilles heel." The article further explains that natural language is a symbol system and therefore a rhetorical and ethical endeavor. And as such, if AI is to become proficient in natural language, AI must acquire an ethical disposition on top of understanding syntactic validity and of linguistic cues. The article notes instances of AI systems who have learned and propagated ideas of biases, racism, and sexism, further emphasizing the point that as AI ultimately has to make decisions and explain the decision making process, the future of AI is not only computational, but deeply intertwined with an ethical and rhetorical understanding of human affairs, which sounds a bit similar to the conversation we just heard. Dr. Maher further argues that all stakeholders involved in AI development, or anyone affected by AI, must gain this understanding of the inherent rhetorical nature of AI for future accountability. And that's all for this Campus Connection.

Thanks so much, Jean, for connecting us to this great content. And thank you for tuning in today to listen to our conversation. As always, as you ponder whether this is my real voice or something that's been churned out by an algorithm, I encourage you to always keep questioning.

Retrieving The Social Sciences is a production of the UMBC Center for Social Science Scholarship. Our director is Dr. Christine Mallinson, our Associate Director is Dr. Felipe Filomeno, and our undergraduate production assistant is Jean Kim. Our theme music was composed and recorded by D'Juan Moreland. Find out more about CS3 at socialscience@umbc.edu and make sure to follow us on Twitter, Facebook, Instagram, and YouTube, where you can find full video recordings of recent CS3 sponsored events. Until next time, keep questioning.