Show notes
BRENDAN LYNCH, HOST: We live in a time where nothing is true. An era where reality and hoax look the same on the internet. Whoa, wait a second. There are people who actually know what they're talking about dangerous people. We call them experts. We're giving these experts a megaphone to drop some truth bombs. If you can handle the truth. I'm Brendan Lynch, and I'm the host of "When Experts Attack!." For Kathryn Conrad, artificial intelligence is the elephant in the classroom, one that can no longer be safely ignored. It's better, she believes, to try to establish some parameters for its use. That's why just before the school year started, the University of Kansas English professor published a blueprint for an AI Bill of Rights in education in the new scholarly journal Critical AI. For Conrad it's an attempt to establish guideposts in a wild and wooly frontier. While her scholarly expertise in modernism lately centers on 19th century Irish writers like Oscar Wilde, she tells "When Experts Attack!" correspondent Rick Hellman her AI proposal comes from months of study and discussion of the challenges posed by the disruptive technology. THEME MUSIC RICK HELLMAN, CORRESPONDANT: You noted in your article we haven't even reached the one-year anniversary of the day that a private company called Open AI introduced ChatGPT and thereby kicked off a national debate about so-called generative AI. And we were just discussing that we feel like futurists have already raced ahead to imagine artificial intelligence as sentient robots and things which they have already in sci fi movies. And yet, it's real life. It's creeping into every facet of our life today, and yet a lot of people don't know about it, what it is, what it means, what it implies. Let's talk about some of the major terms and entities that are out there in the field of artificial intelligence, and hopefully people can see their relevance to the field of education, where your expertise lies. For instance, what is a large language model, like ChatGPT? Is it a chat bot? What can it do and what can't it do? KATHRYN CONRAD: That's a great question. I love that you started with the robots, because of course, that's our first thought when we hear AI. There's H.A.L. and there's Skynet. We've seen a lot of alarmist language around AI, especially since last year. I think it's important to note there's debate about whether what we have right now — things like chat GPT, large language models and other kinds of generative AI — actually have anything to do with those fantasies of a sentient robot future. I would say that open AI — and I can say because I've read their read their self-descriptions in more detail than Anthropic and Google's Bard and some of the other models — they are hoping for something like AGI. That's artificial general intelligence. It's not clear what the relationship between that dreamed-up future and what we have right now is. You asked what is ChatGPT, what are large language models. They are text-generating models. Generative AI includes other kinds of generation models, but just for your question, yeah, it is a chat bot. That's important to know. And I'll try to remember to come back to that issue. What they are are text-generating models that use what's called a transformer architecture. That's a kind of processing algorithm that considers the context and weighs the elements in the sequence of tokens or words in order to create a kind of natural-sounding answer. Chat bots have actually been around for a long time. I was teaching with some chat bots back in 2015 just to talk about how we interact with technology. Since 2017, there was a real shift. This transformer architecture really made the difference between something that was kind of quirky and something that's much more natural and the kind of impressive outputs that we get with something like ChatGPT or Bard or Claude. Because chat bots are an interface, that's how we interact with the architecture. If you've ever messed around with ChatGPT, experimented with it, you'll see that it often refers to itself as "I." It talks about understanding. Sometimes it will use language that makes it sound like an intelligence. That's a deliberate choice. It's a deliberate choice to have us engage with something that seems like a personality. They didn't have to use that kind of model. Chat bots are an interesting choice for companies that have artificial general intelligence as their sort of arc, what they hope to get. If you want to pretend that you're interacting with Jarvis or with Data from "Star Trek," you can because it uses an AI. But there are other ways you can use artificial intelligence that wouldn't do that. It's a deliberate choice to create something that's known as the Eliza effect. That's when you interact with a computer and you just sort of assume an intelligence you interact with it as if it were human. HELLMAN: But they want us to think that it's a sci fi robot. CONRAD:I would just say, if you go to image-generating models, if you were to enter artificial intelligence into something like Midjourney or Dall E-2, it's pretty likely you're going to get a human face, usually female, with some wires coming out of its head. That's definitely part of the image. It's part of the cultural imagination that they were trained on. They're trained on large datasets. That's a really important other part that we'll probably get to. They're trained on large datasets, and then they're tweaked. They're trained and tweaked by human data workers. There's definitely humans behind the veil throughout the process. That's important to remember. Even though when you're interacting with it, there's not a person right behind screen fixing things. They do fix outputs, for sure, and guardrail and reinforce and direct. HELLMAN: Interesting. Well, I think another thing that people think about when they think about artificial intelligence these days — and certainly in the field of education — is its ability to plagiarize and cheat. This gets to some of the ethical issues with AI that we'll be getting to later. You're right, there is some truth to the media's obsessive focus on plagiarism or cheating, which is the ease with which students can generate ostensibly passable work on a range of assignments. You say teachers have already been forced to adapt to this. How so? CONRAD: Pretty quickly it was clear that this would have potential impacts on what students might produce. As teachers, we were used to adapting. We adapted to the internet. We adapted to Wikipedia. We adapted to the pandemic most recently. We're used to pivots. But this was dropped in our laps, fully formed, without any consideration for its impact on education. Teachers have been figuring out whether and how to work with it, how to change policies, how to change assessments, so that we ultimately get what we want. If there are any students listening, what we're looking for when we ask you for assignments is not so much the right answer as an opportunity for you to learn. If you're giving me something that was generated by ChatGPT, then I need to reconsider that assignment. I don't want a perfect paper. I want you to learn how to write a paper. I don't want you to give me perfect code. I want you to learn how to code. This is one of the things teachers have had to deal with from K-12 through graduate school is how to create assessments that allow us to give students the kind of competencies as well as content that we want them to have. That's how we've been adapting. HELLMAN: What are you seeing in the college classroom today in regard to the use of AI, should we say chat bots, by students? CONRAD: That's a good question. I mean, I've talked to a lot of students over the last 10 months. Certainly there are some teachers that are working with it. There are some students or some teachers who've said students shouldn't use it in their classes. I think a lot of teachers haven't said we're still trying to figure out if there's ethical uses of it and how we might use them in the classroom. What I do sense — again, through students, both in and outside my own classrooms and talking to other educators — is they're sort of at sea. They don't even necessarily know whether if there's no policy, does that mean they can use it or that they can't use it? That's part of the reason I think it's really important to help people understand, to give them opportunities to think about policy and to think about principles for policy. That helps to protect students as well as teachers. HELLMAN: Well, that's why you wrote this article about a Bill of Rights for AI in education, right? Let's talk about some of the things that you raised in the article. You start first by saying AI entails a host of ethical problems, from scraping of data to amplifying stereotypes and bias to surveillance. Can you talk about some of those bad, basic root cause issues? And then we'll get on to your Bill of Rights itself. CONRAD: There are a whole lot of ethical issues. And if people are interested in sort of following conversations about it on social media, the hashtag is usually #AIethics for that. So one of the main issues is that the datasets that have been used to train these models, whether they're Visual Media Generators, or whether they're textual or whether they were trained on data that was not specifically consented to by the creators. When I say data, I'd like to remind people that data doesn't just mean your Social Security number or your medical records. It means poems. It means pictures you may have posted on Instagram or an art website that you have copyright over. Because they're publicly available, they were often scraped. When we talk about scraping, that's what we're talking about is taking that as data to train. And the implications for artists is quite profound. Several models have been shown to reproduce artwork that's very close to the original, even with signature sort of garbled in the corners. That's part of the ethical question: Is that fair use? Most artists, I would argue, at least from my research over the last year or so, are not consenting to that. They feel that they should be remunerated for that training data. So that's one of the ethical questions: Is that consent? HELLMAN: Because the analogous issue with regard to written work, is there not? CONRAD: Absolutely. It's been clear to a lot of us who've been experimenting with it for the last several months is that you can get copyrighted responses. You can get responses that include copyrighted text, for sure. But now there have been people who've done more deep probing that makes it real clear that some very large data sets of pirated works have been used to train these models, which is kind of significant. HELLMAN: Not a little bit creepy. CONRAD: Yeah, for sure. And so you've asked about a couple other ones: bias. It's important to recognize that the data set is what's available on the web. It's available in data clusters. On the one hand, it's a whole lot of data that's been scraped. On the other hand, it's limited to what can be on the internet. There's no doubt there are definitely places where there are sort of data gaps, and that reinforces a worldview based on what's available on the internet. That's one thing that reinforces bias. The other are the people who train the models to make sure that they're aligned. That's a very charged word, "alignment," so I'm not going to use that anymore. But I will say just so they look like what we think that the people who are asking the questions want on the other end of it. There are lots of embedded biases, and there are a lot of people who have written about algorithms and other algorithmic bias that are really important. I'll just mention a couple of names if people want to have some fun like reading over the holidays. Safiya Noble's "Algorithms of Oppression" is one. Cathy O'Neil's "Weapons of Math Destruction." You heard that — math. I love that one. It's good "dad joke." And Joy Buolamwini, who is the founder of the Algorithmic Justice League. She has a book coming out called "Unmasking AI" at the end of this month. Joy, for instance, talks about how facial recognition software was trained primarily on white faces. When she was a grad student, she was trying to work on some of the software and she was testing it, and she put her face in front of it on her screen. It wouldn't read her face. As a Black woman, she was like, "Huh, I wonder what's going on here." And so she brought in one of her roommates and put her roommate in front of it, and it read her face fine. She found that it would only read her face when she put on a white Halloween mask, and she's like, "Yeah, this is a problem." These are used for policing. They're used for identification. And I'm also gesturing out towards AI that's beyond necessarily generative AI. This AI is a big tent, and it's involved in everything from generating marketing copy to surveillant policing. It's kind of important to consider these all in a larger network of kinds of technology. HELLMAN: Next, I wanted to go to the meat of your article itself — the proposed a bill of rights for education. And you say that you were impressed by and modeled it after the Biden administration's Office of Science and Technology Policy, which in 2022 released a blueprint for an AI Bill of Rights. Can you talk about that and some of the main points that you think teachers and students need to be aware of or empowered to apply to their lives? CONRAD: Absolutely. I started talking to people on social media about what our responsibilities were as educators and trying to figure out a way to start a conversation that we could all be part of in the curated AI feed. Now everything on my social media seems to be about AI. Somebody had mentioned earlier in the spring, this bill blueprint for a bill of rights for our AI Bill of Rights — that's what it's called — and what the White House had put forward last year. I was like, "Oh, that's great. Let me look at that." And so I'll just read the main points of it. There's more data, if you go on to the website — it's still on online — you can read some of their examples and fuller description. I think it's important for people to know that this is out there. The first principle is safe and effective systems. You should be protected from unsafe or ineffective systems — probably infected systems, too — and algorithmic discrimination protections. You should not face discrimination by algorithms, and systems should be used and designed in an equitable way. That gets to some of the things we were just talking about: data privacy. You should be protected from abusive data practices via built in protections, and you should have agency over how data about you is used. Not just an explanation; you should know that an automated system is being used and understand how and why it contributes to outcomes that impact you. And human alternatives, consideration and fallback: You should be able to opt out where appropriate and have access to a person who can quickly consider and remedy problems you encounter. Those are great. I loved those, and then I clicked thr…
Full show notes at the publisher 
Our TOPPODCAST Picks 
Stay Connected
