LCL Technical: Book Discussion: "The A.I. Con" by Emily M. Bender & Alex Hanna

[Music] Today we are going to be discussing a book. It's Emily Benders and Alex Hannah's book, The AI Con: How to Fight Big Te's Hype and Create the Future That We Want. This discussion is a continuation of the talk that we had a couple months ago where we were talking about AI generally and the topic of this book came up and we decided it was a great idea to dive in a little bit deeper to it. So today we are going to be going through some of the topics that are raised in some more recent news that's happening. Uh so first let me introduce our panel. We've got Aubrey Amstutz who is a cognitive linguist and currently responsible AI research manager at Grid Dynamics. Katie Swindler, who is a specialist in discourse analysis and social linguistics, and was most recently pro uh program manager at Mother Tongue AI, currently freelancing as a consulting linguist. And finally, Professor Alicia Beckford Wasink, who is a professor of linguistics at the University of Washington, also the director of the Social Linguistics Laboratory. This book, I just want to say at the outset, is amazing. I I think we were all really impressed with the quality of the work that goes into it, but also the accessibility of a book. I have to say when I started reading it, I was a little worried that it wasn't going to be accessible to non-llinguists. It absolutely is, but it's not a pop science book in that it doesn't dumb anything down. It's written very clearly. It um presents a lot of linguistic concepts in a very accessible way. So even if you don't have a background in linguistics, you will be able to understand it. And it really leads the reader through all of the ways in which AI, putting quotes around that right here, is worming its way into every aspect of our lives right now, for good or for bad, and ends on what I think is a really helpful note about what we can do. And I think that's one of the things we want to end on today, too, is we're linguists. what can we as linguists do uh about what's currently happening in AI and the language aspects of it in particular? This book was published in May and it already feels like a million years have passed since it was published. Um, just before we got on here, we were talking about how there's news like every 5 minutes about what's happening in AI and tech and it's really difficult to keep up with what's going on and to evaluate kind of rationally all the news that's happening. So, we'll we'll touch on that a little bit today. Um, I guess my first question to the panelists is what's been top of mind for you very recently with all the news that's coming out every five minutes since the book was published? Because there was stuff yesterday and the day before. There was stuff this morning as I was scrolling through my social media. Um, I I think one thing that's been really interesting is what's happened with the newest version of chat GPT, which um is the languagebased thing that most people interact with. So, I have thoughts about it, but let me turn it over to my panelists, whoever wants to jump in there. What are your thoughts on the situation right now, Saturday morning, August 16th? >> I can start. Um, I have several bones to pick. I guess nobody is surprised by that. Um, but the first thing that stood out to me was the framing, and I'm going to harp on this through the entire discussion, like framing our discussion, framing our public discourse around AI. Um, but the idea that GPT5 has PhD level intelligence, that was like a big selling point. It was a big talking point. um I think fundamentally misunderstands the experiential nature of learning um the experiential nature of trying failing learning from colleagues iterating on research and all of that is almost I don't want to say more important to a PhD than the actual research itself but it's such a tremendous part of it and it's almost insulting that you could say a model has PhD level intelligence when obviously there's no real like understanding comprehension engagement with the content of what it's talking about talking about. >> I would echo that. I think that that has stood out to me as well and I have been insulted by it. It definitely feels like there is um a a complete disregard for what goes into uh the preparation, the work, the familiarity with um the uh really brilliant people whose work came before that you had to engage with in order to actually earn the PhD and the importance of remembering and context textualizing that um community of practice that you've joined. I mean part of the part of getting a PhD is like you said not just knowing a set of facts. Um it's actually about engaging with previous work and knowing where your work fits in. It's also knowing what the shortcomings are of your own work and knowing what its strengths are. every one of us who completed a PhD has to talk about future research and has to talk about, you know, those things that we didn't get to cover in our PhD. And so there's always um a way of acknowledging that there is much more to be learned. And I think that I mean I'm speaking for myself, but you know, the more I worked on my PhD, the more I was aware of the things that I didn't get to and the things that um really did need fuller treatment than I was able to give. So for GPT5 to be touted as having PhD level intelligence, right, we're missing all of that. We're missing all of the situatedness of um of the conduct of research and the community of science that is involved and also the things that we don't know even after finishing the PhD. I think the opposite is claimed, right? That that there's the highest level of knowledge on uh a particular topic. That's what that framing right PhD level re PhD level knowledge suggests to me that you know we know as much as there is to know about the thing and we're going to give it to you quickly and easily. I definitely agree. Um I think the the framing is really important like you were saying. Um I think it makes me want to ask the question. I think this is one of the kind of suggestions in the book is like asking questions about like okay well what is how are you measuring that what is PhD level knowledge to you what does that mean um and I think it is a little ironic that they're at PhD level knowledge and it's kind of doing two things at once right because it's saying okay we understand that PhDs on one hand we understand that PhDs are responsible for starting to create knowledge like you're getting into that creation of knowledge that's the whole point right to learn how to become a researcher and then at the same time they're saying well it has PhD knowledge which means that it has memorized a certain level of facts but it's not able to do that research or is it so that kind of brings us into that next step of saying like it has accomplished this level of knowledge but I would say that part of that knowledge is the knowledge of being able to create research and so that brings us to our next question which is do we think they can really create research um so I think it's kind of an interesting uh choice that they would make to frame it that way as this is a static thing that it has achieved this knowledge but then at the same time we also want it to be these machines to also these machines these systems to also be able to create knowledge and do science and do that next step as well. Yeah. From my perspective the thing that I'm like really thinking about a lot is agentic systems. I know it's like maybe that was like two weeks ago and now it's you know there's something new but I think that um that is really like the instantiation of a lot of the fears and worries that we had about these systems and I think it's really I it's the perfect sort of um example of uh AI hypers wanting to uh create the thing that they are laying out why it will be such a bad thing. So they're saying, you know, if we give these AI systems access to all of our, you know, military or all of our whatever sensitive kind of um, you know, systems that are able to do things in the real world, then that's when we're really going to get into trouble if we give them that access. And then at the same time, it's like we need to develop agentic systems to be able to have the full potential of AI be realized. Um, so it's kind of that interesting um, create the problem and solve it kind of marketing thing. 100%. >> Yeah, >> it um takes me to uh the so two things um which are inextricably intertwined. Chapter five of this book is called uh artifice or intelligence AI hype in art journalism and science. And that particular chapter really spoke to me because it talks about exactly those things. So how do we know what we know? um why are we thinking that this so-called artificial intelligence has any sort of creativity to it? Because creativity is part of all three of those things, including science, right? That's something that people don't think about. But it's like Alicia was saying, creativity is part of what you do when you get your PhD. You're creating knowledge. Like you're you're asking these questions and you're using your human creativity to find out what the answers are. And there's no way that a computer can have that level which and the other part of that you know these things tied together is why do we insist not we but the world is using all these human terms to describe um what they say in the book is a synthetic text extruding machine which I love like that's a great way of putting it. Um I from my point of view that is part of the problem that we as linguists are facing. How do we push back against this constant anthropomorphizing of what is essentially pattern matching, right? Like you you just see it everywhere even and and we tend to fall into it when we're talking about it too. Um and it's it's hard to stop. So I'm just throwing that out there as a a general topic. Like it makes me insane to see people using all these human words to describe what is just pattern matching. I think it's a huge problem. >> Yeah. Yeah. It makes it really difficult to remember what's really going on. Sorry. >> Oh, no. Go ahead, Aubrey. >> Go ahead. Well, I was going to say that one of the things that I actually love about this uh book, and I've gone back and actually reread it several times, is uh the section in um chapter 2 actually, where they talk about why it is that we um interact with it as if we think that it had a mind behind it. So, why do we think that there is a mind behind um the AI systems that we're interacting with? and they talk about the idea that you know these systems um automation systems right and AI is not a thing it's a collection of things that are automated that have been packaged together for marketing reasons as AI but you know they talk about the idea that when we interact with other humans we have joint attention on a shared topic we have the ability to um build meaning in context intersubjectively. And these systems have been programmed to be able to mimic the form of human interaction so well that it's easy for us to imagine that there's a mind behind the system that we're interacting with. And I think that that discussion is so so helpful, right? Um they clarified that there's no real understanding. there is no understanding or care of us. And so when the systems respond to us in ways that are harmful or ways that show that our communicative intent was not understood, that's when the shortcomings happen. That's when the short circuits happen. But um the systems have been programmed so well to give reasonable text on just about any topic that it's really easy for us to imagine that there's a mind there interacting with our own. And so we think that that's you know u the same as when we interact with others. But, you know, I I love I love the phrase that they say that computers have no subjectivity to be intersubjective with. And I think that is, you know, that's going to be a phrase that I'm going to use forever because I love it so much. It really brought home this idea that, you know, systems have no personal feelings. They don't have tastes and opinions. They don't know anything about my history with them in interaction as an interlocular or as a friend or as a family member. None of that is there. Um you know it's just the case that a lot of information has been scraped together and then put together to be output with humanlike um phrasing and uh colloquialisms so that it will feel like there's a mind there. And I think that's one of the things Laurel I think that's, you know, one of the reasons why um we're able to think about these uh systems as being humanlike and do some of that anthropomorph. Wow. >> Yes, it is in the English language for this language to say >> I think so. >> Yeah, I agree. Um, go ahead. >> Go for it. >> I was just going to say, yeah, I agree. Um, and I think another phrase that I really loved as far as like kind of using ridicule or or asking questions, um, was the superhuman abilities. So, I think like not only are we now saying, okay, these can, you know, take over these jobs and things like that, which are human abilities, right? But there are these claims that it is reaching superhuman intelligence or superhuman capabilities in specific realms. And I think that when they uh used the example of a hammer, I think it was, and they said a hammer has a superhuman ability to like drive a nail into wood. >> And I think that to me really helped pop the bubble. And and I think like you were saying Lisia, one of your f favorite phrases is going to be that. One of my favorite phrases is going to be, you know, pick anything and it can have a superhuman ability to do that thing. That doesn't mean that that tool should replace humans. >> That's so good. That's the metaphors in this book throughout this book are just they're brilliant. They're beautiful. They're thoughtful. They're like really great at getting the concepts across. Um I have two points on this idea of like subjectivity, intersubjectivity and then how we um co-opt the idea of intelligence when talking about these these text extruding machines. Um the first point is that just to jump on the points that Alicia made is that not only do we as humans have this instinctual um assumption that there is a mind behind this kind of interaction. The idea of treating these models as human is really heavily pushed by the tech industry by media. Um, I was reading a book co-intelligence by Ethan Mllik and he talks he has four tenants for interacting with AI. Um, and one of them is treat the AI as if it's a person. And he even says we shouldn't be treating it as if it's a person. It's, you know, it's a fallacy that's going to lead to to issues, but, and it's a really big butt, it's easier to interact with these systems as humans when we assume they're human. Um, of course it's only easier if we're willing to accept the outcome of interacting with them like they're humans, but that's a different discussion. So, that's one point. The second is that um by framing this entire interaction as intelligence, we're really first of all narrowing our the implications for what is intelligence, which of course is as they point out in the book doesn't have a clear operationalization. There's no real definition of it. Um, intelligence is so tied to context, to culture, to um, you know, situation, to every specific interaction. Um, and it's really, really, really prone to being affected by systems of bias that we've developed in academia and outside of academia, just kind of in in general discourse. So, when we call these things intelligence, we're eliminating all of those pieces of the interactional puzzle that that Alicia brought up, that's um you know, feelings and thoughts and memories and experiences. Um and all of these pieces that we can't replicate using an AI. Uh something like emotional intelligence is totally removed from the picture. And it's not made clear when we're discussing AI how narrow we are making the definition of intelligence itself. >> Yeah. I see a lot of parallels actually from my time thinking a lot about teaching English to non-English native native um English speakers. And it's a it's a lot of like reframing of what you think is happening in a discussion, which is always kind of what linguistics is doing, right? Um, but when you're starting from potential uh lack of like common world knowledge, um, like cultural world knowledge, norms, things like that, you realize how you really have to build that conversation or that interaction from the ground up. And it's always with this goal of like communicative intent, right? Like getting something done, getting something across. Um, and so you realize like that even the little noises we make, you know, when we're thinking to let someone know that I'm thinking and I'm not done yet. I just need a moment. Those noises are different, you know, cross-culturally. So, you have to learn how to make those while you're thinking and you're always thinking when you're trying to speak, get ready to speak in another language, right? So there are like these little things that um not only in the interaction but also in the like content that you realize you should you really are taking for granted when you're not um interacting with people from outside of like your kind of normal sphere. Um I think just that uh ability to to rely on like certain concepts that are like status quo or common knowledge. Um that is something that you realize like oh I I think I actually need to explain this and we need to have these things in common to be able to even have the conversation that I want to have. We need to like first set these ground kind of basic understanding of what is going on. And that to me just comes rings in my mind over and over again when I'm interacting with um AI systems is that I don't know what the common ground is. Like I don't know what we share um but I know for sure that they've never had a body and they've never been in the world. And that's a lot of what we um a lot of the the kind of just epistemic expectations of like understanding why things act the way they do because of physics, right? Like there was a lot of um ways to trip models up asking them very basic questions like you know there's something sitting at the top of the stairs and like something else hits it and it ends up at the end of stairs. Why did that happen? And it's like no idea. um where for us it would be like even a toddler or like a dog or anything would understand like how cause and effect works in the real world and I think we just take for granted how much of that is actually necessary to for intelligence like intelligence right I'm putting quotes around that um is just common world knowledge like that is actually something you need to build off of and when the AI system will never be able to have that in-person embodied uh experience of the world I feel like there's just always going to be this huge gap in our conversations um that is sort of vague and um often unressed I think. >> Yeah. Um as you I'm thinking so hard as you guys are all talking because it's making my brain bubble. Um I I think at least you use the word interlocutor and I think even that is anthropomorphizing right like we are not having a conversation with AI we are not interloccing like loution is a thing that people do right it's just it's so hard like you can't because we have all these >> we need to invent a new glossery >> we do absolutely so that just pinged me and I was like no that's not what we're doing we're not we're not loing with with AI but um >> um but to to what you were just saying Aubrey and and this definitely comes up in the book that yeah the definition of intelligence is like this big. It's a monoculture and that monoculture is um based on um you know straight white upper class rich guy insulated from the real world. And the roots go back to things like incredibly racist things like IQ tests, right? Like IQ tests don't measure anything really how well you do on a test. It's like SATs and things like that. It and it's all based in trying to separate the good people from the bad people. Um, and that's all based in racism and classism and and many many eugenics. It's all eugenics, right? When you get right down to it. >> So that's what the intelligence of AI grows out of. So all of the biases that are in it, the unexamined biases in most cases come from a terrible place, right? we're we're not coming from a place and then the intelligence is measured as like how many facts can I >> fact again putting it in quotes because a lot of the facts that AI thinks it knows are are not true you know they're they're not actually facts >> and so people create words like hallucination which is again a human word like it can't hallucinate or even lying right like that's a human word too >> I don't know it feels like >> intention behind it >> with intention behind it and there is no intention So, um, it's it's an ongoing thing and it seems like it's getting worse to me because I don't see anybody trying to put the brakes on all these human type words and and it makes me insane when I I see things like um, you know, all the things that we've been talking about and all the the misinformation that gets spewed around and yet chat GPT5 still can't tell you how many blue bees there are in the word blueberry because it doesn't understand anything. Anyway, uh I it it's crazy to me that that this is where things are going faster and faster and and more um anthropomorphized the further we get along. It it's frightening to me. I kind I think it's critical that we see this as a part of that narrative that AGI is coming >> and we're going to, you know, replace humans with this super intelligence that can do all of these things that humans can't do. They're going to solve the social problems that we're not able to solve. And I put that heavily in quotes. Um, it's all part of this narrative because the more that we treat these systems as human, the more that we're overlooking the fact that they don't understand, they don't have comprehension. One of the things that came up for me a lot while reading um was the tie to what we talk about um in social linguistics as the idea of like doing something. So you're doing listening, you're doing being a good student, you're performing this persona that is getting you something with a specific social group. There's a performance that is creating a result in the real world. AI is just doing being intelligent without any substance behind it. Right? And if we allow the narrative to continue to promote the doing of something as the something, if that makes sense, >> people are going to totally overlook the fact that these models don't understand, they don't reason. >> I know we talked about chatbt has reasoning mode, right? There's a whole other, you know, discussion to have on that. Um but functionally the tech leaders treat that as the same thing as completing the task with that very human embodied you know experiential um style of interaction. >> Yeah, I'm really glad that you're tying this in with the AGI. Um I think like I think it is really important like you're saying to contextualize this in this kind of longer narrative that's been going um that proposes like we don't have to accept the assumption that there is a hierarchy of intelligences right I think even in linguistics we can get a little bit into this territory where we're like humans are the only ones with language and we know that for sure and animals don't have real language and this is why these are the things that make a real language and a real you know intelligence right so I think like we have to be constantly reminding ourselves that we only have a certain positionality in the universe as well, like not even just in society, but that like the wisdom, the intelligence of a tree might be a thing, right? Like I don't know. I'm I'm not going to say that that's not, you know, an equal intelligence to me in some way. Um and so I think that like this uh necessity to reinforce and reify the structure that says we are different there. It's the ingroup outgroup thing again, right? like saying it comes from these eugenics and these racist and these just sort of, you know, uh trying to have a supremacy system roots that folks want to okay, we're going to separate these groups out and then we need to find ways to clarify and remind everybody why we're different. And that's where we get into like intelligence, like you said, IQ tests, you know, um measuring heads to say how big the head is has something to do with intelligence, right? these things that people invented under the guise of science to remind and create barriers and create a hierarchy and then I think this is the echoes of that you know we're still trying to find something that may not be there right we're trying to say that there is a difference in in these types of intelligences and so I think like even the presupposition that we could have a super intelligence reinforces this idea that we can't be just different types of intelligences in the world in the universe and that there needs to be district hierarchy that gets also into some religious stuff and like you said, you know, sort of like who who created us as intelligence? You know, if that's God, then are we the new God? How convenient for those people. Um but yeah, I think it's really really important to remember that this is not some sort of um mistake. This is not a a new phenomenon. This is part of like a long long long uh history of propaganda essentially and worse. Yeah. >> And also >> that there's a way in which I'm I'm thinking about human harms >> and um you know there are so many examples in the book about how um there are voices that are being excluded in the production of the uh information that these models are being built on. And I think you touched on it a little bit um Katie, but I want to draw it out more because I think it's super important. You know, there are people whose knowledge is being um rejected or not being represented. Um there's a lack uh as the book really spells out um of information from black and brown people and indigenous knowledge. Um there are ways in which people who are trans are absolutely, you know, not being treated well by these systems. And we keep reifying this information which makes our systems continue to be biased and not less biased um toward these knowledge makers and these humans whose experiences are are really being um left out and whose experiences are being harmed. I think that there's a um a wonderful bit where they talk about how, you know, when we're talking about systems that are actually being used to determine recidivism or or to predict the amount of um risk there will be of leaving a child in a particular home, right? That tends to result in more um family separations for people of color, right? When we have systems that are being used in these ways and we are treating them as if they are um you know giving us superhuman knowledge, right? What we're doing is we are making it more and more possible for people to be harmed um by the reliance on these machines. And we are um we're that's definitely racist. It's definitely built on a history of exclusion and of the mistreatment of certain types of people. And I think it's super important for us to really ask questions about what the inputs are to these systems. Who's being harmed? Who's being left out? These are some of the things that were encouraged to do in the last chapter. And you know, because I do research on this, because I actually study the impacts of the failure of AI to work for certain people, um I see them respond and I understand what the the burdens are on them psychologically to be um left out. you know in the research that I do the number one you know people are invited to tell us about their experiences and um the number one or I would say the top five terms that people use is that they feel invisible that they don't feel like their experience is being represented that they are frustrated and that they are annoyed yeah there are people who say that they love it you know um one example from my um corpus is someone who was hospitalized uh a student and had to complete assignments and they um used chat GPT to help them to dictate their ideas for writing a paper. They finished their paper, they were able to submit it on time. This is one example that we got, you know, from a respondent who is really happy. But, you know, I have other um responses from people who have to interact with AI and their jobs. Um, and this is a person from the Gulf Coast who said, "I always feel dread calling insurance companies for my job because the speech recognition system never understands me and it misconstr. I wish I could opt out of the recognition system because I end up in a negative feedback loop." Right? So we don't get the opportunity that could be a recourse right of opting out but um in some cases we can't even opt out of the use of these tools um and uh so we have to continue to experience being harmed by them and I don't think that that's being recognized as an impact on on real humans. Um right when when people like Sam Alman talk and say you know yes these are stochastic parrot systems and so are we right it is denigrating human intelligence just like you were just talking about Katie and so that is I think super important to recognize that these systems are biased that people don't often have the opportunity to opt out and uh that um that certain types of intelligence are being left out >> completely. Uh chapter 4 in the book is specifically about those harms that are done in the social sector in the medical sector. And it reminded me that um about 10 years ago there was a big movement in um health care to try to not solve the problem but mitigate the problem anyway of all the people from different cultures in the US who are now accessing the health care system. some people who didn't speak English at all and some who didn't speak it well and that we needed to solve this problem with people, not with technology, but with people. So, um I know several companies who were employing linguists and and other folks and translators to try to bridge that gap between patients who were thrown into this horrible health care system we have in the US to try to access care. um taking into consideration things like language, things like culture, things like social status. You know, when it's a a family of immigrants, who is allowed to talk to the doctor, right? That varies from culture to culture. It's just different. Um what are you what can you say to the doctor, right? In some places, in some cultures, you can't just come out and say, "This is my problem." Like, you have to talk to it because that's the way things are done and that's just how it is. So companies recognize this and we're like, "Yeah, we need to solve this problem." And what I see now, especially from the examples that were given in the book, is like, "Well, AI is just going to take care of all that." And obviously, it's not because now you're taking the human element out of it and you're losing all of the context. It's just gone. So what if you are a person who can't work with speech recognition because of, you know, the way you speak, because of your accent? What if your cultural norms are completely different and you're trying to act interact not interlocute with an agent and you can't come out and say what the thing is that's wrong with you because you your culture tells you that's not the way we do things. What's going to happen to all those people if you know health care systems insist on employing um agents, fake agents that aren't actually real people that can't do it. Um, I mean, I I have I am a a a white person, a middle-class white person, and even I can't make those systems work sometimes when I'm trying to get in touch with my healthcare provider and I get stuck in a loop either in a chat or when I'm on the phone. Like, if I can't do it, what about what about everybody else, you know, who people who don't have the fluency and technology that I have after doing it for so long? It makes me insane. And I don't see that stopping anytime soon. Um because the hype is so great, right? Like we have to put this into everything. So, >> and also, >> sorry, go ahead. >> Oh, go ahead. I was just gonna say it ties into our discussion before, right? Our last podcast where we were talking about like where does this go? What do you do when there's no more training data? Like if there's not a bunch of this translated material that translators painstakingly did um available like that's up to date with you know the slang, the terminology like new words are being invented every day. >> Where are you going to go for that? And once you don't have folks who are um skilled at doing that first initial translation, they're only skilled at cleaning up what AI does, that's a different skill set and you're not going to have any more training data. So it's also just like I don't understand where the trajectory of this goes. I I want to drill down on that because one of the things that's so interesting to me is that a lot of these companies now they are running out of data. They can't access data. People are putting you know blocks on their robot text so the web pages can't be crawled. Like people are really waking up to the fact that AI is just taking everything everything that's not nailed down as the book said which I love. It's true. >> Um and so companies are now using synthetic data. Where do we get synthetic data? From another AI. So you have these AIs and and the book uh makes this point quite well is that foundational models um so like your claude your chat GPT they're being treated as kind of general purpose tools right everything is a nail and that's the hammer um but these tools are trained on data sets that are fundamentally biased racist exclusionary from a very specific point of you um I want to plug a book I read recently called Atlas of AI by Kate Crawford that talks about so good. It talks specifically about the um issue with trans representation that you brought up Alicia is that a lot of these models were trained on data that does not include anything except binary gender. That's it. So there's no hope. No matter how many guardrails you put in, no matter how much fine-tuning you do, you're never going to see a a realworld or even like minimal amount of representation of these identities in a foundation model. Then we use the foundation model to generate synthetic data that more specialty models are being trained on. Well, if the synthetic data is biased and not high quality, which it is both of those things in my experience, then your very specialized model is just grooving those patterns of of bias, of low quality, of um, you know, of eraser that were already present in your original training data set. So, the problem is it's it's even worse than we think it is. And and as Laurel said, it's very hard to see how we stop that. the foundation models are built. They're constantly tweaking them, but they're not going to train them on a whole new data set. Like that's not >> feasible. So, >> yeah. And I saw some research recently that that said that that showed that um the like synthetic data and I hope I get this right because it was it was a while ago. the synthetic data um that was produced that was about or like meant to be by certain groups like marginalized groups usually represented the perspectives of the majority or dominant if you're if you're yeah using that um paradigm the dominant class's view of what they thought that that group would say or do or think. So it was basically extremely stereotypical. It was like them trying to write as if they were pretending to be from that group. That was what the model was essentially doing when looked at compared to like uh language actually produced by that group. >> It's ideological puppetry essentially. >> Exactly. It's not. >> It also makes me think about how the book talks about the idea of sampling in silicone that you know we even uh attempts to replace human uh respondents or participants in a study with uh AI um faximiles in scare quotes, right? and and trying to do a study as if those were humans who were providing responses to help us solve questions. I think that that's absolutely insane. And for all of the reasons that we're talking about, um the models are biased. They're they're, you know, using stereotypes of what they think that these people might say. And by the way, a lot of transcription is inaccurate. So what they're using is not necessarily a faithful representation. Even if they say, "Oh, we got this from a real person, right? The what the real person said or did might have been mistranscribed um misrepresented." So, the idea of actually using these um types of of automation to replace humans in in research samples scares me. It really frightens me. >> Yeah. >> Could not agree more. uh the the areas where it it's so concerning. I mean, healthc care is a huge one, right? Just because that's people's lives on the line there, but in law as well, and that's something that, you know, I I had studied for a while, Lisa, to your point, when you look at transcriptions that get used in like court cases from interviews or from wiretapping, and it's not straight white people who are talking, they're so wrong because the people that they have doing the transcriptions aren't linguists. They don't know. they're just guessing at what people are saying. And of course, a transcription has no context whatsoever around it. So, you don't know if somebody's joking or, you know, whatever it you can never know from a transcription what's happening in the conversation. So, if those sorts of things are going into the the training databases, which I'm sure they are, that's horrible. And then when you know people are already using um fake AI lawyers to help them and they get caught and they're wrong, but what happens when they don't get caught? What happens if people aren't doing their due diligence when there are madeup court cases or um citing laws that don't exist? You know, it's happening all the time. Again, faster and faster. It's I I see more of it in the news. I I think less than a year ago was the first time I'd seen something in the news about someone actually using uh an AI bot to help them in a court case. And now it's like I see something every week about people trying to do it. So the the speed of it is overwhelming and I kind of wonder um as we hurdle towards oblivion, you know, like it it does feel like that though like like we're shooting towards >> I think it's intentional too, right? like trying to make us feel like it's inevitable, right? >> Yeah. So, so we're going to come to that wall eventually and then what's going to happen like is it is it that everything is going to fall apart? I don't know. I I feel like maybe we can't stop things. We can mitigate things, but I also feel like linguists are going to be really important in picking up the pieces when things actually fall apart and the center can no longer hold. Yeah, I think that this is a great place to plug uh an earlier career cast episode called you need a linguist for that, right? Because there is a there is a way in which um that you know a lot of the the information that's being created, a lot of the experts that are being consulted are not scientists who actually spent time studying the phenomenon that is being represented. Right? So the the um the book talks about uh I think this was also in the definition of intelligence and um the work that was being represented was uh rather than work produced by psychologists, it was a definition that had been uh given in the Wall Street Journal or something like that. And so there was a there was you know we're we're trying to say that we're doing science on a particular area of expertise that people work on but we are consulting um definitions that are in the popular press or we're taking the output of other um searches online searches and using that instead. And it is absolutely something that I think linguists need to do to be able to say, "Hey, you need a linguist for this because we've studied this. We understand this." Um, and it's going to take us kind of inserting our voices more. But I do think that there are places where our expertise isn't even known. You know, people don't know that we exist, that we do the work that we do. And you know, I've encountered this in in some of the consultation work that I've done outside of my teaching position. Uh people don't even know what linguists do or that we have knowledge in certain topics like language variation and change or like language attitudes or ideologies or intelligibility in what makes speech intelligible. um that you know we we we perhaps need to do a better job of of uh getting out there and saying we work on this. Um you're not consulting experts. Um you would you need to be able to consult experts on this, you know. Uh and one of the things that I was I was actually thinking a second point that I'd like to make is that um you know when you do a search now and uh the uh generative AI um kind of overview shows up at the top of your search. I don't know how many times I wish that it would say I don't know anything about this but you should consult these experts or this field of study or this you know and that's not something that happens you you provide a query and it gets a response every time >> and that is one of the things that you know might point people in a direction where they can actually get to the science but that's not something that happens You might get some citations, right? It may give you some information about where parts of the definition were drawn from and then you can follow some links, right? But uh but knowing uh the science and how it is produced by real people who have been working for generations um on particular issues is not something that we can expect AI to point us to. And I actually hesitate know we've been using the term AI a lot. I would like to say automation instead maybe because I don't believe that AI is a thing. Um, and so I'm gonna catch myself and try to and try to change my own usages in this conversation. But >> yeah, I think a lot of times we're talking about chat bots too, so we may be kind of like having that one use case in our mind for a lot of these conversations. Um, >> agreed. Yeah, I think I I really love that you brought that this point up um Alicia about how you can always expect it to respond unless it has like very like you know with a lot of effort been trained not to respond and say hey I'm not going to engage with that which we see how that's really tough to actually implement in like the situations you want it to implement be implemented in. Um but I think that as humans we have we we kind of abide by these maxims right these maxims of um like conversation and how to have a you know productive conversation and one of those maxims is like us self-editing or sort of self um you know the self-restraint to not give an answer when we are not confident in our epistemics or we give those kind of markers right like I think or last I heard but you know you probably want to double check so like we're very good at like deciding whether or not we even should respond and then if we do giving the hedging that's necessary like the caveats and I think that that like kind of um that planning that processing and planning of how you're going to respond is something that people are trying to get these chat bots to do a little bit on the back end like they're prompting them on the back end. Sometimes you'll see their prompt engineering is like, you know, asking it to do these things in the same steps that you would see like uh in high school when you're learning how to like formulate your ideas for a paper and you're, you know, looking at a rubric and you're saying it needs to have these things. So, I think that like they're trying to teach uh via prompting trying to teach or instruct the um chat bots to like plan what they're going to say ahead of time, decide, you know, how how likely is it that, you know, I actually have this information or something like that and then decide to respond. Um, so I think it's going to be interesting to see how those uh efforts play out. Yeah, there's been a real turn towards that like chain of thought style like reasoning mode. >> Um, which of course we could get into the weeds about whether or not that's actually reasoning, which of course it's not. Um, but I think the concept of friction from user experience is really useful to me here. Um, one of the things that they discuss in the book is that these systems are so embedded. We're talking about AIO reviews, right? It gives you a response every time. You can't opt out of AI overviews. Um, they mentioned uh Meta has replaced the Instagram search bar. It's like it's only AI. That's your only option. Um, and that follows the trend of the industry of user experience where the goal is to reduce user friction towards the towards the tool that the company wants them to be using. Right? So, we reduce friction into the use of AI. people are going to naturally use more AI. And by reducing friction, um, for people who maybe aren't familiar, it just means making it easier to click on the AI, making it more interesting, removing, um, like having to sign up for it, having to pay for it, removing barriers to using it, that kind of thing. Um and one of the points in that last chapter about um what are some things that we can do is is is that we should seek to have some friction in points of information access because they require us to consider where the information is coming from um how it's situated in the broader field in the broader conversation the broader context. So, we talk about something like AI overviews where you get links, right? And that's helpful. We like sources. That's a positive step for sure. But sometimes the links that it brings up are either not relevant, they're from non-authoritative sources. Um, they're the output of the language model misconstr. Um, and I read this somewhere that Google has done research that shows people click on links far less than they used to. I'm not sure what the percentage is, but there is a a measurable decrease in the amount of um like continuing research that people are doing into these links, which of course when you were just using Google search with no AI overviews, uh you had to click the links to figure out what was going on in them. Um, so this idea of how do we increase the friction to using these systems, which of course is completely counter to everything that a lot of the tech industry is trying to accomplish. And then um how do we encourage people to use other avenues to access information? Something like libraries, even just like doing their research online in a more comprehensive way. I think that's a really tough question because we don't want to reduce the friction to zero. As we can see, that leads to kind of a a lack of critical examination of the sources. Um, but you know, if we follow user psychology, we we do want to reduce the friction towards information sources that are more likely to be accurate, that are more likely to lead to um a more nuanced picture of of whatever the information is the user is trying to access. That point about friction is so interesting. As you were talking, I was remembering, you know, this is an off told sort of paradigm. Friction produces knowledge, right? Because friction is a thing you didn't expect. So the classic example is you run an experiment and you get an answer that you did not expect and you go, >> huh? And that's what leads to scientific breakthroughs, right? The like that's not what I thought it was going to be. So framing framing friction as a bad thing as they do I think in in many places is wrong like like you don't want so much friction that you can't do the thing you're trying to do um make an appointment or whatever access your bank account but reducing it to zero takes away any kind of impetus to find out more or investigate or or just look at things from a different perspective um and then >> gives you a chance to have those expectations in the first place. >> Exactly. Yeah. Yeah. Um but but the thing I I I keep coming back to that anthropomorphizing and there's another point I want to make before we wrap up today. Um the the ability of the large language models to just spew nonsense is very much in the paradigm that we are kind of trained as western people to res respect. Right? So uh I will give a stereotypical example because this is just how it works. But some white man gets up and says a thing confidently that's 100% wrong, people are more likely to respond and go, "Well, that man knows what he's talking about." Rather than somebody else who is not a white man who goes, "I'm not sure. I'll have to think about that." So, we've been trained over a long period of time to respond positively to people who instantly respond to what we have to say. Even if it's 100% wrong, it's like, "Well, they answered me quickly. They must be right. They sound confident. They sound like they know what they're doing." And as as you were saying, Aubrey, it's been this long history of kind of funneling people into this mode of behavior. And and that's one of the keys of that sort of behavior. And I think that's why people are so um impressed, I guess, by what the large language models actually do because they do give you an answer right away and and you're like, "Oo, that's great." Um Corey Doctoro has written something. He's such a good writer. I can't recommend his stuff enough. Um he's the one who coined the phrase enshitification which is just wonderful about how everything in technology is just getting worse and worse and worse. But he he used the metaphor of comparing large language models to um uh vending not vending machines but you know coin operated jackpot machines where you put the coin in you pull it and mostly you get nothing but sometimes you get the jackpot and that makes you think there's something magic about this like if I just keep pulling it enough times eventually I'm going to get what I want and you forget all the times when it was wrong or it didn't give you the right answer or it gave you partially and then you had to do all this work. The one time out of 10 or a hundred or a thousand that you get the right answer >> is the time that you get impressed and that's what you remember. What's it called? Katie >> intermittent Alex just posted it in the chat. Intermittent reinforcement is very powerful. Really, really powerful. >> Yeah. >> Like gamification is like exactly right. >> So that's partly what we're getting. I don't know how you fight back against that because it's such a human thing, right? >> Yeah. >> Well, you need more people in the room. >> It's a lot of speaking of like intelligences, right? Like Pavlov's dog, like that was a dog. So, I think it's it's not just human. It's just Yeah. It's something about the way that we work, our reward system. I think the solution is you need more people in the room who are um being thoughtful about how we apply these things in the room in tech companies where this development is happening. Um and this is something we touched on in our our last chat together is just the lack of diversity not only um of of gender, of race, of ethnicity, of culture, but of thought. there's just there's no diversity of thought because if you're not on the hype train, you're gonna get fired. Like you don't they don't want people who aren't on the hype train. Um and and I don't really know what the solution is to that except to seek out organizations who encourage diversity of thought because it's not going to be a lot of the corporations that are driving the development of this tech unfortunately. Um, >> yeah, I'm in a similar place where I feel a little bit cynical about how it will go, but I feel re like I do feel sort of reinvigorated when I learn about people doing like particip participatory design practices and like thinking about how to kind of >> surface that knowledge from groups and not just have those decisions be made by engineers who are that's not their specialty is like you know how humans interact, how societies work and grow. Um, yeah, >> it is out there. There's people doing that work with people who are out there doing it more. And I think you're talking about, you know, the the I love what Heidi Harley said in the chat, you know, about um about uh let me see if I can go back and find that friction being weight training for your mental and >> I love >> it is weight training and we need the exercise. We don't want things to be simplified for us and engaging. You know, I think maybe this is a great place to talk about the fact that the book spends quite a lot of time talking about um resources being pulled away from our higher education institutions. Um and so you know at the same time that we are listening to those really confident voices and being um impressed by the things that they're saying you know we are finding that people who are in communities of practice that are developing this knowledge together. Um we're we're removing resources from them and making it harder for them to do their jobs. Um, and I think that, you know, having having uh this focus in the book is super helpful to say libraries are a place where that we should be encouraging rather than pulling resources from because they are a place of access to uh authoritative information. Um, and we should be encouraging um the use of of them rather than defunding them. And I love the the emphasis on that. So would really encourage people who haven't read the book to to look at that. Um you know and also just the humanity of science. I think that needs to be elevated as well. And we um we might talk a little bit about this, but they they talk about the the use of um citation practices as being something that you learn when you are trained in in college. Um you learn what authoritative um resources are and then you learn how to give credit where credit is due and to engage with those people who have written and thought and done experiments and uh engage with that past research. Um you know and there are all all kinds of things that they say you know we already have worked hard on how to do science well on how to gain knowledge well we don't need to move away from that we need to double down on it um you know we double down on peer review on human subjects protections on careful sampling um and uh you know all of these kinds of practices that serve as a kind of positive gatekeeping um is what they in in knowledge creation. And so slow that means slowing down. It means um you know allowing ourselves to be okay with that friction that will help build our muscles right in these areas where um we might feel an impetus and urge and impulse to just get the easy answer. Uh because uh this is another one of my favorite quotes. Science is not a collection of facts. it is a set of processes um and a community of practice engaging around those. And I don't think we can afford to lose that. So I feel like one of the things that um I I feel really encouraged to do is to be old-fashioned in that way. If it's being old-fashioned, I'm happy to be old-fashioned. I don't need to, you know, red headlong into a a set of practices that that uh make me, you know, uh avoid that friction. I I want to encourage that and I want my students and my kids and my friends to know that those practices are important and so I think continuing to stay engaged in those conversations feels really important to me. >> Would you call yourself a neolite? Haha. Talks about lites and you know gives us some um some some good info, some details about how they weren't just people who resisted technology. They were avoiding they wanted to um push back on how technologies of the day were harming people. And I love that section of the book too. So thanks Aubrey for bringing that up. >> Yeah. And a neolite. I think >> it sounded like what you were describing%. >> Yeah. That's totally agree with all of that. Um, you know, it I was thinking about how this works in in the context of careers for linguists because that's kind of what we would talk about a lot and um this this bubble it seems like is going to burst or crack or or something is going to happen. And I feel like the linguists in our our society are going to be more important than ever when it comes to pushing back and correcting and that we as linguists just need to be really vocal about a lot of this stuff. Um, you know, Emily Bender has been on LinkedIn and I've seen her pushing back against some people who disagree with what she says in the book and she does it in a great way. U very calm and correcting just misinformation. And uh I saw her just the other day telling some guy who said, you know, oh, there's no references. And it's like, well, flip to the back of the book, actually, there's loads of references, and here's why they're good references and why these are better than listening to press releases from Microsoft or Open AI. Uh, and we can do that, too. I I know engaging with people is not fun, but sometimes it's really necessary when we're advocating for our rights as human beings to live in this society. And we can do it as we've been saying for this whole hour from a point that not everybody has because we are linguists and we understand how the the language works. So I think as far as calls to action um people should push back where they can call things out where they can maybe not call it AI and call them large language models or or whatever. Try to get all that anthropomorphized language out of the discussion because it's it's simply wrong. Um, and saying no to AI when we can. You know, you can't always, like Alicia was saying before, it's there in the AI overview. You can't turn it off, but don't look at it. You know, just go to the the actual results, not the >> sponsors. Yeah. >> Because every time it it gets generated, it's also a climate impact, right? Like we know that. So now like we're also brought along in this sort of like um you know we're brought along whether we want to or not in impacting the climate in this way just by searching which is like yeah another area of um that would be better to have an opt out as we've said. >> But it's okay because AGI will solve climate change for us. So >> yeah as Sam said it's going to solve physics whatever that means. Yeah. Just he just says stuff stuff just comes out of his mouth. Um, I've been >> like like something else we know well. >> I I've been reading um I I want to recommend um a journalist who's named Ed Zitron and he has a podcast called Better Offline and he also has a blog and a newsletter. He's great. Uh he he is profane but really gets to the heart of a lot of this stuff and doesn't buy into any of the AI hype. Um he's got a great background in technology. So want to recommend him. Yeah. Um, and from what he says and what I've read other journalists saying that companies now are at the point where they're recognizing that trying to force AI into their systems is actually causing more work. Right? So for a long time it's like oh AI is going to code a lot better and now they're just starting to realize actually it's making more work because the code doesn't actually work. And even in things like copywriting, when >> it stuff is generated by large language models, they're realizing it actually takes more work for people to correct what was written than to do it from scratch. So maybe there's a a a little bit of a a push back that's rising to the surface because people are actually seeing that it's not making things better. It's not helping. And you know, again, as linguists, we can kind of >> confirm that and push that and and maybe elucidate it for people. >> Yeah, I absolutely love that. And I think that we can um we can and should just question the promise of AI's benefits for all out loud. And you know sometimes you know we can make fun of it like they say in the book we can criticize it. It's our right to criticize it. Uh and we can point out the ways that it's not working. Uh and so that's that's something that I think is really important Laurel because we are getting just a it's the the airwaves are a wash with information about how it's doing so much good. And so we may question our own um experiences and we shouldn't. I think that if people feel like it's not working, they should believe that experience, not what we're being told um that AI is actually working for us, that these large language models um can do things uh that involve intelligence, reasoning, and so on. We should we should not question our own experiences when we have experiences to the contrary >> and those systems to the same bars right that we've had from previous like systems that were doing that thing. I think it's I don't know why but I've seen a lot of people sort of throw out the old you know metrics that that it needed to reach the threshold that something needed to reach for accuracy before it could be launched as a product. like it's sort of like well this is AI so you know 30% is not bad right that's actually that's pretty good and it's like we never would have settled for 30% success rate with a previous system and so I think being really strong for those of us who do work in technology continuing to hold those same bars and say this is the principle of what this product needs to do however that gets accomplished can be up to y'all if you want to try to use AI sure but we're not going to budge on like what our you know policy is what our intended use is what our expected behavior is and like >> if if you can't seem to get it to work, that's fine. We're going to hardcode, you know, some conversational design or something in there and then we're kind of going back to what we had before. But if that needs to be the fallback because it's not working often enough, like so I think like just reminding folks of like where and why we had these standards in the past >> um and that those don't go away just because it's a new type of technology. >> Absolutely. They talk about that um in the context of machine translation in the book of like people are trying to use AI to do translation generative AI uh generative synthetic text extruding to to do language translation um and transcreation. And it turns out that this whole field that we had before that's decades worth of research papers deep actually does a better job with the assistance of you know a human who is immersed in the culture and aware you know of the pitfalls and can can monitor the technology. Turns out that actually works better and it's a lot more efficient. We hear the word efficient thrown around a lot. I like I like have a twitch every time I hear the word efficient. Um the other piece of this we're talking about like in terms of educating people and obviously not all linguists are educators but a lot of them are a lot of us are um and even like not only educating at a high level. So talking to if you can get a hold of like policy makers even on the local level explaining to them how these things work. um talking to people, journalists, talking to people that you know in media, talking about, you know, what the actual applications of the technology are, what it can be expected to do, why it's dangerous to anthropomorphize. Got it. Um and then also your community, just immediate communities of people. I mean, the number of times that I have friends who who know I work in AI, they go, "How do I get a better result from like AI overviews?" I had somebody ask me that the other day and it's like well there is an answer to that but I also think you should know why using the AI overviews is not actually a useful way to spend your time and here's a different way to go about solving the problem. Um, and it seems small, but like those micro level grassroots interactions are the ways that we make people more resilient against falling through the hype because it really is. You have to develop a resiliency. You have to develop this instinct to question, this instinct to find more information, this curiosity. Um, and you know, there's a lot there's a lot that's working against us there. Obviously, we're living in a time of like pretty substantial anti-intellectualism and um you know, a push for efficiency, productivity, all of those things. And AI is being touted as the solution for those. But I do think I mean, you know, if you can even make a couple of people think for like an extra 30 seconds about using AI, that does matter. It has to matter, you know, because it it it snowballs. You can have a larger impact that way than you think. >> 100%. I I in the first chapter of the book, the authors take a good deal of time to explain how large language models actually work, which I thought was really good. And I I have tried to do Katie what you were just saying to explain to people who don't really understand as simply as I can that it's just pattern matching, right? That's all that is happening. It is matching patterns of words. There's no thinking, there's no considering, there's no nothing. It is simple pattern matching. And people are very surprised by that. >> They're surprised. They're shocked. Yeah. >> They they think that there's something else that's going on in the background and it's just no. Just think of really really fast computers and all they're doing is matching patterns. That's it. That's the only thing that it knows how to do. So, um, that chapter, the first chapter is great for going through the history of computing and things like machine translation, like you were just saying, and and all of the steps that got us to where we are now. So, um, I know we're we're coming up on some time here, and I want to say again, it's such a good book. Like, people should read this book. you should recommend it. You should give it as gifts. Don't buy it from Amazon, but um definitely give it out. And I wanted to say one more thing um which is um the whole anthropomorphizing thing. I'm sorry I'm obsessed with it, but it's in my brain. And two things came up recently, which is one people are using slurs for AIS for large language models. And the one I've seen most often is the word clanker, which is taken from a Star Wars movie. Okay, I guess although the form of that is really troubling, um the others are worse and I don't want to repeat them because some of them are really incredibly offensive and they're based on real slurs for real marginalized people. And I think it's horrible. Like the fact that people jumped to using slurs so quickly as if it were an inferior human being, right, is is like it's blowing my mind a little bit. So that's bad. And then the other thing that I just saw like yesterday was someone referring to their again big quotes AI boyfriend as wireorn. Wirebbor >> I've seen that >> Pikachu this is my surprise Pikachu face like what what what is happening? Like what I don't I don't I I can't and I don't and I don't like it and I I think we should speak out about this in the strongest possible terms. It's awful. like we can't let this continue. It's really bad. >> That's so interesting. I hadn't even thought of the like when you brought up the slurs thing like my first reaction was like that's kind of I don't know. I was like, "Well, maybe like this is an >> Yeah, me too. That was my first reaction." But then I thought about it some more and I was like, "No, this is bad. We shouldn't be using >> slurs." Like acting like it's Yeah. another like and and I think that that ties into the way that um AI was previously sold as more of like a subservient assistant, right? And like very gendered and all that like Alexa and all that kind of stuff, right? >> Um and so I think it's same thing that they're wanting it to be both, right? It needs to be like a subservient >> but also superhuman. Um and so I think >> yeah, using slurs, it's kind of interesting to me. Um, I'm not sure if I if it anthropomorphizes or not. Like I'm trying to think of an example of a slur for something that is not alive. I don't think it exists. So I guess it would be a different class of things. But I feel like we do want to encourage resistance, but just like via what method? So it's very interesting. >> Not name calling, man. I think that's bad. >> Yeah. >> There was a whole trend of people claiming that they got better results from chat GBT when they abused it. When they were like, "You're so stupid. Why don't you give me the correct answer? That was a whole thing for like a couple of months, like a couple of months ago. It moves so fast now that, you know, people are on to something else. But that was a thing. And it's like, man, I mean, I know we're not anthropomorphizing these, but I don't know. It's it's it's very human of us, >> right? >> Completely. Um, so, uh, we're going to take some questions if we have some questions. There have been a ton of amazing comments going by, especially from Heidi in the chat here, and we can read some of those out for people who might not have been following. But I do want to say, um, you know, we're not doomers in the in the language of the book. We're not we're definitely not boosters, but we're not doomers. And the last chapter of the book does go over some of the things that we've been talking about with asking questions, not accepting things, doing our little acts of protest, educating people, uh, most importantly about this. So by the time you get to the end of it after all this there are strategies for how we as linguists especially can help push back against it and u maybe guide the the path a little bit but at least you know have people have a more realistic sense of what large language models are and that it isn't just magic. Um so at this point Alex um is anybody got their hand up or do people have specific questions that they want to talk about? And please feel free to read some of the comments if if you want to highlight a few things that have been said. Like I said, so much interesting stuff going by. >> I want to thank Heidi and Grace in particular as well as other people who have been so active in the chat in providing resources and links. And these links will remain live for uh the podcast that will later be released and in the YouTube video. We will have the chat saved so that you can access these links. I want to, you know, we haven't had a specific question yet, but we've had really good comments, including just fundamental questioning of why use AI. As Grace has been saying in the chat, that's question zero is why use it? uh and trying to push back as as you've been saying against forces, leadership of organizations and uh marketers and people who would want us to buy these products, you know, push back and say, you know, why use this and where's the evidence that this will be of any um positive effect. So, I feel like I can take that to heart in my classroom and, you know, say, you know, why use anything? Why use any of these tools? and in my own way try to push back a bit. I'm going to I'll I'll read out a couple of the really good comments especially um coming from education and policy experts in the chat as we as we go continue on. Um I want to share an example that was brought up by Grace referring to the use of AI in the school context. So reading a comment, the stakes are higher with chat bots adopted in the context of school advisory. In a forum here in UAE, school counselors shared how marveled they were that their students were revealing to the counselor bot more things than to their counselors. Counselors were starting to use the bot as their assistant as the workload was high and accepted this trade-off. So many red flags, no guardrails. And that's something that was absolutely addressed in the chapter four of the book uh in the you know if it quacks like a dock AI hype and social services and we saw in the chapter many more examples of these AI therapy bots that have been brought in and used with very dilitterious effects. So that's something to guard against when we, you know, see um perhaps we have colleagues who are in these service professions who um some may view this as as time-saving or as useful, but we can perhaps, you know, call into question the use of these services and why they're being pushed within their organization. So I I appreciate that comment very much and that example. Shout out also to the state of Illinois for becoming the first state to ban AI therapists, which is great. So exciting. >> And I apologize, there's hammering happening outside my window. >> Yeah. And that brings up the question of regulation. I was putting in the chat a couple of links to some of our uh federal agencies that have been working on AI metrology and standards and testing and push back. uh and also developing types of guard rails for consumers and citizens and you know the government is basically doing things that private companies will not do at least that should be the business of government doing the things the gap filling the gap of the things that private industry will not do because there's no profit motive there's no shareholder profit there's nothing so filling that gap you know we need these um non-industry uh entities to to take on a role at a very large level. Uh to provide you know develop these guardrails with the help of you know experts and um you know try also to to resist the overtake of these tools and services that are being pushed on us without us asking for them. Grace notes in the chat, also check out our center for AI and digital policy, C AIDP. We run AI policy clinics every semester where we study all the a policy frameworks in the world. I um it raises the the issue that uh even though there are policies and laws in fact many of the very large companies like Meta for example just lie about what they've been doing right like it's like oh we're not scraping this information it's like no they absolutely are so I think as consumers and as um people who just live in in our western society we have to question everything right? When when a company like Meta or Open AI or whoever is claiming that they're respecting people's privacy, things like that, I I don't think we can ever assume that they're telling the truth. It's it's unfortunate, but we've had so many examples of them just flat out lying about what they're doing. I I don't think you can trust what they say anymore. >> And the nature of like generative AI being non-deterministic as well, I think is really important to keep in mind. Like people are used to systems being designed away and working that way. But you can't control like you can't control generative AI like it will always be nondeterministic. So I think that's something that like no matter how many policies you put around it like it also same thing it's not like it's a human who is trying not to do do the wrong thing like it doesn't even know that it's doing the wrong thing or violating a policy that you wrote for it. It's just going to keep outputting whatever is probabilistic. Right? So I think like reminding people that these systems don't work the same way as they they used to. You can't depend on them just because you write a rule for it to follow. It won't necessarily follow that rule every time. So I think that's another thing like just kind of uh shifting that like expectation. >> Really good point. And it puts a huge burden on the user, the consumer of these services to do all the leg work. like if they want to use them properly, it's an enormous burden to assume the tech company is lying to you. Assume that their narrative is not, you know, the actual reality of the world. Um, assume that the AI doesn't know what it's talking about. The text extruding machine doesn't know the content of the text it's producing on the screen. Um, and to then go back and fact check everything. And we know that because the friction to using these is so reduced and it feels so seamless and it feels like an interaction, right? That that you feel like you're getting inter subjectivity despite the fact that you're not. Um, you know, people are it's going to take a lot of work for people to actually do the factchecking, the leg work um that they need to do to use these tools safely. Well, I think one on that line, I love that the book talks about Karen and how um and the the role that journalists play in this type of factchecking and investigative reporting. It's absolutely critical for us. And I love that um she's, you know, she's got this uh collaboration with the Pulitzer Foundation. And I think uh it was described to train journalists on how to ask better questions about AI and do more of this investigative reporting because you know we consumers we we don't have the training and the skills to do it. We're getting better. Um and I think that if we improve our tech literacy ourselves, we'll be able to do more of that, but we need to support the work of of journalists who are doing it and doing it well. Uh and I I love the fact um that uh that there are journalists who are really on the ball with this. >> Yeah. Another suggestion um for those who didn't who missed it, Empire of AI by Karen How also really great book um just pulling together all of her research and experience with OpenAI over the last like I think decade or so something like that. Um and I just saw her post on LinkedIn today like oh I'm so excited. um I think is it Adam Garfield? I'm not sure the name of the actor. Um but she was like there's like a picture of him holding her book and she was like he's reading the book because he's going to be Sam Alman in the movie. And so like I can't wait to see the movie version of that book and I think that will help a lot with hype and kind of demystifying some of this uh some of these claims and yes personalities. Yeah. So very excited to see that come out. I think it's called artificial I also think that it's really interesting to see uh you know that in the in the in the wake of the release of of GPT5 right so many people saying no it's not better I like GPT4 better and um I think Alex made this point in the chat you know that that that people who were um who were paying right for the premium level uh have thought you know no this is not working for me I would rather go back and use uh an earlier release I'm not going to want to use anything right and this this idea of the business model saying you know as we develop better and better tech um we're going to take things that we gave access to you to for free and we're now going to make you pay for them and you're going to then be paying for something that you don't like you It's absolutely ridiculous. And you know, yes, encouraging that push back um to say no. Um I'm I'm going to go back to something that worked better for me. >> Totally. We are a little bit over time and we could go on. I I think we all agreed that this could be like a three-hour conversation. There's so much to talk about. This has been so fantastic. Um, just to tell everybody this will is being recorded obviously and we will have it up on our YouTube channel for linguistics uh career launch, but it will also be at the LSA's resource hub for LSA members. Probably in a couple of weeks we're going to do a little bit of editing and put it up there. Um, and now I want to throw it back to Alex to sort of wrap things up for today. Thank you so much, Laurel, and thank you so much to our wonderful panelists for this incredibly stimulating, very useful discussion. I so appreciate this interaction you've had today and please everybody read the book, get the book from your favorite local independent bookstore. I I can't thank you enough for for joining us here today. I would like to turn the floor to the current president of the Linguistic Society of America, Professor Heidi Harley, who has a couple things to say about joining the LSA. >> Hello. Hello. I also would just really like to thank all of our panelists and Laurel for an incredible discussion. Sorry about my voice. Of course, it chooses right now to do this. Um anyway, really I learned so so much from your guys's interactions over this book which I have not read but I'm going to read it right now. Um I would just like to encourage anyone who would like to have an embodied interaction with actual PhD level uh experts in linguistics like these beautiful women that you see before you. Please consider coming to the LSA's annual meeting in New Orleans this year, January 9 10 uh 8th to 10th or 9th to 11th. And uh we will all be there. There will be many panels of this kind. There will be um activities for people of course in academia also many many things of relevance to people outside of academia in government in industry in education we welcome everyone to please come and contribute and have embodied interaction it's so good I can't even tell you and of course you save a lot of money on your annual meeting registration if you're a member so please consider joining the actual linguistic society of America get access to the um resource hub where this and many other resources will be future webinars yada yada. Anyway, thank you guys all again so so much. It has been fantastic. >> Thank you so much, Professor Harley. And can't wait to see you all in New Orleans for the Linguistic Society of America meeting. We hope. And I would also like to invite you if you want to continue this discussion, we have a networking session immediately following. So if you would like to hop over to our proximity chat platform, Gather, we have a space where you can join and talk to our panelists for just a bit and we have um the link join information and password in the chat. And you can also look up very quickly information on how to join. But promise you it's very easy. You'll you'll join as a you'll spawn as a little avatar and you will get onto a map of our linguistics career launch space. Super low friction joining and inter and interaction and it will be just as if we are in a room together interacting in small groups. So please please join. This is positive low friction to encourage our interaction. >> Thank you. >> I I personally want to thank our panelists. You are all so smart and so amazing. So, thank you so much for um devoting the time to reading the book and then coming here today to talk about it. This has been super super great. Love it. >> And with that, I'm going to close out our recording. And I will also put our our panelist uh LinkedIn profiles into the chat. Thank you. Thank you so much, everyone. [Music]