FINALFINALIFAN

The transcript is a conversation between the host of a podcast and Dr. Alex Hanna, the Director of Research at DARE, a research institute focused on mitigating the harms of AI systems. Dr. Hanna discusses the impact of AI hype on various aspects of society and the need for critical thinking. She also talks about her background in sociology and computer science, her work on AI ethics and fairness at Google, and her current focus on the social aspects of AI technology. She explains the field of computational sociology and how it can provide insights into social questions using data analysis methods. She discusses the challenges of bringing social sciences into AI development and the importance of responsible technology. Dr. Hanna highlights the need to avoid harm and the limitations of terms like "responsible" and "ethical" in the corporate setting. She suggests considering notions of justice instead. I think the hype has just completely just taken over in a way that it's affecting how we work, how we play, how we get social services, and, you know, we need to really break that down if we're going to have any chance of really pushing back against this and seeing that the emperor, as it is, OpenAI, has now closed. Welcome to The World We're Building, a podcast on a mission to infuse the AI hype cycle with hope, agency, and critical thinking. I'm Allie, and I'll be your host as we hear from an eclectic group of guests ranging from annotators to activists and designers to policymakers, each pursuing radical work that challenges dominant narratives about AI development. Today we'll be hearing from Dr. Alex Hanna. Dr. Alex Hanna is the Director of Research at DARE, an interdisciplinary research institute working to mitigate the harms of our AI systems and provide tools for safer tech work. She's previously worked in AI ethics and fairness at Google. She's a sociologist by training whose work centers on the data that's used in AI technology and how this data can exacerbate racial, gender, and class inequality. She's also the co-host of a podcast called Mystery AI Hype Theater, 3,000, thank you, must be confused with 2,000, a commentary on the AI hype narrative and who it serves. I'd highly recommend you go check it out, but of course, after listening to today's episode. So today we'll be hearing from Dr. Hanna about a whole range of topics, including the meaning of a responsible data set, the challenges of multidisciplinary work, and what is the right project to focus on when the research agenda can be so broad as it is in the field of good AI. But before we get into it, Alex, I'm hoping you can start by sharing a little bit about yourself so that we can situate your knowledge and experience for today's conversation. Are there elements of your identity or personal experience that you'd like to help contextualize this conversation for our listeners? Yeah, sure. Thanks for having me, Allie. Really appreciate having me on. I can tell you a little bit about my disciplinary background. So I am a sociologist. As an undergrad, I majored in computer science, math, and sociology as a means of both touching on kind of the technical aspects, which I have been interested in since I was very little, always have been into computers, but then was really interested in the different social aspects and especially kind of social conflict and as well as different kind of social justice movements, including labor movements. And so I, you know, after I got my undergrad, I went to grad school and got a master's and a PhD in sociology, focused on a lot of work initially on computational sociology and social movements. And then after that, I was a professor at the University of Toronto for two years in the Faculty of Information and then went to Google for three and a half years, was on the various projects, but spent the most time on the Ethical AI team. And then after Google left and joined D.A.R.E. and have been at D.A.R.E. for about two and a half years now. I'm trying to think about what else would be helpful for folks. I am Egyptian American, trans woman. My family came to the U.S. in the 80s, so first generation or second generation immigrant, depending on how you measure those things. How would you describe computational sociology and maybe the outlook that it gives you as you pursue your career in the field of computer science? I think a lot of the ideas behind it are trying to leverage some of the methodologies that are called data science or big data or whatever, and try to leverage them and try to either answer some of the different dimensions of social life. You know, can we use some of the methods that we have developed, whether that's large scale data processing or novel data collection techniques, and then try to think if that provides us any insight into outstanding questions or hypotheses in the social sciences. And there's lots of different approaches to it, whether that's NLP or doing stuff with network analysis or doing any kind of image analysis. And I think that field is still, you know, beginning, but it's growing at a fast clip. And then there's the other side of it, which I find myself doing much more these days. I, you know, I write very little code these days and doing less computational analysis. And I think that's more about, and I think people wouldn't necessarily call this computational social science, but it is more about the analysis of computing from a social science perspective. This is much more of the science and technology studies type of analysis where you look at institutions, you look at different networks of power, you try to understand the practice of science and the practice of technology. And I, you know, I would sort of pair that with computational social science, maybe in an uneasy relationship with it. But I do think those are perspectives that are complementary rather than in conflict with each other. I think they can be in conflict at times, especially when it comes to things like data scraping and data collection and the ethics around that. But I do think that they do inform each other. How do you compare those two? And do you see them being very similar in process and strategy and values, maybe? Or how do you see those two being different? I see how a lot of technological processes have been infused into the social sciences. But I think what we're seeing now is how do we infuse the social sciences into technology? One of the things that I've done, spent a bunch of time on with University of Toronto Professor Alan Berry, is that we have been collecting data on student protests, leveraging some NLP methods to discern whether newspaper articles mentioned protests, but then also developing some different UI to code those protests, have humans actually code those protests for different information, and then doing kind of different analyses on those. And so that at the end of the day, we're still very interested in some fundamental questions about the social sciences, Terry, with, you know, are certain schools more oriented towards protesting? Are there certain students that tend to protest more? Are there certain issues that are policed more? Do universities respond more violently to particular issues? So the example I was thinking about, you know, we've seen in the, you know, the pro-Palestine protests, the encampments, that those have been really policed intensely. And so that's, we also saw a lot in Montreal, actually, in 2012 during the Maple Spring and the tuition strikes where those were intensely protested. Those are social science questions, which I think have persisted and will persist. The other way is trying to think about how to bring the social sciences to bear in AI and in machine learning. And I think that's more, that's a bit of a different move. That's more like, what are ways that we can bring kind of insights that social sciences have to make sure that machine learning or AI isn't doing, frankly, bizarre things? But the thing is, it's a bit, it's a bit of a, it's a bit of an unholy alliance because often that doesn't work, especially at corporations. They're going, we're going to build this anyways. Can you sort of tell us how to do this to minimize harm? And we're like, well, maybe just don't do it. And, you know, I think there's some drive to mitigate that. And I think a lot of things get brought over from the social sciences. Things like, you know, positionality, I think is very common to talk about. And the positionality is about acknowledging your social position, both in terms of employment and identity, but also geographical. Are you in the West? Are you, you know, in a place that has resources? And, you know, there's kind of limits to positionality. You can say, you know, I acknowledge all these different things and still we're going to build this particular product, which may actually work very poorly for a group of people or doesn't work at all or would make kind of, you know, have these secondary knock-on effects for a particular industry. And so I think it really depends on what the goal is at the end of the day. What is the impact one is trying to have? And, you know, how both the social sciences and computing can draw on each other. I think that's a really good segue to the next question, which is really just about responsible technology. It's a word that's being thrown around a lot right now without much in the way of tangible commitment or specificity. How do you define what is responsible technology? And then maybe we can talk more about how that might look different depending on the type of institution, the type of incentive structures you're adhering to. But to start, how would you define responsible AI technology? I think responsible technology is a very mealy-mouthed way of saying that one is really going to try to avoid doing harm, but is still going to build what they want. There's an interesting history to the phrase responsible technology, responsible AI in particular. I know when I was at Google, a lot of the discussions were trying to avoid the use of ethics, the terminology of ethics. You know, a lot of people don't like the term ethics. And to some degree, I mean, I understand the criticism because ethics itself, you know, tends to focus on the study of what is good or what is a good life or how to live a good life. And it does have a very long history in all areas of philosophy, whether that's Western philosophy, going back to Aristotle or Chinese philosophy or Islamic philosophy. And so, you know, like ethics does have those connotations. And some people have said, well, we should get away from saying ethics because that, you know, we have this notion of ethics washing and we should talk instead about notions of justice. And I accept that critique. I think that's valid. Within the corporate setting, you know, there's no talk of ethics nor talks of justice. It's talks of responsibility. You know, there's an analog there to prior movements of ethics and social responsibility, of corporate social responsibility, things that want to keep the kind of corporate goal in mind, but wants to try to bring responsibility into the ambit. I often see this discussion of responsibility leveraged as a means to avoid regulation or paired with claims by companies that they could self-regulate efficiently and effectively. And I don't believe that's possible. I believe that you need to bring some kind of thing to account because when companies say we're developing or we are innovating responsibly, they are really forcing the public and regulators and investors to take it as an article of faith that they actually are doing the right thing. And we have no proof of that, right? Like if you have no artifacts, if you have no reporting, if you have no transparency, you really can't say you're doing anything. Responsible technology, to me, seems like a pretty big misnomer. I don't like this language of responsibility. I think it really doesn't hold anyone responsible. It has no one accountable other than the institutions themselves, which is not sufficient. We've got a cat on the scene. I know. She was on the table and knocked over my glasses. No, she just wanted to be heard. And do you think that we should be striving for something more than this concept of responsible AI? Where do you think we're headed on that front? Well, it's a question on what a corporation can realistically do, right? Whether they should use their language of responsibility is a little moot to me. It's more like, what are they actually doing? Are they actually reporting their data sets? Are they actually going to be responsive to this legislation? Are they going to be responsive to regulatory inquiries? Are they going to actually work in the public benefit? You know, the vast majority of corporations don't do that. They don't have the missions to. You know, there's some kind of structures such that they might try to do that, as in structuring as like a B Corp or having some kind of different kind of institutional structure. But most corporations are not oriented towards that. So, you know, I don't care what language they're using, really. But responsible is definitely one of those hand-wavy ones I despise. Right. And do you think then that the only thing that really can be done to address the likelihood of irresponsible tech being developed is through regulation? It's not the only one, but it is certainly one of the sticks that we have. Legislation is another. Having workers organizations that can collectively bring particular organizations to heal is helpful, including, you know, for instance, the writer's strike that happened in which the WGA and also the after strike and SAF, SAF, AFRA have at least put the movie studios and the streamers in check as a way of benefiting their workers. It is not going to be the market itself, though, that brings these organizations to heal and forces them to behave in a way that is in the public interest. You know, you have this really interesting perspective because you've worked both in industry at the height of AI tech development within Google and you've also you're currently working at DARE, which is a nonprofit research institute. How has that approach to developing, quote unquote, responsible AI has been unique in each of those spaces? Well, it's very much night and day, right, because at Google, the way it operates is that they're going to build products and then they bring in the responsible or ethical AI team to say, OK, how can we mitigate any harm that comes from this product? You know, but the kind of idea is to have an inevitability of bringing a product to market. And that's the starting premise. And so if you're starting from there, then already you're setting it for failure. I mean, it's sort of acknowledging that these things have the potential for harm and there needs to be some kind of mitigation for those. And so, you know, it's very common at Google to have teams coming to us in the 11th hour saying we need to launch. Can you tell us how to make sure that this doesn't completely go off the rails? And we're like, wow, you should have came to us a lot earlier and then we could have maybe intervened or designed this in a different way. And in some places, you know, we successfully were like, well, you shouldn't release it. You know, this is in no condition to release or the fundamental goal of this product would not be safe. You know, and there are some instances of that, but sometimes they just have not been thoroughly tested and they are put on the market and then they are revealed to do, you know, wild stuff. I mean, there's lots of examples of Google that happening. The AI overviews thing that's in search now is an example of that. The BARD example in which BARD said, you know, the Hubble Space Telescope had taken the first pictures of exoplanets when that was just a lie, complete fabrication. And so, yeah, there's plenty of examples of that. And there's a lot of examples at Meta, OpenAI, you know, etc. At D.A.R.E., we take a very, very, very different approach. We don't have to productize and if when we do focus on things that are products, they have evolved a much more hands on, you know, vision of what they are that are based in particular communities that are going to be using the products. And so a good example of that is, you know, one of our collaborators is Aswan Lashteka and he has a, you know, startup called Lisan.AI and Lisan provides machine translation and automatic speech recognition tools for people in the Horn of Africa, so Ethiopia and Eritrea. And, you know, the tools that are being developed are being done with, you know, compensating people who are giving their data and annotating those data, you know, paid at a good rate, consenting to what their voices or what their labor are being contributing to. And then is trying to do it for, you know, for a community that needs these tools in a way that's much better than what Facebook or Google can offer. So it starts with that community in mind. It starts with people from that community. I mean, Aswan is from that region. He's developing because there is a need that's community needed. And so, you know, we don't use this. We don't use the language of responsible AI at D.A.R.E. We don't use even necessarily the language of ethical AI. I mean, we use the language of, you know, community. We use the language of, you know, who are you accountable to? How is this going to work for this group of people? You know, what are the what are the tech visions or the tech futures that you all want that we need? And we start from there. I love that you're pointing this out because I find that this was the biggest realization for me in working on socially beneficial applications that a lot of the considerations that come in at the end with the ethics teams, especially in the context of for-profit projects, a lot of those reflections are sort of baked in naturally when you bring social scientists and users into the product life cycle. And you can almost mitigate them much more proactively, mitigate some of those AI ethics, responsible AI risks so much more proactively, but also much more meaningfully if you're working directly with that community to the point where steps around responsible AI and AI ethics might seem weirdly inapplicable, redundant, ineffective when you are working in that context, which is so close with end users and communities themselves. How do you think do you think it's important to continue reconciling those fields, like bringing in AI ethics? Is that language at all helpful to you or do you almost associate that with a corporate whitewashing of products that they're building and that the true socially beneficial AI doesn't need that type of reflection? I mean, I think more reflection, the better. I guess one way I'm going to choose to interpret this question is, you know, would it be helpful to bring kind of community based methods into the corporate setting? It's very common for companies to bring in folks and say, we want to have you at the table and consult with you in a community oriented way, but actually do nothing with that. And then, you know, use it as kind of a PR tactic. Is it helpful to bring it in? Well, again, it's something that's actually going to be actioned on. Otherwise, I don't think you should waste people's time. And do you think you can create a meaningful social intervention without that type of process? Should you be very wary of companies that are not looking to build very closely with the community? And part of the reason I say this is because so many companies are looking to scale, deploy AI products that are just applied to communities all over the world that have not been integrated into that process. I mean, having community involvement is a necessary but not sufficient condition to doing this. But actually listening to them, you know, and actually changing things gets you to sufficiency. You know, it's a power relation. If you're, you may need to actually cede power to actually, to actually, you know, engage in a meaningful community engagement process. We have actually gone through a three-day workshop on engaging with one very specific type of community. I think that's also an important takeaway for me is that engaging with communities is not necessarily going to look the same across communities. So having that type of training on how to not tokenize that very particular community that you're looking to work with, I think, is really important. But I wanted to actually transition and ask a couple of questions about one of your recent publications, which is titled, Do Datasets Have Politics? In this publication, you analyzed 113 different computer vision datasets. So these are datasets of images that train the models to, quote unquote, see things, whether, you know, it's a facial recognition, it's a visual recognition, it's a visual recognition, it's a visual recognition of a person, it's a visual recognition of a car, it's a facial recognition tool, being trained with images of people, or a dataset used to train autonomous cars with images of the road, for example. So these are datasets containing all kinds of different images. And this project that you had done was actually one of the first large-scale studies of these computer vision datasets. Now, you seem to focus a lot on datasets in your work. Why do you think it is important to study datasets? You describe in the paper high-quality dataset development as being one of the most undervalued components of the machine learning lifecycle. Why not only is it so important, but why do you think it is so undervalued, given its importance? Yeah, so I came to study datasets because that's the start of any machine learning, right? It's the data. And so many things are implicated with the data, whether that's privacy, whether that's representation, whether that's the labor that goes into actually annotating those data, or whether it is the rights of the individuals in those data, and also copyright, and whether one can use those data legally at all. This is a place where social science, I think, connects really well here, where in social science, if you're doing a PhD in sociology, you're actually expected to collect a lot of your own data. So whether that's doing a survey, or doing an ethnography, or doing interviews, some people, especially demographers, will focus on existing datasets. But the status of data feels like it's never really settled. And of course, it's not settled in computer science. The fact of data seems much different in computer science, where people say, well, I'm going to test this on ImageNet, or I'm going to test this on CIFAR-10, or I'm going to test this on MNIST, or whatever, you know, like those are datasets that are kind of rattled off, and there's widespread agreement that people understand what those are, and they're settled artifacts. And so that's basically why I came to focus on data and why we need to study data. In terms of why dataset development is undervalued, there's a few different hypotheses on this. One of these is, from studying this, I found that it tends to be that model work is valorized, and a clever method is much more important within AI. Nithya Sambhavasian, my former colleague at Google, has a great paper that's called Everybody Wants to Do the Model Work, No One Wants to Do the Data Work, which talks a lot about this. And she did a bunch of interviews with machine learning practitioners. Data work tends to be pretty monotonous, tends to be pretty rote, and it is not given much institutional credence. There's some ways that this is changing. Sudden Erupts has a data and benchmarks track now, which is helpful in a welcome condition. And there are some kind of moves to reward data work, but it is still nowhere near the kind of rewards one gets if they train a huge model and meet some benchmark score. So that's by and large the reason why it's devalued. Could you tell that it was devalued by analyzing these datasets? Yeah, yeah. I mean, we talked in the paper about, you know, just in terms of like how much of a paper like dedicates the description of the dataset. You know, you might be talking about a new dataset and dedicate two paragraphs to talking about the dataset in the paper. As part from very large datasets are very well known, like ImageNet, which has two full papers that talk about it. Most papers dedicate maybe two paragraphs. We went ahead and we literally counted that there were paragraphs dedicated to how much it was described. We also included, you know, websites and any kind of documentation on it. So that was one thing we really got very little detail about, you know, the collection, the time, the annotators, the datasets themselves were often unavailable. And in hindsight, it would be a five year old paper. The website would be down. It was not stored in any kind of institutional repository. You know, it wasn't stored in any of these places that assign a DOI. I think we only found one dataset out of the 113 that was in an institutional repository. And it was probably because it came from a psychologist rather than a computer vision or a machine learning person. And I think only three had DOIs. And can you talk just quickly about the relevance of that? Why is it important to be stored in an institutional repository or to have a DOI? Yeah, I mean, there's a few reasons for that. I mean, one of the reasons is that a DOI is a, you know, it's a digital object identifier. So you can find it. I mean, that's like the very base to find it, you know, otherwise, otherwise they're, you know, ephemeral or they're, you know, on a, you know, a lab website. That's also the kind of virtue of having an institutional repository. You know, even if you move jobs, I mean, it happens all the time. You know, professor goes from, you know, UMass Amherst and then goes to Google and then goes to CMU or whatever. These things need to live in a place that are persistent. And then the other part of it, which is very unappreciated, is the kind of access level. Not everyone should have the data, right? You need to guarantee that people are going to use the data in a way that's not going to abrogate the kind of initial intentions of that. The one dataset that was in the institutional repository is actually great. It was in something called the NYU Data Prairie and it was images of children's faces and facial expressions. And I asked for access to the data and they went, oh, you work at Google? Google doesn't have an institutional review board. We're just going to deny this request outright. And I'm like, that makes sense. Because effectively they needed to have some kind of IRB, you know, treating human subjects like human subjects. There's people in these data. IRBs are by no means perfect. They have lots of problems. But, you know, it is a floor at which institutions have such that human subjects are protected. Meanwhile, most of these datasets, I think two thirds of the datasets in the study, have images of people in them. None of them, you know, went through, I think very few, if one or two maybe went through an IRB process. Who is creating best practices in this space? Is it libraries who really, you know, are valued in the way that they manage, collect, categorize the data? Yeah, I mean, there's a huge literature on data librarianship and data stewardship. I mean, it's not hard to find and there are institutions that support data stewardship. You know, one of the leaders, it's ICPSR, which stands for the Inter-University Consortium of Political and Social Research. And they, you know, they've written a lot about data stewardship. And so these principles exist. Is it, you know, is it restricting the field? Maybe, I mean, to some degree. But I think that kind of ideology that LLMs and, you know, image and diffusion models have is, well, why do we need to be good data stewards? We're just collecting everything, right? And, well, OK, you're collecting everything, but you don't know what's in the data. I mean, and we've seen this, you know, we've seen this with Lyon 5b, where it just has, you know, has CSAM in it and they've taken it offline. That's the CSAM is a child sexual abuse material. And so, you know, they took, after an investigation from Stanford, they found that they took it offline. It also has a lot of very, has a lot of adult pornography in it, mostly depictions of, you know, people who are in, who are, you know, kind of very stereotypical kind of gendered modes. It's, and so Abebe Burhani has done a bit of analysis on this too. I'd be surprised if there wasn't just an incredible amount of violence and gore in it. There's all these things in these data sets. And I think the ideology is more like, well, we're just getting as much as we can. And because we're getting as much as we can, like, we don't really need to create these models or create these data. And, yeah, it is doing a disservice to whatever is happening. And I mean, language models and diffusion models have larger problems. But it's definitely, you know, not helping the case. You, you talk in the paper about some of the values that it seems like the data set developers have, and you're sort of pushing back in the paper against these values or against this culture. You talk specifically about the values of efficiency, universality, impartiality and model work. I'm wondering if you can help make sense of some of these terms and how you would like that culture to be shifting. Yeah, I mean, so we talk about four tradeoffs. So one of them is universality versus contextuality. And I think there's a big desire within computer science or rather within AI to build these everything machines, these things that sort of know everything. This is something my colleague and co-author Emily Bender talks a lot about, about like the building of everything or general purpose machines, which these things aren't, you know, they trained on a limited set of data. That data has limitations. If you're just saying you're getting as much as possible, you're going to be coded for people of the West, typically in English. And how do we actually make data sets which depend on context or depend on particular domains? We talk about positionality versus kind of like the view from nowhere effectively. It's effectively saying, well, you have a position, you can't say that you're coming from nowhere. And acknowledging that will really emphasize what you're not seeing. How do you connect that with your work, if at all, at D.A.R.E., where the focus is more on AI applications? And at D.A.R.E., it seems like you do have a pretty wide mandate. So how do you decide where to focus your attention and what seems like the most high impact use of your time? Yeah, I mean, that's a great question. I think connecting the stuff with the work at D.A.R.E., I mean, D.A.R.E., it's not restricted to AI. I mean, we have a pretty large understanding of what falls under how to study AI. And I know on this podcast, you talked to Crystal Kaufman, for instance, who is an annotator and a data worker and a data worker advocate. And I mean, I think a lot of that is, you know, we focus on working with data workers because there would be no AI if there was no data workers. Right. I mean, this is and they encounter intense, you know, intense working conditions and poor pay and everything like that. And so, yeah, like, why wouldn't you talk to a data worker, right? The same thing when it comes to people who are monitored by AI, whether that's Amazon workers, which is what my colleague Adrian Williams focuses on, or refugees, which is our other fellow, Mara Nestafanos, she focuses on, has been a refugee advocate for 20 years and knows that the borders are getting more and more militarized and more and more automated. And so, I mean, these are the places in which AI is pervasive and prevalent in our face. And so these are all related to what we're working on. You know, we're not oriented towards building LLMs or building diffusion models oriented because the AI that exists now is harming people intensely. How do we focus what to work is on? I mean, I try to see what the biggest burning fire is. And for me, you know, one of the things that Emily and I focus on in the podcast, in our book that's forthcoming next year, is the way in which AI has been hyped up, this hype bubble. How do we pop that bubble? How do we how do we really bring it to light and to do it in a way that is going to be accessible for people who are not AI researchers? I mean, it's coming out as a popular press book and lots of people have concerns about AI. I talked about this book on Facebook, you know, where I don't post at all. And, you know, I had a dozen people like, we need this book, you know, I hate talking about AI. It's just generating all this garbage. You know, how do we talk about it? You actually know about it. Like, explain this to me. So it's helpful. And I think, you know, to me, the kind of other thing that's big is thinking about labor, thinking about work and the way that AI is really affecting work and how labor is infused through the entire project, the process. So I think a lot of my work has been focusing on how to support other worker advocates, how to develop resources for them and how to partner with them. And that's also kind of top of mind. I have a lot of different ideas. A lot of people, people ask me, like, how do you like, how do you get ideas? I'm like, I got a lot of ideas. It's more of a prioritization process. And I think the things that are burning the house down are the hype and the labor aspects, which we're already seeing the impacts today, you know, this early on. I totally agree with you. I mean, it's part of the intro to this podcast as well, that there's just so much hype that needs to be cut through. Yeah. Is the harm that that causes very pervasive? And then how do you sort of categorize it in your head? Well, there's a lot of different types of harm, right? I mean, there's, you know, in terms of labor, for instance, I think that the harm is already here. It's that, you know, it's not that AI is going to replace your job. It's that your boss takes a chance and then you get fired. Or your boss thinks it can, and there's no position for you anymore. Or, you know, we've replaced human relationships with ones that are wholly algorithmically mediated. You know, and we see that in the medical field or in the field of law or in government services. And then the other thing is, you know, craft itself has been devalued, art has been devalued. And, you know, that is because art's been devalued, you know, like, how's that affecting artists? How do we think about art? So, I mean, I think the hype is just completely just taken over in a way that it's affecting how we work, how we play, how we get social services. And, you know, we need to really break that down if we're going to have any chance of really pushing back against this and seeing that the emperor, as it is open AI, has no clothes. And I'm sure it's very hard to do because people engage with chat GPT-40, for example, and they say, oh, I mean, it's incredibly efficient. This is the type of answer I was looking for. Actually, it can create something at a fraction of the cost of the artist. Is it about sort of showing that the technology isn't as powerful as we might think it is? Or is it about saying that actually, even with the extra cost, potential inconvenience associated with doing things in a more human centric way, the quality is higher, the value that it brings to the product or to your life is higher? Because I'm sure it's a very difficult argument to make that this hype isn't well founded. I don't think it's a difficult argument to make. No? Okay, good. No, I don't think so. If you talk to people outside of AI, they're like, I'm sick of this shit. Okay, good. I, yeah, I don't, I think only AI people are fascinated by AI, honestly. I mean, you find it differently. But I mean, you know, when I go, you know, I spoke at the, you know, Bay Area Book Festival, and I said, AI shouldn't do this. This is a human activity. And people are like, yeah, I agree with that. And then we have somebody from Google that was like, demonize the best things in sliced bread. And that's interesting. What are you talking about? Well, it's so interesting to me, because it's the computer scientists that should be able to cut through the hype most effectively, because they know exactly how these things are designed and some of the limitations of it. And yet they're the ones that are almost the most impressed by it. And it sort of manifests in this conversation about existential risk, existential crisis, AI safety. Do you see that as being a function of or symptom of AI hype? Or do you think there's a very genuine concern that this technology is putting us at risk by virtue of, you know, some of the harms you were describing earlier? Yes, and yes, it depends on who you ask. So existential risk was either a function of AI hype from people who still want to make a lot of money off of it, or they are true believers. And there's a set of true believers that I think help, you know, help that, right? I mean, you have your, you know, I'd say on one side, you have Eliza, you know, Yudkowsky, and, you know, you know, Dan Hendricks, and then being incredible hypers, or maybe Ilya Tuksever. And then on the other side, you have like your Sam Altman. You have people who are functionally making bank off this stuff, and you have to look at those interests, those needs, so people are ideologically motivated, right? And it's a bit of both, right? No one's, no one's kind of like universally motivated by one thing, right? Are you focusing on this area? Do you think it would be important, or like, what would have to happen for you to think that this is something to, that is, you know, burning house on fire, just as relevant as some of the other work you're doing? Existential risk? No, it's, it's, it's, it's, it's bullshit. Like, what it, what it, the people who are encountering existential risks are, you know, people in Sudanese refugee camps that can't get into Europe because there's biometrics at the border, or, you know, children in Palestine, or black and indigenous kids being torn from their families in Allegheny County, Pennsylvania. I mean, these people, people are having their lives ruined right now. Why would I pay attention to fantasy existential risk? I have, I have zero data on, on, on prior events, so, you know, even as a good empiricist, I should not pay attention to this. Well, this has been so fascinating. Thank you so much, Dr. Alex Hanna. So where can people find you and your work if they'd like to learn more? Sure. So you can follow me on Twitter at Alex Hanna, no H at the end. You can follow the podcast at dairinstitute.org slash M-A-I-H-T-3-K, which is very much a mouthful, but think Mystery, A-I-R, Hype Theater 3000, 3K. Or you can go to just the DAIR website, which is D-A-I-R hyphen institute.org, and then navigate from there. I'm on the other socials, I'm on Blue Sky, Macedon, but I don't post much there. Yeah. Amazing. And just before we close, is there anyone else whose work you'd recommend that folks take a look into? Yeah, totally. Dava Berhane, Crystal Kaufman, Adrian Williams, Rasecha Safala, Asma Nashtika, these are all folks that you can get to from the DAIR website, except Dava. I will give a shout out to Nina De Aura, a Brazilian AI researcher, and her profile says hacker, anti-racist, N-I-N-A-D-A-H-O-R-A dot dev. Yeah. Thank you again, Alex. And to our listener, if you'd like to ask us any questions, please don't hesitate to reach out at theworldwearebuilding at gmail.com. We'll see you in our next episode.