Camille Isnard-Will Fleisher AI Interview

The podcast features Camille Isnard interviewing Dr. Will Fleisher, a philosophy professor at Georgetown University, about AI ethics. They discuss interdisciplinary collaborations between philosophy, computer science, and cognitive science in addressing AI ethical concerns. The conversation delves into the challenges of regulating AI, the role of democratic oversight, and the concept of ethics washing in AI companies. Dr. Flasher emphasizes the importance of participatory design and effective regulatory bodies for ethical AI development. The discussion highlights the need for genuine ethical considerations beyond surface-level metrics in AI systems. Hello everyone and welcome to today's podcast. I am Camille Isnard, a junior at the Georgetown University College of Arts and Sciences, double majoring in economics and psychology, recording today on Thursday, March 19th in Washington DC. Today I will be sitting down with Dr. Will Fleisher, who is an assistant professor of philosophy at Georgetown and research assistant professor at the Center for Digital Ethics, whose work focuses on the ethical and epistemic implications of AI. Some of his most cited papers include algorithmic fairness criteria as evidence and understanding idealization unexplainable AI. We're going to dive into some deep questions with him so let's get into it. Hello Dr. Will Fleisher, how are you today? I'm doing well, and you? I'm doing great, thank you so much for giving me the time and joining us in this conversation. I wanted to start off by asking you, I've read some of your work and it seems like you've published in both philosophy journals and more technical journals discussing the subject of AI. I was wondering if you can talk a little bit about how you find that internees are talking to each other and kind of working together on this issue or if they're kind of deviating, especially as the speed of this technology advances and surpasses the ethical issue. Yeah, this is a very good question. It's actually complicated. So the philosophers have been thinking for a very long time about AI ethics issues that are sort of sci-fi AI ethics issues. And some of those have turned into no longer being fictional. Some of them I think are very much still fictional, like AI consciousness or things like that. But the philosophers have been thinking about this for a long time, about the relationship between intelligence and cognition and computer science. And there's been cross-pollination with cognitive science for a very long time. Between computer science, cognitive science is an interdisciplinary field. They have a philosophy, cognitive science, linguistics, psychology, neuroscience. And so there's been that research that's where people have been talking to each other for a long time. Some of these venues you mentioned are interdisciplinary computer science style conference proceedings. And so these are very much like computer science style publication venues. But they are places where a huge number of people from computer science, philosophy, anthropology, sociology, STF, a whole bunch of people are talking to each other. And it's been a really fruitful, I think, robust interaction, not without its problems, where a ton of people are having conversations about bias, about accountability and transparency for contemporary technology. So this is another way philosophers have been in the mix of having influence on these much more recent conversations. And these are huge populations of researchers. But computer science is enormous here. It's just massive compared to all the other ones. So the percentage of computer scientists that are taking part in these conversations is not enormous. So there is a sense in which although we're having really fruitful research relationships and conversations across disciplines, including computer scientists, nonetheless there's a huge field of computer science that's not as influenced by concerns about ethics, about epistemology, philosophy, science, or about research in psychology. And do you notice a disparity between the amount of people that are involved in discussing these ethical concerns that are capable of taking action to fight those concerns versus those that can only kind of talk about it? Because a lot of computer scientists aren't as involved, and I guess they're kind of at the edge of the hydraulics. Yeah. They're the ones building things, right? So they have enormous influence over them. And they're also the experts that get asked about it for policy purposes and what it is. And they're not at least concerned about things other than profit and effective technology. Yeah. That's definitely a key issue right now. You mentioned Encephalic, and I was actually going to ask you about it. So obviously there's a lot of news right now about Encephalic and the White House. And I guess the main issue is just about how to regulate AI, which it's not surprising that regulation, federal regulation on AI is kind of behind, because I guess it's always a little bit behind technology, and this is advancing so quickly. And so what I'm seeing is a lot of companies have kind of taken it upon themselves, like Encephalic, to build their own ethical constitutions. And it's kind of made us, we don't make us, what we think is most ethical. So it kind of takes it upon itself to do that. But I was wondering what you think, in an idealized world, would be the best way to kind of go about this issue of regulation. Do you think the solution is like a broad federal regulation that the EU kind of is starting to point out, or, because there's also so much friction with like the competition and the race with China. So I don't know how to, like, answer this. I mean, these are really hard questions. In an ideal, in a really ideal world, we would have local communities with effective means for regulating technology. You know, where democratic citizens can come together and say, these are the kinds of things that we want to see with respect to, you know, controls over how the technology functions. With respect to how our kids are going to interact with it. With respect to how our data is going to be used for it. In an ideal world, like, there would be real democratic oversight, where people are actually getting to make decisions about how the systems are designed. But we don't live in a world where that's easy, because the corporations who are making these decisions are multinational corporations, who operate in a lot of different places. Local communities have a limited ability to regulate them. And even large-scale national and multinational bodies have difficulty regulating them. And part of the reason you're pointing out, I mean, China and Russia obviously have very different governments than the EU. And for now, it's the US. And so, you know, the effective regulation is hard. But I do think that right now the best model we have for effective regulation is probably the EU. I'm not a lawyer or a policymaker. I'm a little bit better with the questions about what should be than the questions about what isn't. And there are complicated questions and intersections about how to get effective regulation. But I do think what we want, from a perspective of what's morally required, what would we want as a just society, we want democratic control and oversight over when we make decisions. And there's different ways to implement that. So I think participatory design, where the design of the systems are being influenced by people from, like citizens from society, from a variety of different backgrounds. I think that's really important. I think having effective regulatory bodies that are appointed, you know, they're real civil service units that are overseen by elected officials and appointed officials. It's really important. So the EPA, you know, as it was 10 years ago maybe, has applied to AI technology, for instance, something like that. We have a civil service of experts, lawyers, economists, environmental scientists, in this case it would be lawyers, economists, and computer scientists, and hopefully philosophers and humans, thinking about what kind of regulations would be appropriate. That's what I want to see in the world that we have, that seems possible. And then, you know, there are going to be trade-offs when it comes to, you know, technological races with other societies. But I think that we need to make, we can make those trade-offs in a more democratic way than they are right now, right? NASA won the space race. But that's a government organization, using government funding with democratic oversight. Like forming a World Bank, but for technology, or for AI. Yeah, I think that would be fine. You know, it's not going to be a panacea, it's not going to solve all the problems, but that seems like a better world than the one we have right now. Right now, as you point out, it's the companies themselves making decisions, self-regulation, and that's what we work with as well. Yeah. So you mentioned, I don't know if you said participatory inclusion? Yeah, participatory design. Okay, participatory design. That kind of reminded me of one of the papers you wrote on algorithmic fairness criteria as evidence. So I wanted to mention it because, obviously, we're living in a world where algorithms are forming more and more decisions that set your future forward, which obviously has really strong implications, especially if some groups are disadvantaged because of that. And you mentioned Compact, which was used by judges in Florida, who were more likely of committing a crime. But then you mentioned the news management, which I thought was really interesting, because how they kind of attempted to resolve the issue of more blacks being flagged as way more whites, but that's actually not the goal in the show. And so that kind of made me ask the question, do you think that is the way in which AI companies are kind of faced with an easy way out to appear more ethical by kind of like, how do I say this, setting up their metrics in a more ethical way, when in reality the structures behind the systems are still discriminatory, and the issues are still very much there? I think that's exactly what's going on in a lot of cases. So this is known as ethics washing. So it's a play on greenwashing, which is making something seem green or pinkwashing, like in the original. But anyway, it's the idea that it takes up to make things seem more ethical. And in fair machine learning, there's been a large number of articles written arguing that the whole practice is aimed more at ethics washing than it is at actually improving circumstances for people. And yes, I do consider this a serious worry, that when companies talk about fairness, or when they talk about inclusion, or when they talk about... I mean, if they're just doing it for advertising and spend purposes, it's unlikely that they're going to be doing something really substantive. Now, that's not always the case. Sometimes public opinion forces companies to make more ethical decisions, but again, I think insofar as a company can make more money by pretending to be ethical rather than actually being ethical, they're going to take the ladder up. This is why we need regulation. The capitalism and the way that corporations work is designed then to aim at profit. For better or for worse, that's how they work. And so we just need to recognize that that's the incentive structure for them. And so expecting companies to self-regulate is just a false assumption. I'm curious, and I think you know more about this, but how do you actually measure whether a company is ethical? Because in my head, it's more of these metrics are used to measure whether a company is ethical. And so if they are capable of manipulating those metrics, how do you know if they truly are fixing this structure? Yeah, I think the thing is that ethics is extremely complicated in general. It might not be measurable. That is, when we look at all the values that we think are real values that we really care about, different people's well-being, different people's autonomy, justice for people across society, if we look at this set of values, there might not be a reasonable sense in which you can measure ethics as such. People have tried for many years. When you collaborate with computer scientists about their machine learning stuff, they often become frustrated that they like to measure things, that's how they roll. And look, measurement is very valuable. But when you're trying to trade off the interests of, say, teenagers having autonomy with respect to all kinds of technology and engagement, versus society as a whole's mental well-being, it's really hard to figure out how to assign particular number values to this state of affairs where students get to choose what they want to do but they're more depressed, versus this state of affairs over here where there are autonomous constraints but the rates of mental well-being are higher. How do you put a literal number on that? How do you put a literal number on that? It's actually incredible, and that's part of the reason why we need a research that involves people from different backgrounds actually looking carefully at what a company's doing and holding them accountable. I kind of wanted to talk about, I think this kind of relates to what we're talking about, but in your other paper of understanding idealization and explainable AI, you talk about XAI, which is asking the model why it got the output it got and how to understand it behind the scenes, but you did write this in 2022, and I'm very curious how this applies today, just because when I get a response from Chuck D or Claude or whatnot, I can fully ask the model why did you give me this, and it'll give me a very convincing explanation and it might seem like it's fully explaining it behind the scenes, but I'm wondering if that's actually the rationalization issue that you mentioned, and whether we're actually getting further away from the truth behind the scenes, or close to zero. Yeah, very good question. The short version is, I think, that method of just asking the model what it's up to is taking this further away. There really is a rationalization, and we have real reasons to doubt that what it's telling us about itself is reliable. Several reasons. One is that in real sense, it doesn't have direct access to what it's doing insofar as it knows what it's doing, and I put carefully, insofar as the model is able to say that it's what it's doing, it's not because of a sort of robust ability for introspection. If they have some ability for introspection, again, scare quotes, but they have some sensitivity to their own internal workings, we don't know how that would work. There's a few recent studies suggesting there's some limited sensitivity that they have to their own internal workings. We should not think that they have some special insight into why they produce the output that they just produced. What you do when you ask it a question, you're doing the same thing as when you ask the original question. I ask, okay, what should I have for dinner tonight? Chat, and it says, oh, well, you should have pizza tonight. I'm like, why is she telling me that? Oh, well, I know that you like pizza, or I know in general people like pizza, or whatever. In both cases, it's doing the exact same thing. It's doing sophisticated output. Very sophisticated, and the fact that you can do network prediction and guess the results of guess is shocking and amazing, and it requires a lot more research. But it's doing the same thing in both cases. In both cases, what it's responding to is it's training data, and it's training on the data, the patterns that it can tell, and what you ask it. So just asking the system is a very dangerous way to try to do this, because it's not reliable at telling the answer, and it's very confusing. But there are methods people are working on as a kind of scientific inquiry, trying to do scientific models of the complicated answers. So I think that research is very worth doing. It's very important, but it's at a very early date, and we still know very little about what's going on. And sometimes anthropic over self will be hard, because they're excited about it, and because it benefits them to suggest that they know more about what we do. But there is real work of real value there, so we're hoping to do that. I was wondering in that same paper, if you could talk a little bit about idealization. You mentioned how it's very important to have it, and how science has worked in the past with idealization. Do you think idealization is different in the context of AI, because we're not exactly sure about what we're unsure about, or what we're leaving out of the model, or what can we say about that? That's a very good question. Idealization is when you have a model of a phenomenon, and the model has parts that are false, a sort of misrepresentation, but typically they're intentional, simplified. Building a model that's easier to work with, that we understand better, that tracks the particular causal patterns of the phenomenon we're interested in right now. When we build a mechanistic model, or some other kind of model of the AI system, do we know which parts are idealization? I think the answer to that is complicated. Sometimes the answer is yes. We know when we're making choices about how they seem to align. So, in the case of LINE, which is a blank example on a paper, which is a very simple old SAI method, we know that we're only paying attention to the local area. That it's only decisions that are like the decision we're asking for an explanation for. We know that we are using linear models rather than non-linear ones, even though the underlying model, if it's a deep neural network, will be non-linear, at least in the way that they tend to be implemented today. So there are differences. And while we don't know how to build a non-idealized model, which is, on some false response view, like the goal, we should be able to figure out how to de-idealize the model. Just like the ideal gap law, we know how to de-idealize it. We know the Van der Waals equation is like de-idealizing. We don't have that with AI. Well, that's actually similar to cutting-edge science. There's a lot of idealized models that we know are idealized and we can point to which parts of them are likely to be de-idealization. We don't know how to get rid of de-idealization. I need to have closer to the table more data with explainable and interpretable AI models. Nonetheless, it doesn't mean that the model isn't granting us some genuine understanding. When you were in New China Mechanics in high school physics class, you came to understand the world a lot better than you did before. You and you both. That's even though New China Mechanics was false. So I think that that's the kind of situation. For some of the things we know are idealizations, we just don't know how to get rid of them. In other cases, there probably are things about the models these methods are producing, which are idealizations and we don't realize it, and we need to be better off if we do. First of all, in my class, we have discussed a lot about ethical concerns about surveillance capitalism and how much data these companies are collecting. And how they present it as it's necessary to provide the services they provide and these improved experiences. But as I see it, if I take the example of Google, for example, a major improved experience for me in Google has been the general response that I'm faced with when I do Google search. It's so much more efficient and it's what I look for when I do a Google search now. You have to know how to get a response better than having to look through all the links. I'm curious, does Google need to collect all my data to give me that response? I don't think it does. But then, we mentioned how every Google search I do provides this excess data to these companies that can then become a product that they then sell to other companies. I was wondering what do you think about how to resolve this issue? Do you think it's more about these companies being entirely transparent about how they're collecting the data and how they're using it as a product? Or, we discussed this in class and a lot of people said that would actually just cause so much more chaos and maybe not actually change anything. Is the actual solution to regulate these companies and then dump them somewhere or another? This is a very difficult set of questions, too. But, good ones. First of all, I watch out because Gemini they're not always reliable. But, these are real transparency issues because we don't know what degree of Google's performance are a result of their data collection. There is good reason to think, and you'll find this in the Surveillance Capitalism book, which I suspect you're reading part of? No. Part of Google's competitive advantage from early on was that they provided more relevant results. And, before they launched into the advertising business, that was part of what made them the marquee search system. And so, it's possible that some amount of tracking people's behavior is important to them. Now, which parts? We don't know. We don't know which parts are most important. It's not transparent. It's a trade secret. I kind of doubt that without some kind of regulatory structure that we would be in a position to know which parts are crucial and which parts are just things to think about. And, I do suspect that mere transparency isn't adequate to get genuine informed consent. As far as what's really necessary to get those benefits, it's really hard to know. It's really hard to know. I genuinely don't. I don't know if people outside of Google know. Okay, so to wrap it all up, I am going to ask a very open-ended question that you can answer however you like. It's kind of depressing, but it is important to address. I was wondering what you think is the biggest threat that AI poses to society, and how you think we could start to deal with all that. Yeah, I mean, I'll give you two. It's too hard to narrow it down beyond that. There's a lot of threats. It's not robot apocalypse. It's not that. The biggest threats are I think the effects on our politics and the effects on the environment. So, AI is there's a lot of ways in which varying kinds of AI technologies have significant effects on the information people get or what they believe about the world, on their ability to evaluate factual claims, on how they perceive other people in society, and make them form their values. So, AI is used to make decisions on social media feeds, on what shows up on the news, on what shows up in Google search, and of course now very influentially with what you find out when you have Chats with the Tea or Claude or Gemini, questions about the real world. And environmental effects of AI is also a very significant thing that I worry about. Yeah, that's the first one. It's very scary. I feel like we all have such a different version of reality, and in my social media feed, I see the same three issues and I think everyone else is, but in reality, it's just because it knows I'm interested in that. And everyone has this perception of reality, which is so different. Alright, well that should finish it. Thank you so much. Thank you.