Name: Brainstorm AI 2023: A Look at Accountability and Best Practices
Uploaded: 2023-12-12
Duration: 18 min 45 s
Description: Stuart Armstrong, Co-founder and Chief Technology Officer, Aligned AI Tulsee Doshi, Director and Head of Product, Responsible AI, Google Brian Patrick Green, Director, Technology Ethics, Markkula Center for Applied Ethics, Santa Clara University Moderator: Rana el Kaliouby, Affectiva, Brainstorm AI

Stuart Armstrong, Co-founder and Chief Technology Officer, Aligned AI Tulsee Doshi, Director and Head of Product, Responsible AI, Google Brian Patrick Green, Director, Technology Ethics, Markkula Center for Applied Ethics, Santa Clara University Moderator: Rana el Kaliouby, Affectiva, Brainstorm AI

Brainstorm AI 2023: A Look at Accountability and Best Practices 

Conferences

Stuart Armstrong, Aligned AI Tulsee Doshi,  Google Brian, Patrick Green, Markkula Center for Applied Ethics, Santa Clara University, Rana el Kaliouby, Affectiva, Brainstorm AI

The fear surrounding A I extends beyond job replacement and productivity concerns. There are many who believe that in the rush to commercialize A I safety and ethics are not being prioritized appropriately. President Biden's executive order stated that responsible A I has the potential to help solve urgent challenges while making our world more prosperous, productive, innovative and secure. As we all know, implementing secure and responsible A I has proven a little bit more difficult than we thought here to discuss best practices for mitigating biases, ensuring safety and building responsible A I. Please welcome Stuart Armstrong, co founder and chief technology Officer at Aligned A I A start up focusing on teaching A I to align to human values, Tulsi Doshi, director and head of products for Responsible A I at Google where she has led more than 30 launches to make their products more inclusive. And Brian Patrick Green, director of the Techno of Technology Ethics at Marula Center for Applied Ethics at Santa Clara University. Brian is also the author of Space Ethics where he explored how ethics relate to space exploration. They're going to be interviewed by Fortune Brainstorm A I co chair and co founder and CEO of Afic PVA. Doctor Rana, uh L KUBI. Uh Before we dive right in, please watch this video from one of our partners checkpoints. All right. Thank you all for being here. Uh We're gonna be discussing how to make sure our A I is responsible and safe and Tulsi. I want to start with you exciting news from Google this week launching Gemini. Um Tell us a bit more about, I mean, Gemini is one of the first multi truly multimodal A I models out there. So tell us a bit more about that. And then how do we ensure that it's built in a way that mitigates data and algorithmic bias? Yeah, it's a great question. So yes, we are very excited about Gemini. And when we say multimodal, really what we're talking about is the ability to engage across modalities. So things text, image, video, audio, really being able to bring these efforts together, right? So think about being able to take an image and ask a question about that image and get an answer, right? And these combinations of things open us up to a wide variety of use cases that we're really excited about. And that has a lot of potential that we're really excited about, but introduces a bunch of new responsibility considerations. And so a lot of what we've been thinking about is we're building Gemini is how do we take everything that we have learned in responsibility and adapt it to this new structure in these new capabilities, right? So we released the A I principles, Google's A I principles five years ago now and those A I principles cover things like safety, fairness, transparency, security, and safety. And so now we're saying, well, what does that mean in the context of multimodal? And what does that mean at every stage of the development process? So how do we think about safety in our training data? And what does it mean to ensure that we have really high quality training data? How do we then translate that to the model? And what are the different use cases that might come up in the context of multi modality? So for example, you know, an image by itself might be a perfectly safe reasonable image text by itself might be perfectly safe. But now the combination of the two is actually offensive. And so how do we develop new metrics, new ways of evaluating that account for those concerns? And then also then the various kind of downs use cases and applications? And so for us, it's really been, how do we build those A I principles in how do we build the practice of the work that we've been doing every step of that way? And that's really what we've been working on and are going to continue to work on as we keep expanding Gemini Hours. It's amazing Stewart, you're on a mission to align A I algorithms with human values. Yes. What does that mean? And also are human values universal. Um Let's not get into that. The answer is universal enough for A I purposes. Um Well, it's just the outgrowth of, from the very first program. It's get the machine to do what you want it to do and not do what you don't want it to do. It sounds simple. But that's the failures of that cause all the disasters today. Nobody wants their algorithm to be biased. Nobody wants their algorithm to go crazy when it goes off distribution. Nobody wants adversarial images to break your classifier. Uh Everybody. And the thing is that alignment is a, is a capability. It's not just sort of a safety measure like 1/4 airbag uh on the roof or something. It's why are the consumer service bot so terrible. For instance, it's, you can have powerful ones, you could put a GP four behind it, but GP D four is not reliable enough and no, none of the models today are reliable that so you can't use them for that because then they'll start insulting the customers, this will blow up. So because this chatbot is not aligned with the designer's intent, you can't deploy anything but the most simplistic safe lockdown models. Brian, what concerns you as it relates to A I? There are also comment on like the intersection of space and A I ethics. Lot of time. There, there are lots and lots of things to be concerned about all the way from safety, all the way to what does this mean in terms of human self conception? Because if we externalize what we think makes us human, if we're, if we think that we're human because of our intelligence and then we externalize it from us and it becomes better than we are at thinking, then what does that mean about us? What's left with, with what human beings are? And we can see this in the way it's going to affect labor because people identify with their jobs. There are all sorts of these kind of effects. Once again running the full spectrum of anything that we can apply intelligence towards is going to be, you know, potentially an ethical issue related to that. And so if you want to talk about space, there are, there are plenty of issues related to that in terms of how is A I going to be used in the design process? How is it going to be used in, in how, you know, tracking debris around the earth or tracking satellites or coordinating large satellite constellations? There's, yeah, there's just a lot of things to think about. So you teach ethics and I'm curious, do you see young people interested in, you know, in issues related to ethics and inclusion? Yeah, this is, this has actually been really fascinating over the last four or five years, my class has gone from having just a few interested students, you know, maybe, maybe a half the class fill up, half way, all the way to having a long waiting list. So I think that people have recognized now that there's a lot happening. That is, I think that's great news for the world. So I wanna ask about um your views on Biden's executive order and the EU A I act. So to see what, what do you think about that? I mean, I'll start by saying like, I think it's amazing that we are having this conversation, right? Because we do need really good regulation. And I think that is actually a really critical part of making sure that we are building systems that are safe and consistent across the industry. And I think there's a lot to value in both the EU A I Act and in Biden's executive order in terms of how we think about high risk systems, in terms of the way. For example, the executive order is setting up an appropriate hub and spokes model that actually really takes into account all of the amazing work that organizations like Nist and other organizations have been doing for years to really build out and understand what risk means and how we develop strong systems. I think there's nuances now that we need to figure out, right? So does a dual use foundation model mean? And how does that definition apply across companies and organizations and across the industry? And so I think we're at a really good point to now be able to have those conversations, I think and really figure out what are, what does it mean when the devil gets into the details? I think I'll be taking audience questions soon. So tee up your questions. But Stewart, what do you think about um A I regulation in general and specifically the executive order? Um Well, there's an expression that you can't manage what you can't measure. And that's a big problem with this. So we want to get rid of bias. But what is bias? And that's one of the reasons that today we're launching a product for measuring uh gender bias is the first one and removing it from language models and other generative models. Can you talk a little bit more specific? Give us an example of how it would work. Um Well, I I deployed it in creating stories and it was quite interesting that female protagonists tend to have very passive roles. Uh male protagonists tend to have very active roles. So it was uh there was a princess who was beautiful, found a cuddly dragon who became her protector. There was a prince who found a fierce dragon but wasn't afraid and they became friends. Those sort of biases creep in all the time. We actually found that fiction has a lot more bias than say, even professional things you asked for top professions, for men or top professions for women. That's less biased than asking for a romance story. What does the protagonist do? But the problem is these large language models are trained on all of this data, right? And that's another problem is that you can't just tell them be less biased to some extent, you can tell to a human. You have very biased training data account for this, but you can't tell that to an A I because they don't understand, they don't generalize concepts in the same way we do. And that's another thing that we're working on to allow models to use concepts in the way humans do and present this to humans in a human, understandable way and say which of this approach is better than that? That's, that's another product we're launching today. We're launching lots of products today, Brian, what's your view on regulation? I think that there's more than enough out there to be regulated, that would be the way to look at it, right? There's so many things happening. But the real question is how to apply this regulation. So one of one of the things that I think of as an ethicist is that, you know, if everybody just made the right decision in the first place as an individual within their organization and the organization was had, you know, different ways to verify this or facilitate these good types of decisions. If the organizational structures were set up right within the organization, nothing bad would ever get out and then the government would never have to get involved with regulating it, but that's the world that we live in. So the question is, how do we set up a situation where the people who are closest to the technology and the technological development have the resources that they need in order to make the decisions that they need to make and do the best that they can knowing also that not every organization is going to be as good as some organizations and also some individuals are not going to be as good at this others. So we really need to have a structural solution. You know, in engineering ethics, there's something called the Swiss Cheese model where you line up the piece of the Swiss cheese and some of the holes line up in the problem, you know, get stopped by some layers, but it gets through to others. If you have enough layers, then the problem doesn't get through. But you really need to have the layers and if you don't set those up, then the problems will get through. I want to quickly follow up on that. So um in your classes, like do engineering and computer science students are they required to take your ethics classes. So they are required to take an ethics class. They don't necessarily take my ethics class. OK. That's good. OK. Questions. Do we have any questions? Questions here? All right. No questions yet. So let me follow up on the idea of how to align your organization to be bought into implementing ethics. How do you do that at Google? Yeah, it's a great question. I think there's a few different pieces you have to pull together if you really want to drive organizational cultural change. Right? I think the first is actually having a vision in a North Star for the organization to align on. And so at Google, the A I principles is that vision for us, right? Really setting a North Star with seven principles, we really want to make sure that our products hold themselves accountable to and then valuing that progress, right? Rewarding really great work that drives that work forward, measuring ourselves across that. We've established an A I Principles review process that actually reviews, we review all of these A I launches coming through to actually make sure they're meeting that criteria. Um And we're also building tools and resources to actually make it easier for teams to do this work, right? I think starting and really thinking about how do you measure bias? How do we provide the tools and the resources to help answer those questions? Because responsible A I isn't one size fits all and we don't want to make the mistake of saying, hey, there's a button you can press, right? We want to actually build that in to the design process every step of the way and that comes with building the tools and resources at every stage of that process and raising the right awareness, our teams for how they can actually do this work such that, you know, when they do go through a review, there's no surprises because you've actually been building it in, from the beginning. So for, for, for some of our audiences here who are like either ceos or leading A I efforts within their organizations, do you advise that there be a kind of an ethics team or how do you integrate it across the entire organization? Yeah, it's a good question and it's something we've, we always struggle with that polarity, right? Because on one hand, you want building systems to be everybody's job and it should be everybody's job. On the other hand, you also need to have individuals who really have the space and the time to dive into the hard questions and build out the right expertise and the right best practice. And so that's kind of why we've built a combination of both. So we have central teams like mine that really focus on what are the gnarly, hard questions that are coming out in responsible A I, what are the new challenges we need to understand? How do we answer questions? Like how do you measure bias? And then we work across teams where we have individuals embedded across the organization who are thinking about what does inclusion mean in the context of youtube or what does it mean in the context of search and we partner directly with those teams to kind of share insights back and forth, to make sure we're learning. What problems are they seeing in the products that we should be thinking about and driving more research, what are insights that we have that we can bring back into the product? And I think honestly, that is a very much a setup I would very much start with, right, eventually, as we build more and more state and more and more insight, I often joke that I'll work myself out of a job because we really will have built responsibility into the entire organization. But there's so many new questions that are coming up that I think there's a, there's a certain amount of work to do before we get there, Stewart. How about you? How do you work with organizations through aligned A I? Well, there is a lot of truth in what you're saying and you do want to sort of get in the habit of, of bringing ethics. But we're also, and we have our organization set up so that it requires unanimity amongst the shareholders to pivot it away from AAA I alignment organization. So that's never going to happen. Um But we want, we don't want safety and ethics and alignment to be a separate thing. We think that if your product is unsafe, you have failed, you have not built a successful A I, it's, I think Stuart Russell's expression was we don't have one group of engineers who build bridges and a second one that goes around and points out that they might fall down. So you should change that there. You just build bridges. So you want to build A I that is successful, that is align, alignment is a capability. So the, we want everyone to be doing it because that's their job. If they, if their A I is racist, they have failed. This is um so that's the there is, of course, you're going to have to have some people looking over the ethics and other parts of the team as you describe it. But the fundamental thing that we want is that people understand alignment is success. This is a capability. Yeah, I think actually, you know, building on that, we've been talking a lot about how if you really wanna build bold successful A I, you have to build responsible A I and I think the more we talk about responsibility, the same way we talk about quality, the same way we talk about accuracy, right? These are building a responsible product is a better user experience. Uh And so you actually are building a better experience for your users. And so actually baking it into the core fundamentals of how you build the product, I think is actually a really critical point. And I think it's more of a question of how do you instill that? How do you give teams the resources to do that work such that it does start becoming a part of that in the same way. And just to build on that, that's what the Marla Center tries to do with many of our materials on our website are just free materials we're putting out there to give people the tools that they need in order to make the good ethical choices that they already want to make. So we have, we published a book that June, which is called Ethics in an Age of disruptive technologies and operational road map. And it goes through five steps, it starts off and it's set up like a road trip. First, you decide you want to go. Second, you figure out where you are right now. Third, you figure out what your destination is. Fourth, you figure out what you need in order to get from here to there. And fifth, you make sure you're tracking your progress. So, and all of these things and it's all from an ethical perspective, right? To make sure that you're thinking about how to really implement these, these tools, these ways of thinking into the organizational structure to make sure that you have everything that you need. Yeah, this is great. Well, thank you so much. My takeaway from this panel is that we're at a time where implementing responsible A I and Safe A I, it's actionable, it's very doable and there's a way to implement that across your entire organization. Thank you. Thank you.