{"url": "https://www.youtube.com/watch?v=ilKPYH94JSg", "page_url": "https://www.youtube.com/watch?v=ilKPYH94JSg", "title": "Plenary Panel: Inequality, Labor, Health, and Ethics in AI + Audience Q&A | AI Now 2016", "show_or_channel": "AI Now Institute", "date": "2016-07-25", "duration_s": 4428, "type": "panel", "speakers": ["Mustafa Suleyman", "Alondra Nelson", "Latanya Sweeney", "Lucy Suchman"], "confidence": "high", "notes": "AI Now 2016 symposium plenary panel on inequality, labor, health and ethics, with audience Q&A.", "people": ["mustafa-suleyman"], "id": "2016-07-25_ai-now-institute-plenary-panel-inequality-labor-health", "audio_duration_s": 4428.335625, "worker": "jc-w05", "transcribe_s": 55.6, "model": "large-v3", "segments": [{"start": 1.49, "end": 10.16, "text": "Thank you, Julia, and thank you all. So you've heard a lot, but we have more. It's now time for our plenary panel of the night."}, {"start": 10.16, "end": 39.47, "text": "As Genevieve alluded to, we spent the day in an expert's workshop. This was 100 people from diverse occupations focused on AI's potential impact on the four key themes, social inequality, health, labor, and ethics. So tonight's panel represents expertise in each one of these domains, distilling insights from the workshop, and we hope sparking a discussion among you and among those on the live stream. So join us on Twitter. We have the AI Now hashtag."}, {"start": 39.47, "end": 53.2, "text": "You can tweet out your questions, and we're going to be selecting from among them and addressing them to our panel afterward. At this point, I'd like to introduce Alondra Nelson, Dean of Social Science at Columbia and author of the book The Social Life of DNA."}, {"start": 60.43, "end": 81.52, "text": "Good evening, everyone. I thank you for coming. Thanks, Meredith, for that introduction. And thanks to you and Kate and our colleagues at NYU and the White House for staging this really important conversation and one that Nicole Wong reminded us a few moments ago is really a quite urgent one. It's an important conversation about a social"}, {"start": 81.52, "end": 107.63, "text": "I'm honored to chair a panel of influential experts and of people I greatly admire in the areas of health, labor, social inequality and ethics. So please help me welcome Mustafa Suleiman."}, {"start": 107.63, "end": 134.0, "text": "Co-founder of Google DeepMind and director of DeepMind Health. He's at the forefront of real-world challenges of AI, particularly in the health domain. And while these efforts and his AlphaGo work gets all the press, Mustafa has also been working on issues of social inequality and ethics in relation to AI for quite a while and is a leader in this. Next, Henry Hsu."}, {"start": 136.46, "end": 159.94, "text": "He's a labor economist based at the University of British Columbia at the Vancouver School of Economics. He's the author of the widely circulated report, Jobless Recoveries, and his research focuses on automation and the decline of middle-class jobs. Professor Su. Next, I'll introduce LaTanya Sweeney."}, {"start": 159.94, "end": 190.62, "text": "Professor of Government and Technology and Residence at Harvard University, whose early work in data discrimination has galvanized many researchers and a broader public debate about the potential disparate impacts of data systems. Professor Sweeney, thank you. And next on the stage is Professor Lucy Suchman, a Professor of Anthropology of Science and Technology at the University of Lancaster."}, {"start": 190.62, "end": 214.83, "text": "For over 30 years, she has researched the interface of human and machines. She was a pioneer at Xerox PARC 20 years ago, and she's been awarded the Benjamin Franklin Medal in Computer and Cognitive Sciences. Thank you. So welcome, everyone. So we're going to begin with some short four-minute..."}, {"start": 214.83, "end": 254.38, "text": "Talks by each of the panelists. Then we'll have a conversation and Q&A among us that I will lead. And then we'll have a larger discussion and Q&A that will be drawn from the questions that we'll receive on Twitter. And I'll remind you, as Meredith did, that it's hashtag AI now if you want to send in questions. So we'll hear first from Mustafa. So six years ago, I co-founded a company called DeepMind. And our mission..."}, {"start": 254.38, "end": 280.82, "text": "was to build general purpose learning algorithms and use those algorithms to make the world a better place. And 18 months or so ago, just after we were acquired by Google and were tasked with running Google's general purpose AI effort, we decided that now was the right time in the history of the company for us to start to make good on the second part of our mission to figure out how we could use our algorithms to make the world a better place."}, {"start": 280.82, "end": 306.56, "text": "And so it was in that vein that we launched DeepMind Health. So today I'm just going to tell you a little bit in the next three or four minutes about what our ambition is with DeepMind Health and what we think we're responding to. So I guess to begin with, I guess we're sort of framing the question with respect to the inequality that we see in health and acknowledging the fact that technology hasn't been doing a great job of..."}, {"start": 306.56, "end": 320.21, "text": "Improving productivity and efficiency in the healthcare system. Here's a graph of healthcare spending in the UK. And what we see is that despite the fact that input, the green line, namely cost..."}, {"start": 320.21, "end": 342.54, "text": "Thank you very much."}, {"start": 342.54, "end": 371.82, "text": "We're really struggling to identify the correct associations between cause and effect. How do we intervene in a complex system and feel confident that the intervention we make, whether it's in an A&E department or whether it's trying to reduce diabetes in outpatients, how do we feel confident that those interventions are going to yield the kind of benefits and results that we want them to? And I think we feel hopeful that maybe advanced applied artificial intelligence might help us to learn some..."}, {"start": 371.82, "end": 396.32, "text": "There's something about the nature and structure of these new data environments that might aid us in doing a better job of navigating through some of this complexity. One of the areas that we've been looking at most recently is in using our AI to help us tackle some challenges in healthcare. We know that one in three people in the UK are clinically obese."}, {"start": 396.32, "end": 424.7, "text": "People in the US are clinically obese and that diabetes is significantly on the rise. One of the really remarkable things is that if you do have diabetes, you're actually 25 times more likely to suffer some kind of sight loss. But the most extreme and damaging versions of that, of severe sight loss, can actually be completely prevented if we intervene early enough. But the difficulty is that we just don't have enough radiologists."}, {"start": 424.7, "end": 449.54, "text": "We need 10 times more radiologists to help us to diagnose these conditions than we actually have. And so this represents an opportunity, I think, for machine learning algorithms to help us to do a better job of classifying the onset of diabetic retinopathy in retina scans. So here you see two examples of a 2D fundus image and on the right-hand side, a three-dimensional..."}, {"start": 449.54, "end": 478.51, "text": "And of course it's a distributed and scalable technology that we can take way beyond the Western developed and privileged world."}, {"start": 478.8, "end": 505.81, "text": "So this is very early work, but it's an insight into some of the very practical things that we hope to do with advanced applied AI and how we might be able to use it to serve that second part of our mission. But of course, new technologies need us to completely rethink the way that we do governance and ethics. And if we leave these components behind, then I think we're really going to struggle to succeed in delivering more effective applications of AI."}, {"start": 505.81, "end": 535.09, "text": "I want to say that we really need to be as innovative on the governance as we have been in the last five or six years on the tech. And in many ways, healthcare applications we know need to be very rigorously scrutinized. People's healthcare data is an incredibly sensitive source of personal and private information. And so we need very new models of oversight, both on the regulatory side, but also very new practices that companies and research labs use to..."}, {"start": 535.09, "end": 548.99, "text": "I think trust is a function."}, {"start": 548.99, "end": 565.5, "text": "Transparency and Security"}, {"start": 565.5, "end": 593.57, "text": "Prove in a verifiable way that your particular personal health data has been edited and accessed in only a very specific way by people who you have mandated. And if we were able to do that in a provably transparent, untamperable public log that verifies who has had access to your data and for what purpose, I think this would go some way towards helping to build trust and to verify that your data has been used as you would wish."}, {"start": 593.57, "end": 620.13, "text": "I think another aspect is a governance mechanism that we have been experimenting with over the last few months at DeepMind Health, which is to have a panel of independent reviewers who oversee our technology. So we've appointed nine independent reviewers who are experts in the clinical, the technical, the policy, legal, and ethical aspects of the clinical work that we're doing, and we've given them access to..."}, {"start": 620.13, "end": 646.14, "text": "Our data sharing agreements, our commercial agreements, our product roadmap. We've given them access to my team, so they can interview any member of my health team and ask them questions. They can investigate our technical infrastructures to verify and check that we are doing what we say we're doing publicly. And these are entirely independent, so they're not contracted, they're not paid, and their remit is to scrutinize and review us in the public interest."}, {"start": 646.14, "end": 682.16, "text": "And so far, this seems to be working reasonably well, but it's early days, and I think there's much more that we can do. So I just wanted to give you a little insight into one of the models that we're experimenting with that we hope has more potential to be successful in the future. Thank you. Henry. So good evening. Thanks for having me here. So as an economist, the discussion of AI that grabs my attention is the anxiety around its impact on employment."}, {"start": 682.16, "end": 711.1, "text": "You know, are we headed towards a world without jobs? So for some perspective, I want you to think back to the turn of the century when we saw an automotive innovation that was a real game changer. It had a radically new fuel source, more power, making for travel at ludicrous speeds, and it brought the potential for a new way of driving. And so some of you may already have guessed that the turn of the century that I'm talking about is actually the turn of the 20th century."}, {"start": 711.1, "end": 741.01, "text": "And the innovation was the internal combustion engine, which, as we know, eventually replaced the horse. And so this example, which is courtesy of historian Gregory Clark, I think it gives us three important takeaways. Number one, the combustion engine, it brought huge efficiency gains in the way that we power equipment. It created new jobs in automaking and far beyond. And it made humanity, you know..."}, {"start": 741.01, "end": 772.06, "text": "Much better off. Point number two is that its effects on employment were uneven. So while the vast majority benefited, some workers were made worse off by this innovation. And in this case, it's the poor horse, which was largely eliminated from the workplace. And important takeaway number three is that people are not horses. So now let me take you to a more recent example. I study..."}, {"start": 772.06, "end": 801.89, "text": "The impact of automation on today's workplace. And in the past 30 years this has been driven by advances in industrial robotics and computing technology. And the effects here too have been uneven. Automation has disproportionately affected routine occupations. Those are occupations that perform a narrow set of rule-based and repetitive tasks. And routine occupations"}, {"start": 801.89, "end": 830.48, "text": "As you may know, have been disappearing over time. Obvious examples include machine operators on the factory floor, travel agents, and administrative assistants. And the timing of this disappearance has actually changed the nature of recessions. So most of you are aware of the economy's struggles since the end of the financial crisis, which ended in 2009, and the fact that we're just now exiting a prolonged jobless recovery."}, {"start": 830.48, "end": 858.34, "text": "Income and profits have bounced back, but jobs have been slow to return. And my work with Nir Haimovich at USC finds that this is part of a longer term pattern. So this picture shows employment in routine jobs since 1967. And the ups and downs, the booms and busts are very obvious because routine jobs bear the brunt."}, {"start": 858.34, "end": 884.82, "text": "And so if you look in the 70s and early 1980s, job losses and recessions were quickly recovered. But now look at the three most recent episodes. Routine jobs were lost in the recessions, but unlike before they never came back. So automation helps explain jobless recoveries and where we are today."}, {"start": 885.17, "end": 913.87, "text": "And as you are probably aware, routine jobs are by and large middle class jobs. So automation is a prime factor, actually it's the primary factor in the decline of the middle class. Many of these workers have had a tough time finding new work. And it contributes to the falling labor force participation rates of working age adults that we've seen in the past 15, 20 years. Okay, so now let's come back to the fact that people aren't horses. People aren't put out to pasture."}, {"start": 913.9, "end": 942.38, "text": "And people in routine jobs can change jobs. And not surprisingly, their success in finding new work depends on their ability to retrain and get more education. And so in the last 30 seconds, I just want to kind of look forward a little bit and ask, you know, will AI look like this? It's obviously hard to predict, but its effects will likely move away from routine jobs and push towards more complex jobs in fields like journalism,"}, {"start": 943.09, "end": 970.8, "text": "And as Mustafa said, it feels like medicine. These are just two examples. But what's true about complex jobs, though, is that they typically involve a richer variety of tasks. And many of those tasks, AI in the near term cannot do. And what's more, workers in complex jobs tend to have more education, which will help them adapt, transfer their skills, and find new ways to capitalize on these technologies."}, {"start": 970.8, "end": 1003.89, "text": "So that's all I have to say for now, and I'll turn it back over to Alondra. Thank you very much, Henry. We'll hear next from Professor Latanya Sweeney. I am so not sure why I have to always be the heavy. I heard over and over again tonight about the beauties and the promises of technologies, and yes, I did hear about economic concerns, but I want to address something more than that."}, {"start": 1003.89, "end": 1033.39, "text": "In particular, that the effects aren't just on the economy, they're on every societal value that we have. You know, in my life, I have had two pursuits academically. One is AI, and the other one is harmonizing technology and society. And tonight, I get to bring those two together. When I applied to graduate school at MIT in computer science, my application talked about a passion that I had to build a thinking machine."}, {"start": 1033.39, "end": 1059.07, "text": "A machine that could think and learn like humans. And I got off to a great start. I built a machine that learned the sounds of the system the way a child would or a baby would. And so if you spoke English, you would learn the English phoneme system. If you spoke Korean, you would learn the Korean phoneme system. And it promised to really revolutionize speech recognition. Then one day, my life took a dramatic turn."}, {"start": 1059.07, "end": 1076.11, "text": "Beverly Woodward was an ethicist at Brandeis University, and she was visiting MIT's AI lab. I'm walking past, and I hear her say, computers are evil. That was crazy talk. So I had to stop and straighten out her thinking."}, {"start": 1076.11, "end": 1101.79, "text": "She then introduced this notion. She felt that we were in a new era, that computers had brought us into a new place. The year was 1997, and that personal data was being shared without knowledge and could only lead to personal harms. She had an example of hospital data that had been collected and given away on government employees, judge, policeman, and so forth."}, {"start": 1101.79, "end": 1127.89, "text": "And it didn't have their names, but it did have their demographics and some medical information. And she says, look, this kind of data isn't even anonymous, and therefore computers are evil. So I needed an experiment to prove her wrong. So William Weld was the governor of Massachusetts at that time. He had collapsed. He had gone to the hospital. Clearly his information was in there. Could I re-identify him? Was it really true?"}, {"start": 1127.89, "end": 1147.65, "text": "This position that Beverly took. I went and I bought the Cambridge voter list for $20 and it had the same demographics and it included his name. And so by linking these two, I could put his name back to that record. The question is how unique would that linking be?"}, {"start": 1147.65, "end": 1169.44, "text": "This is an excerpt from the Cambridge voter list. Six people had his birth date. Three of them were men, and only he had that same zip code. And this had a profound effect both on health care policy at that time. It shaped HIPAA and many other privacy policies world round because"}, {"start": 1169.44, "end": 1196.75, "text": "Demographics of most people are also unique. I sort of started out that day trying to fix Beverly Woodward's thinking, but in the end, she ended up implanting in me a lifelong challenge that has led to a lot of groundbreaking work on data privacy. And you know, we're at that moment again, a moment, a sort of watershed moment, but this time it's not about data, it's about AI. In 2003,"}, {"start": 1196.75, "end": 1222.16, "text": "I read news articles about the death of AI and I could not imagine what they were doing because you see AI is that pursuit of building something that's like you. It's humans and you want to build yourself and you can't really kill that. And so I wanted to demonstrate what were people actually calling AI. So I surveyed thousands of AI papers and the leading books on AI."}, {"start": 1222.16, "end": 1249.73, "text": "And I found two definitions. One was people inspiring their work to build human-like machines, but that kind of pursuit was hard and very uncertain. And then there was a new group emerging, and that group followed a mathematical ideal. That is, they would take a task that humans do now, and they would find a mathematical precision alternative that would therefore..."}, {"start": 1249.73, "end": 1279.42, "text": "They promised to be better at humans at doing that task. And once achieved, they would no longer consider that AI because now a machine could do it. It is this notion of AI that we talk about today. It got new life because of the boom in data that Beverly Woodward talked about. And that boom also brought new statistical algorithms. So then, what then are the unforeseen consequences? How could it possibly go wrong?"}, {"start": 1279.42, "end": 1304.78, "text": "So let me tell you about the day my computer was racist. Adam Tanner is a very well-known reporter, and he was interviewing me, and he was in my office, and we needed to find a particular publication. I Googled my name, and up popped these arrest ads. And so I show him the paper, the link to the paper. He says, oh, forget that. Tell me about that arrest."}, {"start": 1304.91, "end": 1333.57, "text": "And I said, well, that's impossible. I've never been arrested. And we poked around more and more, trying to find out what could possibly be going on. And as we did, we started seeing a pattern. And Adam, who is a tall, white, Italian-American, jumps to this radical conclusion that, oh, those ads only come up because you got one of those black-sounding names. And I said, that's ridiculous. That's not how computers work."}, {"start": 1333.57, "end": 1356.3, "text": "But in reality, the more I tried to show him that he was wrong, the more I couldn't, the more problems I kept finding. The first finding was, in fact, names are associated. Oh, the slide didn't work. Names are more associated with one race than another."}, {"start": 1357.52, "end": 1379.86, "text": "And that was really significant that names given more to babies of one race than another can have these discriminating effects. Eventually, though, I had to break down and do a complete experiment. I ended up collecting 140,000 ads, and the ads only showed up with the names of real people. To show Adam I was right, I had to actually..."}, {"start": 1379.86, "end": 1402.14, "text": "Buy a subscription to the service and show that not only did I not have an arrest record, but no one with my name had one either. It turned out that the ads that were suggestive of arrest came up in 80% of the searches of names that were given most often to black babies than white babies, even if the company had no real arrest record for the name."}, {"start": 1402.14, "end": 1430.18, "text": "The findings were statistically significant, and they also passed the adverse impact test for civil rights investigations. So we say, how did this possibly occur? Well, Google is delivering these algorithms. It's a really simple algorithm, actually, the entire network. Advertisers provide multiple ad copies and a search term and the amount of money they'll pay if someone clicks on the ad. So at first, the algorithm starts."}, {"start": 1430.18, "end": 1458.9, "text": "With a random selection of which of these ads it's going to deliver. But over time, when people click on an ad, that ad becomes weighted. And so over time, the more weighted ad shows up more often. So could this actually be the advertiser who had not actually put an even distribution of text down? Or was this the reflection of society making votes in a sense by what they clicked or chose not to click?"}, {"start": 1459.98, "end": 1490.13, "text": "Throughout my career and also when I was the chief technology officer at the Federal Trade Commission, I have found that a single experiment, a single scientific fact can empower those to help us do their job better. Data, technology, and now AIs are all making it harder for our societal norms and those who provide those societal norms to actually be preserved."}, {"start": 1490.13, "end": 1519.17, "text": "So leaving the FTC, I joined up with about 50 other researchers from around the world, and we've been doing this thing called tech science, where we teach students, others, other people, those in government, and so forth, to do a single experiment, and that single experiment gives you a fact. But the problem that we have with AI is the bias can exist. It can exist by the people who invent it. It can exist by the people who might be using it. It can exist..."}, {"start": 1519.17, "end": 1545.02, "text": "Thank you."}, {"start": 1545.02, "end": 1573.33, "text": "We had good reason to believe that there was bias and you could start a litigation to gain access during the discovery process of the algorithms and data to use it for testing. Judicial decisions now make that very almost impossible to do. They want you to prove the harm before the lawsuit can continue. So the most important thing that we can do in this watershed moment is actually engage and empower ourselves."}, {"start": 1573.33, "end": 1601.79, "text": "with ways that we can test and assess AIs. No one AI, there's no one societal norm, there's no one way to do these assessments. But we do things like safety in cars and products, so kind of consumer reports and lab testing. Maybe we need something like that for some AIs. We already require manufacturers of some kinds of goods to provide warranties and compliance statements. Maybe we need something like that for some AIs."}, {"start": 1601.79, "end": 1632.61, "text": "But in general, we need to be able to allow researchers and others to perform bias assessments on AI and introduce new technologies that can detect and thwart that bias. On those online ads, I showed Google a way that they could easily detect those kinds of bias in the ad network that they currently use. They chose not to do so. Oops, another slide missing. Anyway."}, {"start": 1632.61, "end": 1671.65, "text": "So we're now 20 years later from my original conversation with Beverly Woodward. And computers are not evil. We just need to scientifically harmonize them with society. Thank you. Okay. Well, I'm very happy to follow that wonderful contribution. My opening provocation for a discussion, let's see if I can give it to you, is that..."}, {"start": 1671.65, "end": 1691.57, "text": "The proposition that a central ethical challenge for AI now is the field's demystification. So I'm going to give you a sense for what I mean by mystification, and I'm going to illustrate that with an example. And as you've gathered by now, there are multiple AIs now."}, {"start": 1691.57, "end": 1720.85, "text": "From data analytics to humanoid robots. But one of the continuities across those, and particularly in their public representations, is a language that is saturated with anthropomorphisms. So beginning with the trope of artificial intelligence itself, going back to the giant electronic brains of the 1950s through today's..."}, {"start": 1720.85, "end": 1746.94, "text": "Natural language understanding, affective or emotional computing, and very prominently in our discussions here, machine learning. And so I want us to kind of pause for a moment and take that language seriously and think about what it's doing. Now, one rationale for using these analogies is that they translate"}, {"start": 1746.94, "end": 1774.72, "text": "Complex technologies into terms that people can more readily understand. But anthropomorphic terms are also used suggestively and frequently in ways that muddy the difference between actual technical achievements and technologists' aspirations for the technologies. And in that sense, these ways of talking about AI obscure more than they reveal."}, {"start": 1774.72, "end": 1794.59, "text": "So let me make this more concrete with the case of care. Now we regularly hear these days that we face a crisis in elder care, that the increasing numbers of us who are old, I count myself among those heading in that direction, who are old and in need of care."}, {"start": 1794.59, "end": 1824.5, "text": "Coincides with a critical shortage in the availability or the economic feasibility of human caregiving. So this is taken as two demographic, incontrovertible demographic facts. We have an increasing population of aging people. We have a decreasing population of human caregivers. This defines the problem and it also constrains the solution. And according to this logic then, the solution has to be a technological one."}, {"start": 1824.5, "end": 1850.69, "text": "That's how it works. Now, both of the images that you're looking at here were found through an open-ended Google image search on elder care. The image on the left illustrates a blog post on the implementation of the UK's Care Act of 2014. The image on the right accompanies an article titled, Would You Trust a Robot to Take Care of Your Grandpa?"}, {"start": 1850.69, "end": 1874.69, "text": "In which the OpenRobo Ethics Initiative publishes its 2014 findings regarding public attitudes towards robot caregivers. But note that asking the question conveys the idea that robot caregivers are a possibility. And I'll come back to that. So if we look at these images more closely."}, {"start": 1874.69, "end": 1897.89, "text": "It should raise questions for us regarding what care, so in this case giving medication, what does this actually entail? So we could start by looking at the bodily positioning of the caregivers in each case in relation to those who are receiving the medication. So the multiple ways in which the caregiver on the left"}, {"start": 1897.89, "end": 1927.3, "text": "Orients to the woman to whom she's handing a pill through her leaning in to the appropriate distance, her monitoring of the woman's face and her encouraging smile, the alignment of the hand holding the pill in relation to the hand that feels it to pick it up, the glass of water that's waiting for the moment when it will be needed, and those are just to mention the most obvious. On the right, we might note the substitution of a pill bottle for the individual pill."}, {"start": 1927.3, "end": 1946.75, "text": "The reach of the recipient to meet the robot's grasp. The question of how and why this PR2 mobile robot, which is a development platform that's widely adopted in university robotics labs, is imagined as the right tool for this job."}, {"start": 1946.75, "end": 1972.0, "text": "And we might think about what understanding of medication, medication as both substance and medication as a practice of caregiving, what understanding of that does each of these caregivers have and what further labors of care might they each be able to provide in the next frame after this one that we're looking at."}, {"start": 1972.0, "end": 1994.38, "text": "So before we turn to the question of the acceptability of robot caregivers or public trust in them, I would argue that there's a prior question, which is what grounds are there to believe that a robot could engage in the work of care? Clarifying the technology's limits might then open up public discussion of alternative ways."}, {"start": 1994.38, "end": 2016.43, "text": "of addressing the problems of elder care like creative new living arrangements or revalorization of the labor of care so that more people would be drawn to that work. There would certainly be a place for new technologies in such a future but it would I think be a very different one from a robot substitute for human caregivers. Thanks."}, {"start": 2024.27, "end": 2053.87, "text": "Hey, so now the fun part, conversation, put you in conversation together. Thank you for those presentations, which were all quite remarkable in their ability to do a lot of deep thinking and work in four minutes or so. So it's quite impressive. So I wanted to begin with a question putting Mustafa and Latanya in conversation. You know, Mustafa, your presentation was quite persuasive and this vision that you have for what we've been doing wrong in healthcare and how we might use."}, {"start": 2053.87, "end": 2083.09, "text": "to do better, do a better job and to save lives. And recently, you've been allowed to use NHS data and the UK health records to work on this towards this effort. And Natanya, your early work is also about health records in some regard. And I think so putting your presentations together, I guess I wanted to ask both of you to help us think about the balance that one might have to strike between privacy and better preventative health care."}, {"start": 2083.09, "end": 2099.26, "text": "Is that a balance that we can strike? What are the costs and benefits of using electronic healthcare records and machine learning and data processing and these things in this moment for healthcare?"}, {"start": 2099.26, "end": 2128.66, "text": "There are obviously enormous sensitivities and challenges with respect to what data we have access to, for how long do we have access to it, and what can we actually do with that. And I think one of the difficulties is that there's always a challenge in responding to the existing regulatory framework when you're trying to push the envelope and develop new technologies that are sometimes not always what was imagined or considered by those who initially"}, {"start": 2128.66, "end": 2141.86, "text": "I think the prospect of being able to do really useful things and demonstrate utility is something that"}, {"start": 2141.86, "end": 2161.36, "text": "Both drives us and I think will also encourage lots of other people to recognize that there are practical benefits and that there's an immediate necessity which is driving our collective interest in this field. But at the same time, as I said in my talk, we sort of have to be as innovative on the governance framework as we are."}, {"start": 2161.36, "end": 2190.67, "text": "And that's why we've taken steps to set up an independent panel of reviewers that have an unprecedented amount of access and oversight and influence in how we're handling such sensitive data. And I think there's so much more that we can do as well. So we're convening a patient engagement forum which will meet four times a year that's open to anybody in the UK who wants to participate. It will be streamed live on YouTube as well. They can ask any question. It can shape our design. It can encourage us to pursue different"}, {"start": 2190.67, "end": 2215.12, "text": "Different directions, they can express concerns, give us feedback. And in parallel, we also have a clinical engagement forum as well, trying to get the feedback from nurses and doctors and actually have our development process driven by what they tell us they need. So I think they're just some of the things that we're trying to do to think proactively about how we innovate on the governance in this space."}, {"start": 2216.59, "end": 2229.33, "text": "It's really clear that you are not in the United States. Because in the United States, you just get the data. I mean, we don't have the same privacy regime."}, {"start": 2229.33, "end": 2258.61, "text": "That the UK has. We don't have a comprehensive privacy. We all look to HIPAA as whenever we talk about medical data. But HIPAA only controls some aspects of medical data. To the point that the AIs are used in the care of the patient. We all want to live longer. I want my grandmother to do better. So there are no real major restrictions of access to data in the care of the patient. And outside of the care of the patient,"}, {"start": 2258.61, "end": 2284.11, "text": "We started a project called thedatamap.org that has thousands of people and companies and business analytic groups and all of these places who are able to get data because HIPAA isn't really a kind of comprehensive privacy law. So you could probably jump start a lot faster if you moved to the United States."}, {"start": 2284.11, "end": 2311.49, "text": "So I have a question for Henry and Lucy, but before we move on, I wanted to come back to Mustafa and Latanya around the question of re-identification. So what your project showed from a decade ago and certainly what AI today allows quite easily actually is re-identification of data. And so how should we be thinking about that with regard to trying to harness the power of these big data sets of healthcare records, but also that..."}, {"start": 2311.49, "end": 2341.3, "text": "Very real possibility that you can figure out quite readily who people are. Well, that actually is the goal of many AIs. I mean, the truth of the matter is business analytic companies is one of the fastest growing areas. And what those companies are doing is they get every scrap of data they can on every individual or as many individuals or as targeted individuals and build together profiles that they then make into products that they sell. So that actually is the goal. The goal is..."}, {"start": 2341.3, "end": 2367.44, "text": "What in privacy we would call re-identification, in their world they call linkage. And it's really interesting when we talk about it in privacy, everyone gets all concerned around a simple re-identification experiment I might do. On the other hand, linkages are happening all the time in a hidden way, and it's not very transparent. And I think this parallel to AI is just spot on."}, {"start": 2367.44, "end": 2391.66, "text": "And that is that the technology moves so fast and our policies move so slow. And this mismatch is not only just temporal, but also knowledge. And so as a result, if you have to chase it, then 10 years from now, AIs will have come and gone, just like data privacy has sort of in many ways come and gone."}, {"start": 2391.66, "end": 2409.7, "text": "So you're absolutely right. I agree with this. And I think the opportunity or our response should be to empower the patient with control of their own information and make it the patient's responsibility and within the patient's control. Because after all, all of these things are a question of control."}, {"start": 2409.7, "end": 2433.94, "text": "I mean, the more control an individual has to provide or withdraw a particular use of their data, be it for medical research purposes or any other purpose that they choose, the more empowered I think they will feel and in control of their own destiny. So I think that will be one step on the way to sort of..."}, {"start": 2433.94, "end": 2454.93, "text": "Limiting, I think, some of the sort of more cynical potential interpretations of risk that I think that we all fear and that none of us really want. I think the most important thing is that we give the opportunity for our patients to really have control of their data again, and I think from there we have a little bit more hope. I just have to say, there you go being a UK guy again."}, {"start": 2454.93, "end": 2480.02, "text": "I mean, you do not have control of your medical data in the United States. Whether you sign the HIPAA document or not, the data goes. The regulation already has decided that the data will be shared. And we had to fight, when I was on the federal HIT policy committee, I had to fight really hard along with others that individuals could actually even get a copy of their data."}, {"start": 2480.02, "end": 2507.76, "text": "Before that, before high-tech, and before we forced, in many ways, the conversation with ONC, we basically, you couldn't even go and get a copy of your data. Now it's kind of annoying that you have to pay for it, but that's another story. Yeah, well, I mean, the UK health system costs 50%, or half of what it does in the UK, and outcomes are significantly better, and coverage is universal. Now he's bragging."}, {"start": 2509.9, "end": 2530.37, "text": "OK, so let's talk a little bit about work and labor. So Henry, you've done a series of influential papers that have identified the implications of AI and automation for the employment sector. And much like Jason Furman this morning and David Edelman a little bit this evening, you"}, {"start": 2530.37, "end": 2559.7, "text": "Show us, even in your talk as well, that we're left with this sort of trough of lovely and lousy jobs, as you call them, and sort of scooping out the middle, these middle-class routinized jobs. And Lucy, you are a pioneer of human-computer interaction, and you've been actually thinking about automation and humans and computers for a very long time, and labor is certainly one of the themes about what you were presenting tonight. So I want you to talk, I guess,"}, {"start": 2559.7, "end": 2588.22, "text": "How can we be more purposeful and deliberate in making a sustainable AI that fosters workers and employment rather than making them obsolete? That is a very good question, a very difficult question. First I just wanted to say that the phrase lousy and lovely jobs, that's not from us. It's from a wonderful paper by"}, {"start": 2588.22, "end": 2616.21, "text": "Martin Goose and Alan Manning from a few years back. But what is true about the type of technological advance and automation that we've seen in the recent past, in the past 30 years, it has had this effect of disproportionately hitting the middle class and leaving very good jobs at the top and not so good jobs at the bottom end. This isn't the..."}, {"start": 2616.21, "end": 2644.7, "text": "Pattern that you see every time that there is a major technological innovation. As an example, if you think about the assembly line that was kind of revolutionized manufacturing back at the turn of the 20th century, what that did was it allowed cars to be made by a small group of skilled artisans who would build the car from start to finish."}, {"start": 2644.7, "end": 2667.79, "text": "at obviously very high costs. It allowed auto manufacturing to be done by low-skilled individuals. I mean, we're talking about the immigrant communities of the Midwest, of Polish descent, of Italian descent. And what that really did was it allowed immigrants who otherwise would have been marginalized in the economy to really..."}, {"start": 2667.79, "end": 2695.9, "text": "Fill in the American middle class. So it's not the case that technological innovation will always affect certain people. It's just that the most recent wave has. And if we look forward, I think, I don't want to pretend to be a guru or anything, but I do think that as we look forward into the future in terms of what AI will do, it will start affecting more high-paying jobs in the ways that I described."}, {"start": 2695.9, "end": 2725.18, "text": "But because high-paying jobs tend to do a variety of tasks, the effects will be more subtle. So if I could give an example, we were talking earlier today about the possibility that AI will allow me as a professor to not spend so much time in office hours with students diagnosing where their shortcomings are in terms of the course material."}, {"start": 2725.18, "end": 2745.3, "text": "Our technology will be able to go through their homework assignments, go through the midterms, and using that information, piece together specific areas of weakness. And then with that information, it will allow me in office hours to do what human beings are good at, which is interacting with them, trying to figure out, okay, if this concept is difficult for you the way that I taught it."}, {"start": 2745.3, "end": 2772.05, "text": "In the lecture or the way that you read about in the textbook, what are the other ways that we can address it to make it more understandable? And so there's a real sense in which technology in terms of what we're seeing coming forward, it could be a real boost to employment, what it is that we do and how it is that we do things in ways that are hard to imagine."}, {"start": 2772.24, "end": 2800.56, "text": "Yeah, I mean, I guess I think we need to think about automation as a political question. I mean, there is a kind of, as I was suggesting, a kind of discourse of economic necessity that I think obscures the degree to which these are social and political value propositions."}, {"start": 2800.56, "end": 2829.42, "text": "The skill levels of jobs gets defined is not an objective characteristic of the jobs. It is an effect of an ordering. So that's why I think care work is such a good example. I mean, we could think of care work as routine, but it's actually, I mean, one of the things that I think is interesting about the failures of robotics is"}, {"start": 2829.42, "end": 2849.98, "text": "is that one of the most difficult domains has turned out to be housework care. Because those are eminently, they're very heterogeneous, they're very open-ended, very contingent, they are of course also interactional. So the fact that we value care work,"}, {"start": 2849.98, "end": 2879.55, "text": "In the United States particularly, care work, teaching, particularly teaching at the primary and secondary school level. The fact that you get paid a huge amount of money to write code and you get paid nothing to take care of people's children is not a reflection of the relative skills on some objective scale, but rather it's a reflection of the valuation"}, {"start": 2879.55, "end": 2908.54, "text": "That we make of those different jobs within a particular political economy. So I think we, it's not that of course I think computerization has an enormous part in the relief of certain kinds of tedium. There are all sorts of appropriate places for automation, but I think we have to really slow down the presumptions of"}, {"start": 2908.54, "end": 2938.67, "text": "Labor shortages and ask, are there really not enough people to do those jobs or are those jobs systematically devalued in such a way that they don't provide people with a sustainable living? Yeah, I just wanted to add to that. I mean, so I agree completely with you, Lucy. I mean, if you look at how primary school teachers are paid in Canada relative to other occupations and compare that to the United States,"}, {"start": 2938.67, "end": 2968.14, "text": "It's very different. And so this is, I think, a statement about, potentially about social values. And I think what we should think about in terms of AI is it's a technological advance. It's a tool. And we've been creating tools forever. And different tools affect different people in different ways. And when that happens, as a society, it's our job to say, wait, this tool has the potential for amplifying"}, {"start": 2968.14, "end": 2998.13, "text": "Social injustice or amplifying these differences in social value that perhaps now, perhaps before we get there, we should be stopping and thinking and talking about those things and making decisions about what we value, making sure that our government codifies those values in a reasonable way. And I guess that's why we're here tonight, too. Thank you both. So if I had to do my quick classificatory system of your presentations, I would say, you know,"}, {"start": 2998.13, "end": 3026.77, "text": "Optimist, neutral diagnostician, and pessimist, you know, pop the hype bubble, you know, for LaTanya and Lucy, perhaps. I'm the only computer scientist. What are you talking about? Yes, I know computers aren't evil, but you have a quite critical, you're clear about that, but you have a quite critical perspective. So, I guess I wanted to ask you from your different..."}, {"start": 3026.77, "end": 3043.33, "text": "Perspectives, if you accept my classification. What does a good AI look like? I mean, so what if, as Nicole suggested earlier, AI is here, you know, the horse is out of the gate, whatever, choose your metaphor. I guess we'll use horses tonight, little horses."}, {"start": 3043.33, "end": 3072.05, "text": "If we're stuck with these AIs that we have and the sort of new version that's more algorithmic and machine learning, what do we want the best AI to look like? What should it look like? And how can we make that come into being? And also, what would a just AI look like? What's an unbiased, fair AI that supports the flourishing of most of us, many of us?"}, {"start": 3072.24, "end": 3100.3, "text": "For all of you, any of you want to take that up? So, I mean, I think the first thing to say is that we should try to avoid talking about AI as though it's a person or that it has status just like, you know, a human does. I mean, I think we have a tendency to project that into these systems. We have to remember these are machine learning systems which we design, which we control, which we direct, and they're problems that we point them at."}, {"start": 3100.3, "end": 3129.58, "text": "We're in control of that. So I would say that a just or a good machine learning system is one that we have control over, that is reasonably transparent in that we can see the training data that goes into those systems, that we keep attempting to do a better job of understanding the models that have been learned and looking at the classifications and the predictions and the recommendations that are generated, and that the benefits of these systems as we develop them over the coming..."}, {"start": 3129.58, "end": 3158.13, "text": "Thank you very much."}, {"start": 3160.43, "end": 3188.24, "text": "Yeah, so I think in terms of the perspective I think I'm supposed to bring in terms of employment effects, I don't think that you can imagine a just AI in terms of its effects of employment because it is a technological progress. And as I've said, technological progress has uneven employment effects. But what I think we need to require and make sure at a minimum is that these machine learning algorithms do not exacerbate."}, {"start": 3188.24, "end": 3217.14, "text": "We have good evidence that human beings are biased. We discriminate. We do so in the workplace. There's very good economic studies where there are applications sent out to the same job interview with different ethnicities of names, and you see very different results in terms of the success rate of getting an interview. So human beings..."}, {"start": 3217.52, "end": 3245.97, "text": "Aren't good at some of these things and we want to make sure that as we develop these technologies that they are better than us at doing that and you know speaking you know hearing from from the workshop this morning it seems to me an important piece of You know machine learning is the ability to get access to training data That is essentially the key input into developing"}, {"start": 3246.61, "end": 3275.41, "text": "And the implementation of this technology. And so to the extent that that training data is an important input, it's an important public good to society, then we should really be pushing for public funding of these things in the same way that we did so with the genome project. It should not necessarily be in the hands of individuals or corporations."}, {"start": 3275.41, "end": 3308.77, "text": "To the exclusion of others, especially in the realm of making sure that there is no bias or discrimination in the way that we use these technologies because obviously the more people get to do it, the more chances you get to discover these difficulties, these problems. So I think pushing for more public money towards these ends would be a good thing. Hey, I want my thinking machine."}, {"start": 3308.77, "end": 3336.53, "text": "I want everything technology can do. I want the new tomorrow. I want people to live forever. I want us to have a fantastic technical today, tomorrow, and the future. I want everything that I could squeeze and get every success out of technology. The problem isn't the technology side. The problem is it's out of pace with society and that we live in a technocracy."}, {"start": 3336.53, "end": 3360.0, "text": "Basically, an arbitrary decision made by a designer in Silicon Valley dictates the policies that we live our lives by. We don't know their names, we didn't vote for them, but yet what they allow the technology to do or not do determines what values we really have. And the people who would normally protect us are not able to do so."}, {"start": 3360.0, "end": 3373.47, "text": "And if we don't fix that problem, we're going to be left with wherever the technology takes us. And that's what we need to change. We need to provide mechanisms for oversight."}, {"start": 3373.47, "end": 3392.24, "text": "Thousands of my students have gone to work to Silicon Valley. We didn't teach them about how to design in for a societal impact, how to do those kinds of risk assessments. Even if we did, the final product that comes to market is not just a technology. It's a technology within a business package."}, {"start": 3392.24, "end": 3410.75, "text": "They come together. Somebody's making money on it. And all of a sudden, when it's gotten that hardened, you can't separate. Like at the FTC, it was often impossible to separate the arbitrary business package from the technology itself. It all came together by that point."}, {"start": 3410.75, "end": 3427.66, "text": "So we really do need mechanisms that will, and which AI it is and which social norm it is determines what that mechanism looks like. But in the end of the day, it's always going to be a form of access and accountability and transparency."}, {"start": 3433.49, "end": 3452.18, "text": "Yeah, well I would I would definitely follow up on LaTanya's pointing out the conjoining of I mean we have artificial intelligence as a field that was you know initiated in the middle of the last century and it's conjoined with its funding."}, {"start": 3452.18, "end": 3479.15, "text": "Institutions, largely government, overwhelmingly military in the United States and also industry. So I think we have to understand AI as part of that history and that set of relations. I'd actually like to see the field of artificial intelligence and its technologies completely re-describe themselves."}, {"start": 3479.34, "end": 3505.84, "text": "So we could have a kind of temporary prohibition on the use of analogies to human, and just see what kind of re-descriptions could come out. I mean, so for example, machine learning would be described as some kind of increasingly sophisticated forms of the iterative adjustment of parameters in a system. That rolls off the tongue, Lucy."}, {"start": 3505.84, "end": 3535.65, "text": "And then we might have to ask some questions so we understand better what that means, but I think we would have a much clearer understanding of what's actually going on in the field than referring to it as machine learning, which invokes all of the associations that we have with human learning, and I think in a way that can be very, very misguiding. And I guess the other thing for me is that the projects would be"}, {"start": 3535.65, "end": 3558.24, "text": "embedded in, they would be done in relation to the practitioners in different arenas for whom those technologies should become part of their worlds, part of their lives, part of their work. And it would be a joint project of figuring out kind of"}, {"start": 3558.24, "end": 3585.79, "text": "What made sense and what would actually be useful? So I could imagine fabulous kind of design projects around elder care, and they might involve some forms of robotics. You know, there are unquestionably really valuable places for data analytics. But, you know, it would be worked out in context and in relation to, it would be informed by."}, {"start": 3585.79, "end": 3611.15, "text": "The knowledge and skilled practices of those into whose lives these technologies were going to become a part. So it would be the whole organization of the project of making the technology would be radically transformed. And it would be really interesting to see what came out of that."}, {"start": 3611.15, "end": 3640.27, "text": "Thank you. So we've got some questions that came in via Twitter. I want to get to as many of these as we can. So in the interest of time, I just asked one or two of you to respond to each of them. So this first question, how do we deal with the basic facts that we often don't know how deep learning systems work? They are not the usual algorithms we can check."}, {"start": 3640.27, "end": 3654.13, "text": "This is the million dollar question, right? This is super difficult. It's a major area of research for us. And I think one of the things that we should think about is to distinguish the set of applications that we're content."}, {"start": 3654.13, "end": 3683.47, "text": "Using models that are not interrogable, not transparent. And I think there are some that we are already using in production today, like image search or speech recognition that we appear to be reasonably comfortable with using models that we can't parse so well. And then there are other applications that I think carry a greater risk and consequence that I think we need to be much more careful about. But this is a massive area of research for us."}, {"start": 3683.47, "end": 3717.1, "text": "I'm confident that we can make significant progress towards this over the next few years. The second question is a little bit similar, so let me read this one and then you can weigh in. Shall we as a society outlaw opaque machine learning networks if we can't examine them? AI abolitionism. Let me just make a small comment by analogy. Oftentimes in epidemiology and medicine,"}, {"start": 3717.1, "end": 3745.44, "text": "We understand that there's a correlation between action X and outcome Y. We don't always know why that correlation exists. We don't even know if it's a causal relationship between X and Y. Yet, given the information that X is correlated with Y, oftentimes we get medical advice, you know, don't do X or do a lot of X. Same thing in economics, you know, we know that, you know,"}, {"start": 3745.44, "end": 3771.12, "text": "Deflation on a large scale was associated with the Great Depression. As economists, we have ideas as to why deflation caused bad economic outcomes, but we're not sure of it yet. Yet we still use as a rule of thumb in central banking, it'd be great to avoid deflation. We don't exactly know why yet, but we want to know why, and I'm sure that in the computer science community..."}, {"start": 3771.12, "end": 3797.52, "text": "There is a desire to know how the deep learning machines are coming to these conclusions. So it's a matter of pragmatism. Do you just not do it until you know, or do you do it in the interim with a grain of salt? Most of our policies. I mean, to whatever extent, I'm not a lawyer. I don't particularly care for the law myself."}, {"start": 3798.32, "end": 3813.84, "text": "But in my job here as a computer scientist who's trying to harmonize technology, trying to apologize to the world for these ill fits, in that regard, whatever the societal norms are,"}, {"start": 3813.84, "end": 3834.08, "text": "I consider it a complete success if I can deploy a technology that preserves those norms. If society wants to debate and Congress wants to decide it's actually going to pass a law or something like that, and they produce new norms, then the technology should be able to account to the norms. I want the people we elect..."}, {"start": 3834.08, "end": 3863.6, "text": "Controlling those norms, not the technology itself. So what does that mean? That means where we have laws, where we have barriers, like in the case that I gave on the civil rights or on hiring or on sentencing, all of these places where we have expectations or we have explicit laws or regulations, the algorithms have to be able to be transparent, tested, have some kind of warranty or be available for inspection. Though I don't think it's so much inspection."}, {"start": 3863.6, "end": 3885.9, "text": "A lot of times, like in the civil rights law investigation, it really didn't matter who was responsible for those ads. At the end of the day, the fact that it happened was the problem. And so being able to measure that just simply says that's a problem, that has to stop, and now we can figure out what you're going to do to stop it."}, {"start": 3886.0, "end": 3917.86, "text": "Thank you. So another question. How can we make it so that AI doesn't intimidate people into devaluing their personal data and identity? I mean, I guess Lucy gave us a start. I mean, one way might be to sort of dissipate this halo effect or, you know, through language that's very precise about, you know, what one is getting into when one participates in data. I mean, maybe one of the things that question is kind of..."}, {"start": 3917.86, "end": 3935.87, "text": "Pointing us toward is the ways in which we are enrolled. One argument would be that we are increasingly constituting ourselves in the way that our systems need us to be."}, {"start": 3935.87, "end": 3964.19, "text": "So we experience that with service encounters. When we get automated service systems, we learn to make ourselves into the kind of interlocutors that are most likely to have a successful communication with the system. And similarly, in our increasing self-identification,"}, {"start": 3964.19, "end": 3993.82, "text": "In relation to our data persona, our data traces, there is a kind of mutual reshaping going on there that we might want to at least, we need to at least recognize that that's going on and open up the possibility of saying, of resisting it in whatever ways make sense."}, {"start": 3993.82, "end": 4021.42, "text": "Go off of that in two different directions. One is most of the data that we leave behind is really not something that we can control how others are using it. But the harms we can definitely feel. So it's not just the devalue. But you can't prove it if you can't show the data sharing happened or that it was the consequence. That hiring, firing, or promotion decision was based on data that you don't even know was used or have access to."}, {"start": 4021.42, "end": 4043.02, "text": "But at the same time, this idea of to what extent can individuals get control of a copy of their data and build another infrastructure that allows them to have more account, that the apps and so forth will be more accountable to their data might be a way forward. I think it was a fantastic question."}, {"start": 4043.18, "end": 4063.87, "text": "Thanks. So this one looks queued up for Mustafa. You've kind of answered it already. So with great power comes great responsibility. What should be the responsibilities of engineers and designers of AI systems? So I might add to that, what's unique about what you've been doing at DeepMind is really taking the leadership and trying to do..."}, {"start": 4063.87, "end": 4091.23, "text": "You know, to have an ethical relationship to data and to the communities that you're drawing on, involving communities and that. So maybe another way of answering this, since you've talked about some of the things you're doing around the design, around systems, is what you see in your field or your industry that you're responding to. I mean, why did you feel like you, you know, what are you responding to when taking the lead to create these different ethical structures? So I think our responsibility is to..."}, {"start": 4091.23, "end": 4117.73, "text": "Innovate as fast as we're innovating on the technology with the governance and that means that we have to have the courage to get it wrong and say when we've made mistakes and the courage to try new forms of oversight and transparency and accountability within the structure of a company as we are limited by the conventional form that that imposes on everything that we do and I think if in say"}, {"start": 4117.73, "end": 4128.8, "text": "Five to ten years time, I can look back and say, you know what, we actually created a constructive semi-permeable membrane."}, {"start": 4128.8, "end": 4158.34, "text": "of really smart and engaged people who really care about these questions of ethics and oversight. And we've given them access to the things that we work on, the algorithms that we build, and the data that we have. And we're shining a light of transparency on some of the stuff that we do. And people have responded in a really constructive and productive and engaged way. I think that would be pretty cool. That's what we're trying to get towards. But it's going to be tough."}, {"start": 4158.34, "end": 4178.99, "text": "It's tough. And how does that become disseminated? I mean, is it Google's the multinational international leader and others will follow suit? I mean, is part of your mission to sort of filter this kind of practice through the industry?"}, {"start": 4178.99, "end": 4208.93, "text": "What we need to first do is figure out what works. So we're still a relatively small group. We operate autonomously from the rest of Google, and we have about 300 or so people. And when we start to demonstrate that there are positive signs that we're achieving something like this semi-permeable membrane of oversight, then I think we can start to think about how other players in the industry might want to learn something from it, but that'll be up to them. So this is running out of time. Do any of you want to weigh in on this question of..."}, {"start": 4208.93, "end": 4240.08, "text": "Great responsibility for engineers and designers. Other than Superman is my favorite hero. So thank you all. This has been a tremendous honor and a pleasure. I've learned a lot and I'm sure many of the folks here did as well. Please join me in thanking Mustafa Suleiman, Professor Henry Hsu, Professor Nassana Sweeney, Lucy Suchman for a very terrific panel."}, {"start": 4240.08, "end": 4270.08, "text": "So we're almost at the end of the show. I want to thank you all for being present for what I think has been a truly extraordinary discussion. We've had so many rich ideas and quite complex challenges thrown at us tonight. Obviously, we can't come up with all of the answers in the here and now, but I think we've got a sense now of what it might take. And to me, I can certainly see a lot of potential for..."}, {"start": 4270.08, "end": 4294.29, "text": "Cross-disciplinary collaboration and we've seen so much of that tonight. Probably also cross-industry research and I think a lot of public participation. The only way I think to know if the sorts of technologies we've been talking about tonight are actually meeting the needs of diverse populations is if we include the widest spectrum of participants in the design and implementation of AI."}, {"start": 4295.63, "end": 4308.94, "text": "So this is not the end. This is just the beginning of a much bigger project. And so ideally what we would like is that events like this will inform where and how AI will be deployed in the next five to ten years."}, {"start": 4308.94, "end": 4336.88, "text": "And finally, speaking of participation, many of you know that the White House currently has an RFI, a request for information, around artificial intelligence. That's open to everyone in this room to actually send your thoughts and ideas. Perhaps they've been sparked further by tonight to the White House, and the closing date for that is July 22. So as with any event of this scale, we have a lot of institutions and people to thank. First..."}, {"start": 4336.88, "end": 4365.79, "text": "Our sponsors, the John D. and Catherine T. MacArthur Foundation, Google Open Research, and Microsoft Research. We also have all of our speakers and our chairpeople. Thank you so much. That was really wonderful. As well as our program committee, Eric Horvitz, Helen Nissenbaum, Latanya Sweeney, Mustafa Suleiman, Nicole Wong, and Ryan Calo. Our research team, who did so much work behind the scenes, Madeline Clara Lisch, Solon Barakas, Aaron Plasek, Khadija Ferry,"}, {"start": 4365.79, "end": 4395.68, "text": "Ferryman, pardon, and our producers, Mariah Peebles and Alejandro Calcagno. And finally, to the NYU Information Law Institute and, of course, to the White House for their vision and support. And I just want to say one more thank you, actually. I'm going to sneak in. One of the great privileges of curating a program like this is that we get to bring the speakers who've inspired us in our work. So I want to say a big extra thank you to all of the speakers who've made us do the work that we do. This is to you."}, {"start": 4395.68, "end": 4423.06, "text": "Thank you. And last but absolutely not least, my co-chair and co-pilot in this entire process, Meredith Whitaker. Thank you for making it happen. Oh, wow. And thank you, Kate. It has been a real pleasure putting this dream team together with you. And I'm leaving tonight inspired, and I hope all of you are too. Thank you and good night. Thanks and good night."}]}