WEBVTT - Smart Talks with IBM: Scaling AI With Purpose 0:00:04.519 --> 0:00:12.719 Welcome to tech Stuff, a production from iHeartRadio. This season 0:00:12.840 --> 0:00:16.280 on smart Talks with IBM, Malcolm Gladwell and team are 0:00:16.320 --> 0:00:19.560 diving into the transformative world of artificial intelligence with a 0:00:19.600 --> 0:00:24.239 fresh perspective on the concept of open What does open 0:00:24.680 --> 0:00:28.000 really mean in the context of AI. It can mean 0:00:28.120 --> 0:00:31.720 open source code or open data, but it also encompasses 0:00:31.800 --> 0:00:36.120 fostering an ecosystem of ideas, ensuring diverse perspectives are heard, 0:00:36.280 --> 0:00:40.000 and enabling new levels of transparency. Join hosts from your 0:00:40.040 --> 0:00:43.640 favorite Pushkin podcasts as they explore how openness in AI 0:00:43.880 --> 0:00:49.160 is reshaping industries, driving innovation, and redefining what's possible. You'll 0:00:49.159 --> 0:00:52.440 hear from industry experts and leaders about the implications and 0:00:52.479 --> 0:00:56.280 possibilities of open AI, and of course, Malcolm Gladwell will 0:00:56.280 --> 0:00:58.280 be there to guide you through the season with his 0:00:58.440 --> 0:01:01.960 unique insights. Look out for new episodes of Smart Talks 0:01:02.040 --> 0:01:05.440 every other week on the iHeartRadio app, Apple Podcasts, or 0:01:05.440 --> 0:01:09.000 wherever you get your podcasts, and learn more at IBM 0:01:09.160 --> 0:01:21.400 dot com, slash smart Talks. 0:01:18.720 --> 0:01:26.720 Pushkin Hello, Hello, Welcome to Smart Talks with IBM, a 0:01:26.800 --> 0:01:33.120 podcast from Pushkin Industries, iHeartRadio and IBM. I'm Malcolm Gladwell. 0:01:33.160 --> 0:01:36.400 This season, we're diving back into the world of artificial intelligence, 0:01:36.760 --> 0:01:39.760 but with a focus on the powerful concept of open 0:01:40.319 --> 0:01:46.399 its possibilities, implications, and misconceptions. We'll look at openness from 0:01:46.440 --> 0:01:49.200 a variety of angles and explore how the concept is 0:01:49.240 --> 0:01:53.280 already reshaping industries, ways of doing business, and our very 0:01:53.360 --> 0:01:58.639 notion of what's possible. In today's episode, Jacob Goldstein sits 0:01:58.640 --> 0:02:02.640 down with Rebecca Finley, the CEO of the Partnership on AI, 0:02:03.320 --> 0:02:07.360 a nonprofit group grappling with important questions around the future 0:02:07.360 --> 0:02:11.560 of AI. Their conversation focuses on Rebecca's work bringing together 0:02:11.880 --> 0:02:15.800 a community of diverse stakeholders to help shape the conversation 0:02:16.320 --> 0:02:21.800 around accountable AI governance. Rebecca explains why transparency is so 0:02:21.960 --> 0:02:26.960 crucial for scaling the technology responsibly, and she highlights how 0:02:27.000 --> 0:02:30.520 working with groups like the AI Alliance can provide valuable 0:02:30.520 --> 0:02:35.079 insights in order to build the resources, infrastructure, and community 0:02:35.480 --> 0:02:40.400 around releasing open source models. So, without further ado, let's 0:02:40.400 --> 0:02:41.680 get to that conversation. 0:02:48.800 --> 0:02:50.480 Can you say your name and your job? 0:02:50.960 --> 0:02:54.120 My name is Rebecca Finley. I am the CEO of 0:02:54.160 --> 0:02:57.840 the Partnership on AI to benefit people and society. Often 0:02:57.919 --> 0:02:59.800 referred to as PAI. 0:03:00.760 --> 0:03:03.280 How did you get here? What was your job before 0:03:03.320 --> 0:03:05.160 you had the job that you have now. 0:03:06.160 --> 0:03:11.079 I came to PAI about three years ago, having had 0:03:11.080 --> 0:03:16.320 the opportunity to work for the Canadian Institute for Advance Research, 0:03:16.880 --> 0:03:21.560 developing and deploying all of their programs related to the 0:03:21.600 --> 0:03:27.160 intersection of technology and society. And one of the areas 0:03:27.360 --> 0:03:31.680 that the Canadian Institute had been funding since nineteen eighty 0:03:31.720 --> 0:03:35.560 two was research into artificial intelligence. 0:03:35.840 --> 0:03:37.840 Wow, early, they were early. 0:03:38.880 --> 0:03:43.200 It was a very early commitment and an ongoing commitment 0:03:43.320 --> 0:03:48.720 at the Institute to fund long term fundamental questions of 0:03:48.760 --> 0:03:57.720 scientific importance in interdisciplinary research programs that were often committed 0:03:57.720 --> 0:04:01.240 and funded to for well over a decade. The AI, 0:04:01.480 --> 0:04:04.560 Robotics and Society program that kicked off the work at 0:04:04.560 --> 0:04:10.280 the Institute eventually became a program very much focused on 0:04:10.880 --> 0:04:16.000 deep learning and reinforcement learning, neural networks, all of the 0:04:16.680 --> 0:04:21.960 current iteration of AI, or certainly the pregenerative AI iteration 0:04:22.080 --> 0:04:26.599 of AI that led to this transformation that we've seen 0:04:26.720 --> 0:04:29.960 in terms of online search and all sorts of ways 0:04:30.000 --> 0:04:32.880 in which predictive AI has been deployed. So I had 0:04:32.920 --> 0:04:36.280 the opportunity to see the very early days of that 0:04:36.440 --> 0:04:40.680 research coming together, and when in the early sort of 0:04:40.720 --> 0:04:46.479 two thousand, twenty and tens, when compute capability came together 0:04:46.680 --> 0:04:51.480 with data capability through some of the Internet companies and otherwise, 0:04:51.560 --> 0:04:55.239 and we really saw this technology start to take off. 0:04:55.960 --> 0:04:59.360 I had the opportunity to start up a program specifically 0:04:59.400 --> 0:05:04.040 focused on the impacts of AI in society. There was, 0:05:04.160 --> 0:05:06.839 as you know, at that time, some concerns both about 0:05:06.839 --> 0:05:11.279 the potential for the technology, but also in terms of 0:05:11.320 --> 0:05:14.280 what we were seeing around data sets and bias and 0:05:14.279 --> 0:05:18.839 discrimination and potential impact on future jobs. And so bringing 0:05:18.880 --> 0:05:23.200 a whole group of experts, whether they were ethicists or 0:05:23.680 --> 0:05:29.200 lawyers or economists, sociologists into the discussion about AI was 0:05:29.279 --> 0:05:31.920 core to that new program and continues to be core 0:05:32.000 --> 0:05:35.839 to my commitment to bringing diverse perspectives together to solve 0:05:36.279 --> 0:05:39.159 the challenges and opportunities that AI offers today. 0:05:40.279 --> 0:05:42.839 So specifically, what is your job now? What is the 0:05:42.839 --> 0:05:44.920 work you do? What is the work that PAI does? 0:05:46.000 --> 0:05:49.839 I like to answer that question by asking two questions. 0:05:50.279 --> 0:05:53.680 First and foremost, do you believe that the world is 0:05:53.920 --> 0:05:57.839 more divided today than it ever has been in recent history? 0:05:58.680 --> 0:06:02.480 And do you believe that if we don't create spaces 0:06:03.080 --> 0:06:06.440 for very different perspectives to come together, we won't be 0:06:06.520 --> 0:06:09.240 able to solve the challenges that are in front of 0:06:09.279 --> 0:06:13.240 the world today. My answer to both of those questions is, yes, 0:06:13.720 --> 0:06:17.279 we're more divided, and two, we need to seek out 0:06:17.400 --> 0:06:22.159 those spaces where those very different perspectives can come together 0:06:22.640 --> 0:06:25.880 to solve those great challenges. And that's what I get 0:06:25.880 --> 0:06:29.440 to do as CEO of the Partnership on AI. We 0:06:29.440 --> 0:06:34.159 were begun in twenty sixteen with a fundamental commitment to 0:06:34.400 --> 0:06:40.920 bringing together experts, whether they were in industry, academia, civil society, 0:06:41.040 --> 0:06:45.280 or philanthropy, coming together to identify what are the most 0:06:45.279 --> 0:06:49.000 important questions when we think about developing AI centered on 0:06:49.080 --> 0:06:52.480 people and communities, and then how do we begin to 0:06:52.600 --> 0:06:55.400 develop the solutions to make sure we benefit appropriately. 0:06:56.400 --> 0:07:01.719 So that's a very big picture set of ideas. I'm 0:07:01.760 --> 0:07:04.000 curious on a sort of more day to day level. 0:07:04.080 --> 0:07:06.400 I mean, you talk about collaborating with all these different 0:07:06.480 --> 0:07:08.800 kinds of people, all these different groups, what does that 0:07:08.880 --> 0:07:11.600 actually look like. What are some specific examples of how 0:07:11.640 --> 0:07:12.480 you do this work? 0:07:13.040 --> 0:07:16.560 So right now we have about one hundred and twenty 0:07:16.680 --> 0:07:22.280 partners in sixteen countries. They come together through working groups 0:07:22.360 --> 0:07:25.320 that we look at through a variety of different perspectives. 0:07:25.360 --> 0:07:29.239 It could be AI, labor and the economy. It could 0:07:29.280 --> 0:07:33.960 be how do you build a healthy information ecosystem. It 0:07:34.000 --> 0:07:37.239 could be how do you bring more diverse perspectives into 0:07:37.280 --> 0:07:41.720 the inclusive and equitable development of AI. It could be 0:07:41.840 --> 0:07:46.120 what are the emerging opportunities with these very very large 0:07:46.160 --> 0:07:50.120 foundation model applications and how do you deploy those safely? 0:07:50.600 --> 0:07:54.160 And these groups come together most importantly to say, what 0:07:54.240 --> 0:07:57.680 are the questions we need to answer collectively, So they 0:07:57.720 --> 0:08:00.520 come together in working groups. I have an amazing staff 0:08:00.560 --> 0:08:04.120 team who hold the pen on synthesizing research and data 0:08:04.440 --> 0:08:10.000 and evidence, developing frameworks, best practices, resources, all sorts of 0:08:10.080 --> 0:08:12.480 things that we can offer up to the community, be 0:08:12.600 --> 0:08:16.280 they in industry or in policy, to say this is 0:08:16.360 --> 0:08:19.000 how we can well, this is what good looks like, 0:08:19.120 --> 0:08:20.520 and this is how we can do it on a 0:08:20.560 --> 0:08:22.560 day to day basis. So that's what we do, and 0:08:22.600 --> 0:08:25.920 then we publish our materials. It's all open. We make 0:08:25.960 --> 0:08:28.200 sure that we get them into the hands of those 0:08:28.200 --> 0:08:30.920 communities that can use them, and then we drive and 0:08:30.960 --> 0:08:33.520 work with those communities to put them into practice. 0:08:34.160 --> 0:08:37.120 You use the word open there and describing your publications. 0:08:38.040 --> 0:08:40.800 I know, in the world of AI, on the sort 0:08:40.800 --> 0:08:44.640 of technical side, there's a lot of debate, say, or 0:08:44.720 --> 0:08:49.679 discussion about kind of open versus closed AI, And I'm 0:08:49.720 --> 0:08:53.880 curious how you kind of encounter that particular discussion. What 0:08:54.040 --> 0:08:56.040 is your view on open versus closed AI. 0:08:57.080 --> 0:09:02.920 So the current discussion between open and closed release of 0:09:03.080 --> 0:09:08.280 AI models came once we saw chat, GPT and other 0:09:08.559 --> 0:09:12.960 very large generative AI systems being deployed out into the 0:09:13.000 --> 0:09:18.679 hands of consumers around the world, and there emerged some 0:09:18.840 --> 0:09:23.640 fear about the potential of these models to act in 0:09:23.720 --> 0:09:27.240 all sorts of catastrophic ways. So there were concerns that 0:09:27.280 --> 0:09:32.319 the models could be deployed with regard to different development 0:09:32.320 --> 0:09:37.679 of viruses or biomedical weapons, or even nuclear weapons, or 0:09:37.720 --> 0:09:41.800 through manipulation or otherwise. So this are emerged about over 0:09:41.840 --> 0:09:47.640 the last eighteen months, this real concern that these models, 0:09:47.760 --> 0:09:52.079 if deployed openly, could lead to some level of truly 0:09:52.200 --> 0:09:57.800 catastrophic risk. And what emerged is actually that we discovered 0:09:58.360 --> 0:10:00.240 that through a whole bunch of work that's been done 0:10:00.280 --> 0:10:03.600 over the last little while, that releasing them openly has 0:10:03.679 --> 0:10:06.040 not led and doesn't appear to be leading in any 0:10:06.040 --> 0:10:11.320 way to catastrophic risk. In facts, releasing them openly allows 0:10:11.360 --> 0:10:16.120 for much more greater scrutiny and understanding of the safety 0:10:16.160 --> 0:10:18.800 measures that have been put into place. And so what 0:10:18.920 --> 0:10:22.560 happened was sort of the pendulum swung very much towards 0:10:22.600 --> 0:10:26.120 concerned about really catastrophic risk and safety over the last year, 0:10:26.160 --> 0:10:28.760 and over the last year we've seen it swing back 0:10:28.840 --> 0:10:31.559 as we learn more and more about how these models 0:10:31.840 --> 0:10:34.440 are being used and how they are being deployed into 0:10:34.480 --> 0:10:40.439 the world. My feeling is we must approach this work openly, 0:10:40.800 --> 0:10:44.080 and it's not just open release of models or what 0:10:44.120 --> 0:10:48.080 we think of as traditional open source forms of model 0:10:48.120 --> 0:10:51.280 development or otherwise, but we really need to think about 0:10:51.320 --> 0:10:56.040 how do we build an open innovation ecosystem that fundamentally 0:10:56.080 --> 0:11:00.000 allows both for the innovation to be shared with many people, 0:11:00.280 --> 0:11:03.960 but also for safety and security to be rigorously upheld. 0:11:04.440 --> 0:11:07.880 So when you talk about this kind of broader idea 0:11:07.880 --> 0:11:12.480 of open innovation beyond open source or you know, transparency 0:11:12.480 --> 0:11:15.840 and models, like, what do you mean sort of specifically 0:11:15.840 --> 0:11:17.320 how does that look in the world. 0:11:17.679 --> 0:11:21.320 So I have three particular points of view when it 0:11:21.320 --> 0:11:23.960 comes to open innovation because I think we need to 0:11:24.000 --> 0:11:27.400 think both both upstream around the research that is driving 0:11:27.440 --> 0:11:30.560 these models and downstream in terms of the benefits of 0:11:30.600 --> 0:11:34.320 these models to others. So, first and foremost, what we 0:11:34.400 --> 0:11:37.000 have known in terms of how AI has been developed, 0:11:37.000 --> 0:11:39.040 and yes, I had an opportunity to see it when 0:11:39.040 --> 0:11:42.160 I was at the Canadian Institute for Advanced Research is 0:11:42.200 --> 0:11:48.360 a very open form of scientific publication and rigorous peer review. 0:11:48.640 --> 0:11:51.720 And what happens when we release openly is you have 0:11:51.800 --> 0:11:56.040 an opportunity for the research to be interrogated to determine 0:11:56.040 --> 0:11:59.320 the quality and significance of that, but then also for 0:11:59.400 --> 0:12:02.240 it to be picked up by many others. And then secondly, 0:12:02.679 --> 0:12:06.800 openness for me is about transparency. We released a set 0:12:06.800 --> 0:12:10.360 of very strong recommendations last year around the way in 0:12:10.440 --> 0:12:14.680 which these very large foundation models could be deployed safely. 0:12:15.360 --> 0:12:19.400 They're all about disclosure. They're all about disclosure and documentation 0:12:19.600 --> 0:12:22.880 right from the early days pre R and D development 0:12:22.920 --> 0:12:25.400 of these systems, right in terms of thinking about what's 0:12:25.440 --> 0:12:27.960 in the training data and how is it being used, 0:12:28.120 --> 0:12:32.160 all the way through to post deployment monitoring and disclosure. 0:12:32.720 --> 0:12:35.960 So I really think that this is important transparency through it. 0:12:36.000 --> 0:12:38.880 And then the third piece is openness in terms of 0:12:38.920 --> 0:12:42.319 who was around the table to benefit from this technology. 0:12:42.720 --> 0:12:44.840 We know that if we're really going to see these 0:12:44.880 --> 0:12:49.920 new models having being successful deployed into education or healthcare 0:12:50.120 --> 0:12:53.920 or climate and sustainability, we need to have those experts 0:12:54.000 --> 0:12:57.120 in those communities at the table charting this and making 0:12:57.160 --> 0:12:59.800 sure that the technology is working for them. Those are 0:12:59.800 --> 0:13:01.520 the free ways I think about openness. 0:13:02.640 --> 0:13:06.040 Is there like a particular project that you've worked on 0:13:06.160 --> 0:13:10.080 that you feel like, you know reflects your approach to 0:13:10.240 --> 0:13:11.240 responsible AI. 0:13:12.320 --> 0:13:15.280 So there's a really interesting project that we have underway 0:13:15.280 --> 0:13:19.480 at PAI that is looking at responsible practices squarely when 0:13:19.520 --> 0:13:23.040 it comes to the use of synthetic media. And what 0:13:23.080 --> 0:13:26.680 we heard from our community was that they were looking 0:13:26.720 --> 0:13:30.240 for a clear code of conduct about what does it 0:13:30.360 --> 0:13:33.280 mean to be responsible in this space. And so what 0:13:33.440 --> 0:13:36.480 happened is we pulled together a number of working groups 0:13:36.480 --> 0:13:40.359 to come together. They included industry representatives, They also included 0:13:40.640 --> 0:13:46.400 civil society organizations like WITNESS, a number of academic institutions 0:13:46.440 --> 0:13:49.520 and otherwise. And what we heard was that there were 0:13:49.760 --> 0:13:55.680 clear requirements that creators could take, that developers of the 0:13:55.720 --> 0:13:58.800 technology could take. And then also distributors. So when we 0:13:58.880 --> 0:14:03.360 think about those generative AI systems being deployed across platforms 0:14:03.400 --> 0:14:06.560 and otherwise, and we came up with a framework for 0:14:06.640 --> 0:14:10.000 what responsibility looks like. What does it mean to have consent, 0:14:10.160 --> 0:14:13.800 what does it mean to disclose responsibly, what does it 0:14:13.840 --> 0:14:18.120 mean to embed technology into it? So, for example, we've 0:14:18.120 --> 0:14:21.160 heard many people talk about the importance of water marking 0:14:21.240 --> 0:14:23.640 systems right and making sure that we have a way 0:14:23.640 --> 0:14:25.920 to water mark them. But what we know from the 0:14:25.960 --> 0:14:30.240 technology is that is a very very complex and complicated problem, 0:14:30.480 --> 0:14:33.680 and what might work on a technical level certainly hits 0:14:33.720 --> 0:14:36.720 a whole new set of complications when we start labeling 0:14:36.760 --> 0:14:39.840 and disclosing out to the public about what that technology 0:14:39.880 --> 0:14:43.520 actually means. All of these I believe are solvable problems, 0:14:43.520 --> 0:14:47.160 but they all needed to have a clear code underneath 0:14:47.160 --> 0:14:49.400 them that was saying this is what we will commit to. 0:14:49.720 --> 0:14:53.000 And we now have a number of organizations, many many 0:14:53.000 --> 0:14:56.240 of the large technology companies, but also many of the 0:14:56.720 --> 0:14:59.840 small startups who are operating in this base, civil society, 0:15:00.040 --> 0:15:03.840 media organizations like the BBC and the CBC who's have 0:15:04.000 --> 0:15:07.520 signed on. And one of the really exciting pieces of 0:15:07.640 --> 0:15:11.160 that is that we're now seeing how it's changing practice. 0:15:11.440 --> 0:15:14.200 So a year in we asked each of our partners 0:15:14.240 --> 0:15:17.440 to come up with a clear case study about how 0:15:17.480 --> 0:15:20.720 that work has changed the way they are making decisions, 0:15:21.120 --> 0:15:25.720 deploying technology and ensuring that they're being responsible in their use. 0:15:25.800 --> 0:15:28.720 And that is creating now a whole resource online that 0:15:28.720 --> 0:15:31.080 we're able to share with others about what does it 0:15:31.200 --> 0:15:34.480 mean to be responsible in this place. There's so much 0:15:34.520 --> 0:15:36.680 more work to be done, and the exciting thing is 0:15:36.720 --> 0:15:38.920 once you have a foundation like this in place, we 0:15:38.960 --> 0:15:42.560 can continue to build on it. So much interest now 0:15:42.560 --> 0:15:45.520 in the policy space, for example, about this work as well. 0:15:46.600 --> 0:15:50.440 Are there any specific examples of those sort of case 0:15:50.480 --> 0:15:55.160 studies or the real world experiences that say media organizations 0:15:55.240 --> 0:15:57.560 had that are interesting that are illuminating. 0:15:57.960 --> 0:16:02.960 Yes. So, for example, what we saw with the BBC 0:16:03.560 --> 0:16:06.680 is that they're developing a lot of content as a 0:16:06.680 --> 0:16:10.200 public broadcaster, both in terms of their news coverage but 0:16:10.320 --> 0:16:12.880 also in terms of some of the resources that they 0:16:12.880 --> 0:16:16.600 are developing for the British public as well. And what 0:16:16.640 --> 0:16:19.080 they talked about was the way in which they had 0:16:19.320 --> 0:16:24.840 used synthetic media in a very very sensitive environment where 0:16:24.880 --> 0:16:29.520 they were hearing from individuals talk about personal experiences, but 0:16:29.640 --> 0:16:33.200 wanted to have some way to change the face entirely 0:16:33.320 --> 0:16:36.320 in terms of the individuals who were speaking. So that's 0:16:36.360 --> 0:16:39.440 a very complicated ethical question, right, how do you do 0:16:39.600 --> 0:16:42.400 that responsibly? And what is the way in which you 0:16:42.560 --> 0:16:46.640 use that technology, and most importantly, how do you disclose it? 0:16:46.920 --> 0:16:49.400 So their case study looked at that in some real 0:16:49.520 --> 0:16:53.120 detail about the process they went through to make the 0:16:53.200 --> 0:16:57.160 decision responsibly to do what they chose, how they intended 0:16:57.200 --> 0:16:58.920 to use the technology in that space. 0:17:00.040 --> 0:17:04.280 Describe your work and some of these studies, the idea 0:17:04.320 --> 0:17:08.159 of transparency seems to be a theme. Talk about the 0:17:08.200 --> 0:17:10.200 importance of transparency in this kind of work. 0:17:11.359 --> 0:17:16.320 Yeah, transparency is fundamental to responsibility. I always like to 0:17:16.320 --> 0:17:20.159 say it's not accountability in a complete sense, but it 0:17:20.240 --> 0:17:24.159 is a first step to driving accountability more fully. So, 0:17:24.680 --> 0:17:27.960 when we think about how these systems are developed, they're 0:17:28.000 --> 0:17:33.240 often developed behind closed doors inside companies who are making 0:17:33.280 --> 0:17:37.320 decisions about what and how these products will work from 0:17:37.359 --> 0:17:42.600 a business perspective. And what disclosure and transparency can provide 0:17:42.640 --> 0:17:46.040 is some sense of the decisions that were made leading 0:17:46.080 --> 0:17:48.879 up to the way in which those models were deployed. 0:17:49.000 --> 0:17:54.359 So This could be ensuring that individual's private information was 0:17:54.400 --> 0:17:58.720 protected through the process and won't be inadvertently disclosed, or otherwise, 0:17:59.160 --> 0:18:02.199 it could be providing some sense of how well the 0:18:02.280 --> 0:18:06.000 system performs against a whole level of quality measures. So 0:18:06.040 --> 0:18:08.840 we have all of these different types of evaluations and 0:18:08.920 --> 0:18:12.200 a measures that are emerging about the quality of these 0:18:12.200 --> 0:18:16.080 systems as they're deployed. Being transparent about how they perform 0:18:16.119 --> 0:18:19.240 against these systems is really crucial to that as well. 0:18:19.480 --> 0:18:22.560 We have a whole ecosystem that's starting to emerge around 0:18:22.640 --> 0:18:25.639 auditing of these systems. So what does that look like 0:18:25.720 --> 0:18:28.120 we think about auditors and all sorts of other sectors 0:18:28.160 --> 0:18:30.639 of the economy. What does it look like to be 0:18:30.720 --> 0:18:33.919 auditing these systems to ensure that they're meeting all of 0:18:33.960 --> 0:18:37.640 those both legal but additional ethical requirements that we want 0:18:37.680 --> 0:18:38.840 to make sure that are in place. 0:18:40.160 --> 0:18:44.720 What are some of the hardest ethical dilemmas you've come 0:18:44.800 --> 0:18:46.920 up against in AI policy? 0:18:48.160 --> 0:18:51.520 Well, the interesting thing about AI policy right is what 0:18:51.560 --> 0:18:55.800 it works very simply in one setting, can be highly 0:18:55.880 --> 0:18:59.199 complicated in another setting. And so, for example, I have 0:18:59.240 --> 0:19:02.119 an app that I adore. It's an app on my 0:19:02.240 --> 0:19:05.240 phone that allows me to take a photo of a bird, 0:19:05.800 --> 0:19:08.040 and it will help me to better understand, you know, 0:19:08.080 --> 0:19:10.439 what that bird is, and give me all sorts of 0:19:10.440 --> 0:19:14.639 information about that bird. Now, it's probably right most of 0:19:14.680 --> 0:19:17.160 the time, and it's certainly right enough of the time 0:19:17.200 --> 0:19:20.200 to give me great pleasure and delight when I'm out walking. 0:19:20.760 --> 0:19:24.199 You could think about that exact same technology applied. So, 0:19:24.440 --> 0:19:27.680 for example, now you're a security guard and you're working 0:19:28.240 --> 0:19:31.840 in a shopping plaza, and you're able to take photos 0:19:31.880 --> 0:19:35.280 of individuals who you may think are acting suspiciously in 0:19:35.280 --> 0:19:37.840 some way and match that photo up with some sort 0:19:37.920 --> 0:19:42.119 of a database of individuals that may have been found, 0:19:42.119 --> 0:19:44.920 you know, to have some sort of connection to other 0:19:45.000 --> 0:19:47.360 criminal behavior in the past. Right, So what goes from 0:19:47.440 --> 0:19:50.919 being a delightful Oh, isn't this an interesting bird? To 0:19:51.000 --> 0:19:55.399 a very very creepy What is this say about surveillance 0:19:55.480 --> 0:19:59.080 and privacy and access to public spaces? And that is 0:19:59.119 --> 0:20:02.359 the nature of AI. So much of the concern about 0:20:02.359 --> 0:20:07.520 the ethical use and deployment of AI is how an 0:20:07.640 --> 0:20:13.080 organization is making the choices within the social and systemic 0:20:13.160 --> 0:20:16.919 structure they sit. So so much about the ethics of 0:20:16.960 --> 0:20:20.679 AI is understanding what is the use case how is 0:20:20.720 --> 0:20:24.440 it being used, how is it being constrained? How does 0:20:24.480 --> 0:20:27.520 it start to infringe upon what we think of as 0:20:27.880 --> 0:20:32.320 the human rights of an individual to privacy, And so 0:20:32.400 --> 0:20:35.640 you have to constantly be thinking about ethics. What could 0:20:35.640 --> 0:20:39.119 work very well in one situation absolutely doesn't work in another. 0:20:39.480 --> 0:20:43.000 We often talk about these as socio technical questions. Right, 0:20:43.400 --> 0:20:46.760 just because the technology works doesn't actually mean that it 0:20:46.800 --> 0:20:48.760 should be used and deployed. 0:20:49.520 --> 0:20:54.520 What's an example of where the partnership on AI influence 0:20:54.760 --> 0:20:58.160 changes either in policy or in industry practice. 0:20:59.320 --> 0:21:02.040 We talked a little bit about the Framework for Synthetic 0:21:02.119 --> 0:21:06.520 Media and how that has allowed companies and media organizations 0:21:06.560 --> 0:21:09.720 and civil society organizations to really think deeply about the 0:21:09.760 --> 0:21:12.600 way in which they're using this. Another area that we 0:21:12.760 --> 0:21:18.840 focused on has been around responsible deployment of foundation and 0:21:19.000 --> 0:21:21.679 large scale models. So, as I said, we issued a 0:21:21.800 --> 0:21:26.080 set of recommendations last year that really laid out for 0:21:26.280 --> 0:21:31.120 these very large developers and deployers of foundation and frontier models, 0:21:31.200 --> 0:21:35.439 what does good look like right from R and D 0:21:35.680 --> 0:21:39.520 through to deployment monitoring. And it has been very encouraging 0:21:39.720 --> 0:21:42.520 to see that that work has been picked up by 0:21:42.920 --> 0:21:47.000 companies and really articulated as part of the fabric of 0:21:47.040 --> 0:21:51.880 the deployment of their foundation models and systems moving forward. 0:21:52.280 --> 0:21:55.359 So much of this work is around creating clear definitions 0:21:55.359 --> 0:21:58.600 of what we're meaning as the technology evolves and clear 0:21:58.640 --> 0:22:01.160 sets of responsibilities. So it's great to see that work 0:22:01.200 --> 0:22:04.679 getting picked up. The NTIA in the United States just 0:22:04.720 --> 0:22:09.119 released a report on open models and the release of 0:22:09.119 --> 0:22:11.840 open models. Great to see our work sited there as 0:22:11.880 --> 0:22:15.200 contributing to that analysis. Great to see some of our 0:22:15.240 --> 0:22:19.199 definitions and synthetic media getting picked up by legislators in 0:22:19.240 --> 0:22:22.880 different countries. Really just it's important, I think, for us 0:22:22.880 --> 0:22:26.080 to build capacity, knowledge and understanding and our policy makers 0:22:26.119 --> 0:22:30.840 in this moment as the technology is evolving and accelerating 0:22:30.840 --> 0:22:31.600 in its development. 0:22:32.720 --> 0:22:36.320 What's the AI Alliance and why did Partnership on AI 0:22:36.440 --> 0:22:37.160 decide to join? 0:22:37.720 --> 0:22:41.240 So you had asked about the debate between open versus 0:22:41.359 --> 0:22:45.880 closed models and how that has evolved over the last year, 0:22:46.200 --> 0:22:50.719 and the AI Alliance was a community of organizations that 0:22:50.840 --> 0:22:55.080 came together to really think about, Okay, if we support 0:22:55.600 --> 0:22:59.119 open release of models, what does that look like and 0:22:59.160 --> 0:23:01.520 what does the community the need. And so that's about 0:23:01.560 --> 0:23:06.040 one hundred organizations IBM one of our founding partners is 0:23:06.080 --> 0:23:09.119 also one of the founding partners of the AI Alliance. 0:23:09.440 --> 0:23:12.719 It's a community that brings together a number of academic 0:23:12.800 --> 0:23:17.000 institutions many countries around the world, and they're really focused 0:23:17.119 --> 0:23:22.479 on how do you build the resources and infrastructure and 0:23:22.560 --> 0:23:26.680 community around what open source in these large scale models 0:23:26.720 --> 0:23:30.280 really mean. So that could be open data sets, that 0:23:30.400 --> 0:23:35.200 could be open technology development. Really building on that understanding 0:23:35.200 --> 0:23:38.320 that we need an infrastructure in place and a community 0:23:38.359 --> 0:23:43.600 engaged in thinking about safety and innovation through the open lens. 0:23:44.400 --> 0:23:48.520 This approach brings together organizations and experts from around the 0:23:48.560 --> 0:23:54.520 globe with different backgrounds, experiences, and perspectives to transparently and 0:23:54.680 --> 0:23:59.439 openly address the challenges and opportunities today. I poses the 0:23:59.480 --> 0:24:04.280 collaborative nature of the AI Alliance encourages discussion, debate, and innovation. 0:24:05.119 --> 0:24:08.200 Through these efforts, IBM is helping to build a community 0:24:08.520 --> 0:24:12.160 around transparent open technology. 0:24:12.800 --> 0:24:15.760 So I want to talk about the future for a minute. 0:24:16.119 --> 0:24:19.520 I'm trures what you see as the biggest obstacles to 0:24:20.400 --> 0:24:23.879 widespread adoption of responsible AI practices. 0:24:24.640 --> 0:24:29.959 One of the biggest obstacles today is an inability and 0:24:30.000 --> 0:24:33.880 a really a lack of understanding about how to use 0:24:33.960 --> 0:24:38.280 these models and how they can most effectively drive forward 0:24:38.720 --> 0:24:42.640 a company's commitment to whatever products and services it might 0:24:42.720 --> 0:24:46.479 be deploying. So I always recommend a couple of things 0:24:46.480 --> 0:24:49.040 for companies to really to think about this and to 0:24:49.119 --> 0:24:53.960 get started. One is think about how you are already 0:24:54.040 --> 0:24:57.359 using AI across all of your business products and services. 0:24:57.640 --> 0:25:02.520 Because already AI is integri into our workforces and into 0:25:02.600 --> 0:25:04.919 our workstreams and into the way in which companies are 0:25:04.920 --> 0:25:08.800 communicating with their clients every day. So understand how you 0:25:08.840 --> 0:25:12.760 are already using it and understand how you are integrating 0:25:12.920 --> 0:25:16.159 oversight and monitoring into those One of the best and 0:25:16.359 --> 0:25:19.800 clearest ways in which a company can really understand how 0:25:19.840 --> 0:25:22.919 to use this responsibly is through documentation. It's one of 0:25:22.920 --> 0:25:26.320 the areas where there's a clear consensus in the community. 0:25:26.400 --> 0:25:29.160 So how do you document the models that you are using, 0:25:29.280 --> 0:25:31.679 making sure that you've got a registry in place. How 0:25:31.720 --> 0:25:34.119 do you document the data that you are using and 0:25:34.160 --> 0:25:36.399 where that data comes from. This is sort of the 0:25:36.440 --> 0:25:39.840 first system, first line of defense in terms of understanding 0:25:39.880 --> 0:25:42.439 both what is in place and what you need to 0:25:42.440 --> 0:25:45.920 do in order to monitor it moving forward. And then secondly, 0:25:46.040 --> 0:25:48.560