WEBVTT - How a human-centered approach is building trustworthy AI. 0:00:05.920 --> 0:00:09.200 Hello, Hello. This is Smart Talks with IBM, a podcast 0:00:09.280 --> 0:00:13.039 from Pushkin Industries, I Heart Radio and IBM about what 0:00:13.039 --> 0:00:16.000 it means to look at today's most challenging problems in 0:00:16.040 --> 0:00:21.240 a new way. I'm Malcolm Glabbo. Today I'll be chatting 0:00:21.239 --> 0:00:25.400 with two IBM experts in artificial intelligence about the company's 0:00:25.440 --> 0:00:29.840 approach to building and supporting trustworthy AI as a force 0:00:30.240 --> 0:00:37.680 for positive change. I'll be speaking with IBMS Chief Privacy 0:00:37.720 --> 0:00:42.519 Officer Christina Montgomery. She oversees the company's privacy vision and 0:00:42.560 --> 0:00:46.960 compliance strategy globally, looking at things like immunity certificates and 0:00:47.080 --> 0:00:50.680 vaccine passports. Not what could we do, but what were 0:00:50.720 --> 0:00:52.640 we willing as a company to do? Where were we 0:00:52.680 --> 0:00:55.680 going to put our skills and our knowledge and our 0:00:55.720 --> 0:01:00.800 company brand in response to technologies that could help provide 0:01:00.960 --> 0:01:05.080 information in response to the pandemic. She also co chairs 0:01:05.440 --> 0:01:09.440 their AI Ethics Board. I'll also be talking with Dr 0:01:09.520 --> 0:01:14.200 Seth Dobrin, Global Chief AI Officer at IBM. Seth leads 0:01:14.240 --> 0:01:18.759 corporate AI strategy and is responsible for connecting AI development 0:01:18.959 --> 0:01:22.560 with the creation of business value. Seth is also a 0:01:22.560 --> 0:01:26.400 member of IBMS AI Ethics BORT. We want to make 0:01:26.440 --> 0:01:32.480 sure that the technology behind AI is as fair as possible, 0:01:33.240 --> 0:01:36.640 is as explainable as possible, is as robust as possible, 0:01:37.160 --> 0:01:40.720 and is as privacy preserving as possible. We'll talk about 0:01:40.720 --> 0:01:43.600 the need to create AI systems that are fair and 0:01:43.640 --> 0:01:46.720 addressed bias, and how we need to focus on trust 0:01:46.760 --> 0:01:50.920 and transparency to accomplish this. What might the future look 0:01:50.960 --> 0:01:55.560 like with an open and diverse ecosystem with governance across 0:01:55.600 --> 0:01:59.480 the industry. There's only one way to find out. Let's 0:02:00.000 --> 0:02:10.360 I did. One of the things I'm curious about is 0:02:11.080 --> 0:02:17.320 the origin of this interesting concern about the ethics and 0:02:17.440 --> 0:02:22.399 trust component of AI, or is this a later kind 0:02:22.440 --> 0:02:27.240 of of evolutionary concern. About ten years ago, when we 0:02:27.280 --> 0:02:31.480 started down this journey to transforming business using what we 0:02:31.520 --> 0:02:34.760 think about is AI today, the concept of trust came up, 0:02:34.760 --> 0:02:37.720 but not in the same context that we think about 0:02:37.720 --> 0:02:41.320 it today. The context of trust was really focused on 0:02:41.400 --> 0:02:43.360 how do I know it's given me the right answer 0:02:43.960 --> 0:02:46.560 so that I can make my decision. Because we didn't 0:02:46.560 --> 0:02:48.720 have tools that help explain how an AI came to 0:02:48.760 --> 0:02:51.360 a decision, you tended to have to get into these 0:02:51.400 --> 0:02:54.000 bakeoffs where you had to kind of set up experiments 0:02:54.040 --> 0:02:56.040 to show that the AI was at least as good 0:02:56.120 --> 0:03:00.200 as human if if not better, and understand why over 0:03:00.280 --> 0:03:04.880 time it's progressed as AI has in uh started to 0:03:04.919 --> 0:03:09.600 come up against real human conditions. And I think that's 0:03:09.600 --> 0:03:12.840 when we started thinking about what is going on with 0:03:12.919 --> 0:03:16.680 AI when it relates to bias, particularly you know, about 0:03:16.720 --> 0:03:19.520 five eight years ago there was an issue with mortgage 0:03:19.639 --> 0:03:23.480 particularly related to the zip code, but started giving you know, 0:03:23.520 --> 0:03:29.360 biases against people of certain races um and so I 0:03:29.400 --> 0:03:32.040 think those things combined have led to us to the 0:03:32.040 --> 0:03:34.800 point where we are today plufs. You know, the social 0:03:34.880 --> 0:03:37.800 justice movement over the last two years has has really 0:03:37.800 --> 0:03:42.120 accelerated a lot of the concern mm hmm. Because I 0:03:42.160 --> 0:03:46.800 noticed you're you're a lawyer by trade. It's an interesting 0:03:46.800 --> 0:03:51.880 subject because it seems like this is where AI experts 0:03:51.920 --> 0:03:55.720 like Seth and lawyers work together. It sounds like a 0:03:55.800 --> 0:03:59.280 kind of classic cross disciplinary endeavor. Can you talk about 0:03:59.320 --> 0:04:03.000 that a little bit. It's absolutely cross disciplinary in nature. 0:04:03.400 --> 0:04:07.680 For example, our AI Ethics Board, I'm the co chair. 0:04:07.760 --> 0:04:10.400 The other co chair is our AI Ethics Global Leader 0:04:10.480 --> 0:04:14.920 francescar Rossi, who's a well renowned researcher in AI. Ethics, 0:04:15.360 --> 0:04:18.000 so she comes with that research background. So we had 0:04:18.279 --> 0:04:21.159 a board in place, an AI ethics board in place 0:04:21.240 --> 0:04:24.840 before I stepped into this job, and there were a 0:04:24.839 --> 0:04:27.760 lot of great discussions among a lot of researchers and 0:04:27.800 --> 0:04:31.479 a lot of people that deeply understood the technology, but 0:04:31.560 --> 0:04:36.200 it didn't have decision making authority. It didn't have all stakeholders. 0:04:36.200 --> 0:04:39.480 Are many stakeholders across the business at the table, and 0:04:39.600 --> 0:04:41.880 so when I came into the job as a lawyer 0:04:41.920 --> 0:04:45.000 and as somebody with the corporate governance background, I was 0:04:45.040 --> 0:04:48.640 sort of tasked with building out the operational aspects of 0:04:48.680 --> 0:04:54.440 it to make it capable of implementing centralized decision making, 0:04:54.480 --> 0:04:58.200 to give it authority, to bring in those perspectives from 0:04:58.240 --> 0:05:02.599 across the business and from people with different focuses within 0:05:02.880 --> 0:05:06.920 the IBM corporation, lots of different backgrounds, and we have 0:05:07.400 --> 0:05:14.360 very robust conversations, and we also engage the individuals throughout 0:05:14.400 --> 0:05:19.040 IBM who either from an advocacy because they care very 0:05:19.120 --> 0:05:22.560 much about the topic, or they're working in the space 0:05:22.839 --> 0:05:26.880 individually and have thoughts around the topic, are doing projects 0:05:26.880 --> 0:05:29.279 in the space, want to publish in the space. We 0:05:29.360 --> 0:05:32.560 have a very organic way of having them be involved 0:05:32.600 --> 0:05:37.000 as well. Absolutely necessary to have that cross disciplinary aspect. 0:05:37.560 --> 0:05:42.640 You mentioned beginning your answer, talked book about robust conversations 0:05:42.760 --> 0:05:44.920 phrase I love. Can both of you give me an 0:05:44.960 --> 0:05:49.880 example of an issue that's come up with respect to 0:05:49.960 --> 0:05:55.000 trust and AI. So so, one example might be the 0:05:55.040 --> 0:05:58.960 technologies that we would employ as a company in response 0:05:59.000 --> 0:06:02.120 to the COVID nineteen pandemic. So there are a lot 0:06:02.120 --> 0:06:05.880 of things we could have done, and it became a 0:06:05.960 --> 0:06:10.159 question not of what we're capable of deploying from a 0:06:10.200 --> 0:06:15.839 technology perspective, but whether we should be deploying certain technologies, 0:06:15.920 --> 0:06:20.719 whether it be facial recognition for fever detection, certain contact 0:06:20.720 --> 0:06:24.680 tracing technologies are Digital health pass is a good example 0:06:25.240 --> 0:06:29.359 of a technology that came through the board multiple times 0:06:30.000 --> 0:06:33.080 in terms of like if we are going to deploy 0:06:33.360 --> 0:06:36.880 a vaccine passport, which is not necessarily what this technology 0:06:36.920 --> 0:06:39.240 turned out to be, but looking at things like immunity 0:06:39.279 --> 0:06:43.200 certificates and vaccine passports, not what could we do, but 0:06:43.279 --> 0:06:45.440 what were we willing as a company to do? Where 0:06:45.440 --> 0:06:48.359 were we going to put our skills and our knowledge 0:06:48.480 --> 0:06:53.040 and our company brand in response to technologies that could 0:06:53.040 --> 0:06:56.080 help to either bring about a cure or help to 0:06:56.600 --> 0:07:00.880 provide information in response to the pandemic. COVID is a 0:07:00.880 --> 0:07:05.000 great example because it highlights the value and the acceleration 0:07:05.080 --> 0:07:10.600 that good governance can bring. Because the way that we 0:07:10.680 --> 0:07:14.720 as an ethics board laid out the rules, the guardrails, 0:07:14.760 --> 0:07:17.840 if you will, around what we could would and wouldn't 0:07:17.880 --> 0:07:21.160 do for COVID help people just do stuff without worrying 0:07:21.680 --> 0:07:23.200 that we need to bring this to the board. It 0:07:23.240 --> 0:07:26.160 also laid very clear for this type of use case 0:07:26.200 --> 0:07:29.040 we need to go have a conversation with the board. 0:07:29.640 --> 0:07:32.840 It also provided a venue for us as a company 0:07:32.920 --> 0:07:39.040 to make decisions um and make risk based decisions where okay, 0:07:39.200 --> 0:07:41.760 this isn't a little bit of the of the fuzzy area, 0:07:41.880 --> 0:07:44.680 but we think, given what's going on right now in 0:07:44.800 --> 0:07:47.360 the world and the importance of this, we're willing to 0:07:47.360 --> 0:07:49.240 take this risk so long as we go back and 0:07:49.280 --> 0:07:52.440 we clean everything up later. And so so I think 0:07:52.520 --> 0:07:55.720 that's really important that number one, governance is set up 0:07:55.760 --> 0:07:59.400 so that it accelerates things, not stops them. And number two, 0:07:59.520 --> 0:08:02.760 that there's clear guidance into you know, it's not no, 0:08:03.280 --> 0:08:05.720 it's here's what you can do and here's what you 0:08:05.760 --> 0:08:08.680 can't do. And help the teams figure out how they 0:08:08.760 --> 0:08:11.520 can still move things forward in a way that doesn't 0:08:11.680 --> 0:08:17.120 infringe on our principles. Yeah, I want to sort of 0:08:17.120 --> 0:08:22.240 give this there's a concrete sense about how a concern 0:08:22.280 --> 0:08:27.480 about trust and transparency and such would guide what a 0:08:27.520 --> 0:08:32.360 technology company might do. Now a real example, So, if 0:08:32.400 --> 0:08:35.680 I want to make sure that people are wearing face 0:08:35.720 --> 0:08:37.880 masks and then just highlight that there is someone in 0:08:37.920 --> 0:08:41.360 this area that's not wearing a face mask and you're 0:08:41.400 --> 0:08:45.480 not identifying the person, I think we'd be okay with that. 0:08:46.160 --> 0:08:48.800 What we wouldn't be okay with that with is if 0:08:48.800 --> 0:08:51.760 they wanted to identify the person in a way that 0:08:51.840 --> 0:08:55.040 they did not consent to and that was very generic. 0:08:55.080 --> 0:08:57.160 So I'm going to go through a database of unknown 0:08:57.160 --> 0:08:59.200 people and I'm going to match them to this person, 0:09:00.840 --> 0:09:03.280 and so that would not be okay, and a fuzzy 0:09:03.320 --> 0:09:06.480 area would be you know, I'm going to match this 0:09:06.559 --> 0:09:08.679 to a known person, so I know this is an 0:09:08.679 --> 0:09:11.040 employee and I know this is him. This is something 0:09:11.080 --> 0:09:13.000 that we as a board would want to have a 0:09:13.040 --> 0:09:15.640 conversation with. If this employee is not wearing a mask, 0:09:16.200 --> 0:09:18.000 can I match them to a name or do I 0:09:18.040 --> 0:09:21.360 just send a security personnel over here because the employee 0:09:21.360 --> 0:09:24.080 is not wearing a mask. That's a harder I think, 0:09:24.200 --> 0:09:28.920 and that's a real world example that we face during COVID. Yeah, 0:09:29.400 --> 0:09:33.359 let's talk a little bit about diversity and shared responsibility 0:09:33.360 --> 0:09:38.400 as principles that matter in this world of AI. What 0:09:38.400 --> 0:09:41.440 do what do those terms mean as applied to AI, 0:09:41.520 --> 0:09:45.000 and what's the kind of practical effect of seeking to 0:09:45.080 --> 0:09:49.400 optimize those goals? You know, I think first of all, 0:09:49.480 --> 0:09:54.200 we need to have good representation of society doing the 0:09:54.240 --> 0:09:57.360 work that impacts society. So A, it's just the right 0:09:57.360 --> 0:10:01.080 thing to do. B. There's tons of research out there 0:10:01.120 --> 0:10:04.400 that shows that diverse teams outperformed non diverse teams. There's 0:10:04.400 --> 0:10:07.280 a Mackenzie report that says, you know, companies in the 0:10:07.320 --> 0:10:11.000 top quartile for diversity outperformed their peers that aren't by like, 0:10:12.480 --> 0:10:16.120 so tons of good research. The second thing is you 0:10:16.240 --> 0:10:18.559 just don't get as good results when you don't have 0:10:19.120 --> 0:10:22.080 equal representation at the table. There's lots of good examples 0:10:22.080 --> 0:10:25.680 of this. So there was a hiring algorithm that was 0:10:25.679 --> 0:10:30.400 evaluating applicants and passing forward, but all the applicants in 0:10:30.440 --> 0:10:32.560 the past for this company, you know, the vast majority 0:10:32.559 --> 0:10:35.520 of them were male, and so females were just summarily 0:10:35.520 --> 0:10:38.920 wiped out, regardless to some extent of their fit for 0:10:38.960 --> 0:10:43.440 the role. I wanted to ask Castina. A project comes 0:10:43.679 --> 0:10:49.720 before the board, and so a conversation might be the 0:10:49.720 --> 0:10:52.079 team you put together and the data you're looking at 0:10:52.960 --> 0:10:57.880 is insufficiently diverse, We're worried that you're not capturing the 0:10:57.960 --> 0:11:00.679 reality of the of the kind of world we're operating in. 0:11:01.120 --> 0:11:03.120 Is that is that an example of a conversation might 0:11:03.160 --> 0:11:06.800 you might have at the board level. Well, I think 0:11:06.840 --> 0:11:09.280 the best way to look at what the board is 0:11:09.320 --> 0:11:13.360 doing to try to address those issues of bias. I mean, so, 0:11:13.720 --> 0:11:17.400 for example, we've got a team of researchers that work 0:11:17.440 --> 0:11:21.520 on trusted technology, and one of the early things that 0:11:21.559 --> 0:11:26.240 they've done is to deploy toolkits that will help detect bias, 0:11:26.280 --> 0:11:28.400 that will help make a i AM more explainable, that 0:11:28.440 --> 0:11:32.040 will help make it trustworthy in general. But those tools 0:11:32.080 --> 0:11:35.839 initially very focused on bias, and they deployed them to 0:11:35.920 --> 0:11:39.079 open source so they could be built on and improved. 0:11:39.240 --> 0:11:43.320 Right and right now, the board is focused more broadly, 0:11:43.400 --> 0:11:47.080 not looking at it an individual problem in an individual 0:11:47.240 --> 0:11:50.400 use case with respect to bias, but instilling those ethical 0:11:50.440 --> 0:11:54.880 principles across the business through something we're calling ethics by Design. 0:11:55.480 --> 0:11:59.960 Bias was the first focus area of this Ethics by Design, 0:12:00.520 --> 0:12:03.120 and we've got a team of folks being led by 0:12:03.200 --> 0:12:06.400 the Ethics Board who are working on the question you 0:12:06.400 --> 0:12:10.160 asked Malcolm about, how do we ensure that the AI 0:12:10.200 --> 0:12:13.080 we're deploying internally or the tools and the products that 0:12:13.080 --> 0:12:17.320 we're deploying for customers take that into account throughout the 0:12:17.360 --> 0:12:20.520 life cycle of AI. So through this Ethics by Design, 0:12:20.960 --> 0:12:23.640 the guidance that's coming out from the Board starts at 0:12:23.679 --> 0:12:28.120 that conceptual phase and then applies across the life cycle 0:12:28.320 --> 0:12:32.280 up through in the case of an internal use of AI, 0:12:32.480 --> 0:12:34.840 up through the actual use and in the case of 0:12:34.840 --> 0:12:37.880 AI that we're deploying for customers are putting into a product, 0:12:37.880 --> 0:12:41.560 you not through that point of deployment. So it's very 0:12:41.640 --> 0:12:46.760 much about embedding those considerations into our existing processes across 0:12:46.800 --> 0:12:49.120 the company to make sure that they're thought of, not 0:12:49.200 --> 0:12:51.320 just once and not just in the use cases that 0:12:51.400 --> 0:12:55.080 the Board has an opportunity to review, but in our 0:12:55.080 --> 0:12:58.880 practices as a company and in our thinking as a company. 0:12:58.960 --> 0:13:01.040 Much like you know, we did this and companies did 0:13:01.080 --> 0:13:05.920 this years ago, um with respect to privacy and security, 0:13:06.000 --> 0:13:09.040 that concept of privacy and security by design which some 0:13:09.200 --> 0:13:11.520 may be familiar with that stem from the g d 0:13:11.640 --> 0:13:15.160 PR in Europe. Now we're doing the same thing with ethics. 0:13:15.160 --> 0:13:19.400 How unusual is what you guys are doing. I mean 0:13:19.440 --> 0:13:22.080 if I if I lined up all the tech companies 0:13:22.120 --> 0:13:24.760 that are heavily into AI right now, would I find 0:13:24.800 --> 0:13:27.800 similar programs in all of them? Or are you guys 0:13:28.760 --> 0:13:32.920 off by yourselves? So I think we take a little 0:13:32.960 --> 0:13:36.240 bit of a unique perspective. In fact, we were recently 0:13:36.240 --> 0:13:40.800 recognized as a leader in the ethical deployment of technology 0:13:40.800 --> 0:13:44.280 and responsible technology use by the World Economic Forum. So 0:13:44.440 --> 0:13:48.680 World Economic Forum and the Marcola Center Um of Ethics 0:13:48.720 --> 0:13:51.680 at at Santa Clara University did an independent case study 0:13:51.920 --> 0:13:55.640 of IBM that did recognize our leadership in this space. 0:13:55.679 --> 0:13:58.280 Because of the holistic approach that we take, we're a 0:13:58.280 --> 0:14:01.160 little bit different I think in some other tech companies 0:14:01.200 --> 0:14:05.800 that do have similar counsels in place because of the 0:14:06.040 --> 0:14:09.720 broad and cross disciplinary nature of ours. We're not just researchers, 0:14:09.720 --> 0:14:15.439 were not just technologists. We literally have representation from backgrounds 0:14:15.679 --> 0:14:18.720 spanning across the company, whether it be you know, legal 0:14:18.800 --> 0:14:22.440 or developers or researchers or or you know just HR 0:14:22.600 --> 0:14:25.680 professionals and the like. So that makes us a little 0:14:25.680 --> 0:14:28.600 bit unique the program itself. And then I think we 0:14:28.680 --> 0:14:32.880 hear from clients that are thinking for themselves about how 0:14:32.920 --> 0:14:35.880 do I make sure that the technology I'm deploying or 0:14:36.040 --> 0:14:42.840 using externally or with my clients is trustworthy? Right, So 0:14:42.840 --> 0:14:45.040 so they're asking us, how did you go about this, 0:14:45.800 --> 0:14:47.840 how do you think about it as a company, what 0:14:47.880 --> 0:14:52.880 are your practices? So on that point, we are CEO 0:14:53.560 --> 0:14:57.080 is the co chair of a something called the Global 0:14:57.120 --> 0:15:01.280 AI Action Alliance initiated by the WEF, and as part 0:15:01.320 --> 0:15:05.080 of that, we've committed to sort of open source our approach. 0:15:05.160 --> 0:15:07.240 So we've been talking a lot about our approach. I 0:15:07.280 --> 0:15:09.480 think it is a little bit unique, as I said, 0:15:09.520 --> 0:15:12.160 but we are sharing it because again, we don't want 0:15:12.160 --> 0:15:14.920 to be the only ones that have trustworthy the AI 0:15:15.000 --> 0:15:18.120 and that have this holistic, cross disciplinary approach, because we 0:15:18.160 --> 0:15:21.040 think it's the right approach. It's certainly the right approach 0:15:21.080 --> 0:15:22.880 for our company, and we want to share it with 0:15:22.920 --> 0:15:36.960 the world. It's not secret or proprietary, but if you 0:15:37.080 --> 0:15:40.800 talk to the analyst community that serves the tech the 0:15:40.840 --> 0:15:44.240 tech you know, the tech sector. They say far and wide, 0:15:44.240 --> 0:15:47.240 IBM is is is ahead in terms of things that 0:15:47.280 --> 0:15:51.760 we're actually doing as opposed to talking about it all 0:15:51.840 --> 0:15:56.440 while making sure that it is enforceable and impactful. So 0:15:56.800 --> 0:15:59.680 for instance, you know we were talking about we review 0:15:59.760 --> 0:16:04.640 you cases and we can require that the teams adjust them. 0:16:04.680 --> 0:16:07.680 That's unique, right, Most of the other tech companies do 0:16:07.760 --> 0:16:11.240 not have that level of oversight in terms of ensuring 0:16:11.320 --> 0:16:14.480 that their their outcomes are are aligned. There's a lot 0:16:14.520 --> 0:16:17.120 of good talk, but I think you know, the weft 0:16:17.520 --> 0:16:19.120 use case that came out on I think it was 0:16:19.560 --> 0:16:23.440 September really supports that that we're ahead. And then if 0:16:23.440 --> 0:16:26.880 you look at companies just in general that have AI 0:16:26.960 --> 0:16:29.520 ethics board. My experiences that with all the companies and 0:16:29.560 --> 0:16:33.880 I know interact with hundreds of of leaders and companies 0:16:33.920 --> 0:16:37.560 a year, less than five percent of them have a 0:16:37.680 --> 0:16:41.800 board in place, and even fewer of those kind of 0:16:42.320 --> 0:16:45.640 really have a rhythm going and know how they're gonna 0:16:46.000 --> 0:16:49.680 going to operate as a board. Yet m hmm. I 0:16:49.680 --> 0:16:51.160 wanted to talk a little bit about the rule of 0:16:51.200 --> 0:16:59.400 government here, these government leading or following here. UM, I 0:16:59.400 --> 0:17:03.360 would say they're catching up. I think we're following is 0:17:03.400 --> 0:17:08.720 probably the most improved, right because look, I think over 0:17:08.800 --> 0:17:12.480 the last uh couple of years, as we talked about, 0:17:12.600 --> 0:17:15.320 or maybe it's been almost ten years at this point 0:17:15.359 --> 0:17:18.360 in time, as these issues have come to light, companies 0:17:18.400 --> 0:17:22.960 have largely been left to themselves to impose guardrails upon 0:17:23.240 --> 0:17:26.000 their practices and their use of AI. Let's not be 0:17:26.280 --> 0:17:29.680 to say that there aren't laws that regulate for example, 0:17:29.840 --> 0:17:36.119 discrimination laws do would apply to technology that's discriminatory, but 0:17:36.280 --> 0:17:39.160 the unique aspects, to the extent there are unique aspects 0:17:39.240 --> 0:17:45.240 or issues that get amplified through the application of AI systems. Um, 0:17:45.280 --> 0:17:47.880 the government is really just catching up. So we've got 0:17:47.920 --> 0:17:53.119 the the EU proposed a comprehensive regulatory framework for AI 0:17:53.640 --> 0:17:56.520 in the spring time frame. UM. We see in the 0:17:56.640 --> 0:18:00.960 US the FTC is starting to focus on and algorithmic 0:18:01.000 --> 0:18:04.600 bias and just in general on algorithms and that they 0:18:04.640 --> 0:18:07.120 be fair and the like. So there are numerous other 0:18:07.160 --> 0:18:13.159 initiatives following the EU that are looking at frameworks for 0:18:13.240 --> 0:18:17.760 governing AI and regulating AI, and we've been involved, I 0:18:17.800 --> 0:18:22.159 mentioned earlier on our Precision Regulation recommendation. So we have 0:18:22.240 --> 0:18:26.200 something called the IBM Policy Lab, and what differentiates our 0:18:26.240 --> 0:18:29.760 advocacy through the Policy Lab is that we try to 0:18:29.800 --> 0:18:36.480 make concrete, actionable policy recommendations, so not just again articulating principles, 0:18:36.520 --> 0:18:40.919 but really concrete recommendations for companies and for governments and 0:18:41.000 --> 0:18:45.800 policymakers around the globe to implemented to follow um things 0:18:45.880 --> 0:18:48.960 like you know, out of our precision regulation of AI. 0:18:49.080 --> 0:18:53.040 That's where our recommendation is that regulation should be risk based, 0:18:53.160 --> 0:18:55.960 it should be context specific. It should look at and 0:18:56.040 --> 0:19:00.960 allocate responsibility on the party that's closest to the risk, 0:19:01.040 --> 0:19:03.199 and that may be different at different times in the 0:19:03.280 --> 0:19:06.239 life cycle of an AI system. So we deploy some 0:19:06.320 --> 0:19:11.200 general purpose technologies and then our clients train those over time, 0:19:11.400 --> 0:19:13.480 so you know, bearing the risk it should sit with 0:19:13.600 --> 0:19:15.760 the party that's closest to the risk at the different 0:19:15.800 --> 0:19:19.199 points in time in the AI life cycle. You know, 0:19:20.240 --> 0:19:26.560 one of the interesting things about this issue today, we're 0:19:26.600 --> 0:19:31.360 now in a situation where someone like IBM, I'm guessing 0:19:31.800 --> 0:19:35.000 is that it would be as sensitive to public reaction 0:19:35.320 --> 0:19:37.600 to the uses of AI as they would be to 0:19:37.680 --> 0:19:40.640 government reaction to the uses of AI. And I wanted 0:19:40.640 --> 0:19:43.960 to just way those you know, is that this is 0:19:44.000 --> 0:19:47.240 a this kind of fascinating development in our age that 0:19:47.720 --> 0:19:52.240 all of a sudden it almost seems like whatever form 0:19:52.320 --> 0:19:55.960 public reaction takes can be a more powerful lever of 0:19:55.960 --> 0:20:00.600 of in moving changing corporate behavior. Then what government are saying? 0:20:00.800 --> 0:20:02.399 And do you do you think this is true in 0:20:02.440 --> 0:20:06.360 this AI space? I think the government regulation that we're 0:20:06.359 --> 0:20:11.000 seeing is responding to public sentiment. So I agree with 0:20:11.040 --> 0:20:14.560 you a hundred percent that that this is being moved 0:20:14.560 --> 0:20:17.320 by the public. And you know, oftentimes when we have 0:20:17.400 --> 0:20:21.840 conversations at the ethics board, Okay Christina and the lawyers 0:20:21.840 --> 0:20:23.679 say okay, this is not a legal issue, then the 0:20:23.720 --> 0:20:26.600 next conversation is what happens if this story shows up 0:20:26.600 --> 0:20:28.080 on the front page of the New York Times of 0:20:28.119 --> 0:20:32.920 the Wall Street Journal. So so absolutely we consider that. 0:20:34.359 --> 0:20:35.920 So I would also I would add to that, like 0:20:36.119 --> 0:20:40.360 we've been well, probably think the oldest technology company, we're 0:20:40.640 --> 0:20:43.840 over a hundred years old, and our clients have looked 0:20:43.840 --> 0:20:48.359 to us for that hundred plus years to responsibly usher 0:20:48.400 --> 0:20:52.320 in new technologies right and to manage their data, their 0:20:52.400 --> 0:20:55.280 most sensitive data, in a trusted way. So for us, 0:20:55.359 --> 0:20:58.640 it's it's not just about the the headline risk. It's 0:20:58.640 --> 0:21:01.240 about ensuring that we have of business going forward because 0:21:01.240 --> 0:21:06.960 our clients trust us UM and society trusts US. So 0:21:07.320 --> 0:21:11.560 the guardrails we put in place, particularly around the trust 0:21:11.600 --> 0:21:14.159 and Transparency principles, or the guard rails we put in 0:21:14.160 --> 0:21:17.040 place around responsible data use. In the COVID pandemic, there 0:21:17.119 --> 0:21:21.639 was nothing that from a legal perspective, said we couldn't 0:21:21.680 --> 0:21:24.600 do more. There was nothing that said in the US 0:21:24.680 --> 0:21:29.119 we can't use facial recognition technology and our sites. But 0:21:29.359 --> 0:21:34.000 we made principal decisions, and we made those decisions because 0:21:34.040 --> 0:21:36.800 we think they're the right decisions to make. And when 0:21:36.840 --> 0:21:40.440 I look back at the Ethics Board and the analysis 0:21:40.960 --> 0:21:42.879 and the use cases that have come forward over the 0:21:42.880 --> 0:21:46.359 course of the last two years, I can think of 0:21:46.480 --> 0:21:49.600 very few where we said we're not going to do 0:21:49.680 --> 0:21:54.640 this because we're afraid of regulatory repercussions. UM. In fact, 0:21:54.680 --> 0:21:57.879 I can't think of any where because it wouldn't have 0:21:57.920 --> 0:22:00.160 come to the board if it was a league all. 0:22:00.960 --> 0:22:06.040 But yet we did refine in some cases stop I 0:22:07.480 --> 0:22:12.840 actual transactions right and solutions because we felt they were 0:22:12.880 --> 0:22:18.159 not the right thing to do. Yeah, yeah, A question 0:22:18.200 --> 0:22:20.800 for either of you, can you can you dig a 0:22:20.840 --> 0:22:24.239 little more into this, into the real world applications of this. 0:22:25.040 --> 0:22:27.200 What are some of the very kind of concrete kinds 0:22:27.200 --> 0:22:29.960 of things that come out of this focus on untrust. 0:22:32.359 --> 0:22:35.680 So so you know, some some real world examples of 0:22:35.880 --> 0:22:39.880 how trust plays into what we're doing. Is gets back 0:22:39.920 --> 0:22:43.360 to a couple of things Christina said earlier around how 0:22:43.359 --> 0:22:45.320 we're open sourcing a lot of what we do. So 0:22:45.800 --> 0:22:50.159 our research division builds a lot of the technology that 0:22:50.200 --> 0:22:54.359 winds up in our products um uh. And then, particularly 0:22:54.359 --> 0:22:57.760 related to this topic of AI ethics and trustworthy AI 0:22:59.200 --> 0:23:01.720 are the fall is to open source the base of 0:23:01.760 --> 0:23:04.560 the technology. So we have a whole bunch of open 0:23:04.600 --> 0:23:08.280 source tool kits um that anyone can use. In fact, 0:23:08.359 --> 0:23:11.199 some of our competitors use them as much as we 0:23:11.280 --> 0:23:14.760 do in their products. And then we build value adds 0:23:14.880 --> 0:23:18.159 on top of those and so that is something that 0:23:18.200 --> 0:23:22.280 we advocate strongly for in the Ethics Board helps support 0:23:22.359 --> 0:23:25.359 us with that, as do you know, our our product teams, 0:23:25.440 --> 0:23:30.480 because the value is you know, AI is one of 0:23:30.480 --> 0:23:34.720 those spaces where when something goes wrong, it affects everyone, right, 0:23:34.840 --> 0:23:38.040 so if if there's a big issue with AI, everyone's 0:23:38.040 --> 0:23:40.560 going to be concerned about all AI, and so we 0:23:40.600 --> 0:23:45.479 want to make sure that the technology behind AI is 0:23:45.520 --> 0:23:49.960 as fair as possible, is as explainable as possible, is 0:23:49.960 --> 0:23:54.000 as robust as possible, and is as privacy preserving as possible. 0:23:54.040 --> 0:23:56.800 So tool kits that address those are all publicly available, 0:23:57.200 --> 0:23:59.760 and then we build value added capabilities on top of 0:23:59.800 --> 0:24:02.320 that when we set when we bring those things to 0:24:02.320 --> 0:24:05.720 our customers in the form of an integrated platform that 0:24:05.800 --> 0:24:08.439 helps manage the whole life cycle of an AI. Because 0:24:09.040 --> 0:24:13.280 AI is different than software in that the technology under 0:24:13.359 --> 0:24:16.320 AI is machine learning. What that means is that the 0:24:16.359 --> 0:24:19.879 machine keeps learning over time and adjusting the model over time. 0:24:20.640 --> 0:24:23.240 Once you write a piece of software, it's done, it 0:24:23.280 --> 0:24:26.080 doesn't change. And so you need to figure out how 0:24:26.119 --> 0:24:29.800 do you continuously monitor your your AI over time for 0:24:30.040 --> 0:24:33.560 those things I just described and integrate them into your 0:24:33.920 --> 0:24:37.600 security and privacy by design practices so that they're continuously 0:24:38.400 --> 0:24:42.160 updating and aligned to your company's principles as well as 0:24:42.680 --> 0:24:48.880 societal principles as well as any relevant regulations. Yeah, when 0:24:48.960 --> 0:24:56.040 this question, give me one suggestion prediction about what AI 0:24:56.119 --> 0:25:01.399 looks like five years or ten years for now. Yeah, So, 0:25:01.400 --> 0:25:05.320 so that is a really really good question. And you know, 0:25:05.960 --> 0:25:09.760 when we when we look at what AI does today, AI, 0:25:10.080 --> 0:25:14.200 while it's very insightful, and it helps us realize things 0:25:14.200 --> 0:25:16.080 that as humans we may not have picked up on 0:25:16.119 --> 0:25:19.800 our own. And so to augment our intelligence, A surfaces 0:25:19.840 --> 0:25:24.240 insights and maybe reduce as a complexity from almost infinite 0:25:24.280 --> 0:25:27.920 and comprehensible to humans. Two, I have five choices now 0:25:27.960 --> 0:25:30.159