00:00:02.685 --> 00:00:03.170 [MUSIC] 00:00:03.169 --> 00:00:06.800 KATHLEEN SULLIVAN: Welcome  to AI Testing and Evaluation:   00:00:06.800 --> 00:00:11.520 Learnings from Science and Industry.  I'm your host, Kathleen Sullivan. 00:00:11.520 --> 00:00:16.640 As generative AI continues to advance, Microsoft  has gathered a range of experts—from genome   00:00:16.640 --> 00:00:21.040 editing to cybersecurity—to share how  their fields approach evaluation and   00:00:21.040 --> 00:00:26.160 risk assessment. Our goal is to learn from  their successes and their stumbles to move   00:00:26.160 --> 00:00:31.360 the science and practice of AI testing  forward. In this series, we'll explore   00:00:31.360 --> 00:00:41.266 how these insights might help guide the future of  AI development, deployment, and responsible use. 00:00:41.266 --> 00:00:41.280 [MUSIC ENDS] 00:00:41.280 --> 00:00:45.760 For our introductory episode, I'm pleased to  welcome Amanda Craig Deckard from Microsoft   00:00:45.760 --> 00:00:49.920 to discuss the company's efforts to  learn about testing in other sectors. 00:00:49.920 --> 00:00:54.320 Amanda is senior director of public  policy in the Office of Responsible AI,   00:00:54.320 --> 00:00:57.840 where she leads a team that works  closely with engineers, researchers,   00:00:57.840 --> 00:01:03.680 and policy experts to help ensure AI is being  developed and used responsibly. Their insights   00:01:03.680 --> 00:01:10.400 shape Microsoft's contribution to public policy  discussions on laws, norms, and standards for AI. 00:01:10.400 --> 00:01:12.538 Amanda, welcome to the podcast. 00:01:12.538 --> 00:01:13.160 AMANDA CRAIG DECKARD: Thank you. 00:01:13.160 --> 00:01:17.440 SULLIVAN: Amanda, let's give the listeners  a little bit of your background. What's your   00:01:17.440 --> 00:01:22.240 origin story? Can you talk to us a little bit  about maybe how you started in tech? And I would   00:01:22.240 --> 00:01:27.585 love to also learn a little bit more about what  your team does in the Office of Responsible AI. 00:01:27.585 --> 00:01:36.240 CRAIG DECKARD: Sure. Thank you. I'd say my  [LAUGHS] path to tech, to Microsoft, as well,   00:01:36.240 --> 00:01:42.160 was a bit, like, circuitous, maybe. You know,  I thought for the longest time I was going to   00:01:42.160 --> 00:01:52.720 be a journalist. I studied forced migration. I  worked in a sort of state level sort of trial   00:01:52.720 --> 00:01:57.920 court in Indiana, a legal service provider  in India, just to give you a bit of a flavor. 00:01:57.920 --> 00:02:06.160 I made my way to Microsoft in 2014 and have been  here since, working in cybersecurity public policy   00:02:06.160 --> 00:02:13.440 first and now in responsible AI. And the way  that our Office of Responsible AI has really,   00:02:13.440 --> 00:02:19.360 sort of, structured itself is bringing  together the kind of expertise to really   00:02:19.360 --> 00:02:24.240 work on defining policy and how to  operationalize it at the same time. 00:02:24.960 --> 00:02:29.680 And, you know, that means that we have  been working through this, you know,   00:02:29.680 --> 00:02:37.920 real challenge of defining internal policy and  practice, making sure that's deeply grounded in   00:02:37.920 --> 00:02:43.040 the work of our colleagues at Microsoft Research,  and then really closely working with engineering   00:02:43.040 --> 00:02:48.000 to make sure that we have the processes, that we  have the tools, to implement that policy at scale. 00:02:48.000 --> 00:02:54.400 And I'm really drawn to these kind of hard  problems where they have the character of two   00:02:54.400 --> 00:03:00.240 things being true or there's like, you know,  real tension on both sides and in particular,   00:03:00.240 --> 00:03:05.600 in the context of those kinds of problems, roles  in which, like, the whole job is actually just   00:03:05.600 --> 00:03:10.320 sitting with that tension, not necessarily, like,  resolving it and expecting that you're done. 00:03:10.320 --> 00:03:17.600 And I think, really, there are two reasons why  tech is so, kind of, representative of that kind   00:03:17.600 --> 00:03:22.320 of challenge that I've always found fascinating.  You know, one is that, of course, tech is,   00:03:22.320 --> 00:03:28.480 sort of, ubiquitous. It's really impacting so  many people's lives. But also, you know, because,   00:03:28.480 --> 00:03:34.560 as I think has become part of our vernacular now,  but, you know, is not necessarily immediately   00:03:34.560 --> 00:03:38.640 intuitive, is like the fact that technology is  both a tool and a weapon. And so that's just,   00:03:38.640 --> 00:03:43.520 like, another reason why, you know, we have  to continuously work through that tension and,   00:03:43.520 --> 00:03:47.960 sort of, like, sit with it, right,  and even as tech evolves over time. 00:03:47.960 --> 00:03:52.000 SULLIVAN: You bring up such great points, and  this field is not black and white. I think that   00:03:52.000 --> 00:03:56.800 even underscores, you know, this notion that you  highlighted that it's impacting everyone. And,   00:03:56.800 --> 00:04:02.080 you know, to set the stage for our listeners,  last year, we pulled in a bunch of experts   00:04:02.080 --> 00:04:08.720 from cybersecurity, biotech, finance, and we  ran this large workshop to study how they're   00:04:08.720 --> 00:04:12.240 thinking about governance and those playbooks.  And so I'd love to understand a little bit more   00:04:12.240 --> 00:04:17.200 about what sparked that effort—and, you  know, there's a piece of this which is   00:04:17.200 --> 00:04:23.905 really centered around testing—and to hear from  you why the focus on testing is so important. 00:04:23.905 --> 00:04:28.480 CRAIG DECKARD: If I could rewind a little bit and  give you a bit of history of how we even arrived   00:04:28.480 --> 00:04:34.640 at bringing these experts together, you know,  we actually started on this journey in 2023.   00:04:34.640 --> 00:04:42.240 At that time, there were, like, a lot of  these big questions swirling around about,   00:04:42.240 --> 00:04:45.360 you know, what did we need in terms  of governance for AI? Of course,   00:04:45.360 --> 00:04:51.840 this was in the immediate aftermath of the ChatGPT  sort of wave and everyone recognizing that, like,   00:04:51.840 --> 00:04:56.560 the technology was going to have a different level  of impact in the near term. And so, you know,   00:04:56.560 --> 00:05:00.960 what do we need from governance? What do we need  at the global level, in particular, of governance? 00:05:00.960 --> 00:05:07.840 And so at the time, in early 2023 especially,  there were a lot of attempts to sort of draw   00:05:07.840 --> 00:05:14.800 analogies to other global governance institutions  in other domains. So we actually in 2023 brought   00:05:14.800 --> 00:05:18.720 together a different workshop than the one  that you're referring to specifically focused   00:05:18.720 --> 00:05:26.240 on testing last year. And we, kind of, had  two big takeaways from that conversation. 00:05:26.240 --> 00:05:33.680 One was, what are the actual functions of these  institutions and how do they apply to AI? And,   00:05:33.680 --> 00:05:37.280 actually, one of the takeaways was they  all sort of apply. [LAUGHS] There's,   00:05:37.280 --> 00:05:42.480 like, a role for, you know, any of  the functions, whether it be sort of   00:05:42.480 --> 00:05:47.200 driving consensus on research or building  industry standards or managing, kind of,   00:05:47.200 --> 00:05:52.000 frontier risks, for thinking about how  those might be needed in the AI context. 00:05:52.000 --> 00:05:56.000 And one of the other big takeaways  was that, you know, there are also   00:05:56.000 --> 00:06:04.480 limitations in these analogies. You know, each  of the institutions grew up in its own, sort of,   00:06:04.480 --> 00:06:10.640 unique historical moment, like the one that  we sit in with AI right now. And in each of   00:06:10.640 --> 00:06:15.200 those circumstances, they don't exactly  translate to this moment. And so, yeah,   00:06:15.200 --> 00:06:20.720 there was like this kind of, OK, we want to  draw what we can from this conversation and   00:06:20.720 --> 00:06:27.280 then we also want to understand, what is also very  important that's just different for AI right now? 00:06:27.280 --> 00:06:32.960 We published a book with the lessons  from that conversation in 2023. And then   00:06:32.960 --> 00:06:38.240 we actually went on a bit of a tour  [LAUGHS] with that content where we   00:06:38.240 --> 00:06:43.040 had a number of roundtables actually all  over the world where we gathered feedback   00:06:43.040 --> 00:06:48.800 on how those analogies were landing, how our  takeaways were landing. And one of the things   00:06:48.800 --> 00:06:55.680 that we took from them was a gap that some of the  participants saw in the analogies that we chose to   00:06:55.680 --> 00:07:02.720 focus on. So across multiple conversations, other  domains kept being raised, like, why did you not   00:07:02.720 --> 00:07:09.920 also study pharmaceuticals? Why did you also not  study cybersecurity, for example? And so that,   00:07:09.920 --> 00:07:15.520 you know, naturally got us thinking about what  further lessons we could draw from those domains. 00:07:15.520 --> 00:07:18.400 At the same time, though, we also saw a need to,   00:07:18.400 --> 00:07:23.920 again, go deeper than what we went and  really, like, focus on a narrower problem. 00:07:23.920 --> 00:07:27.680 So that's really what led us to trying to  think about a more specific problem where we   00:07:27.680 --> 00:07:33.040 could think across levels of governance and  bring in some of these other domains. And,   00:07:33.040 --> 00:07:39.280 you know, testing was top of mind. Continues  to be a really important topic in the AI policy   00:07:39.280 --> 00:07:44.880 conversation right now, I think, for really good  reason. A lot of policymakers are focused on,   00:07:44.880 --> 00:07:50.640 you know, what we need to do to, kind  of, have there be sufficient trust,   00:07:50.640 --> 00:07:55.760 and testing is going to be a part of  that—really better understand risk,   00:07:56.480 --> 00:08:01.040 enable everyone to be able to make more, kind  of, risk-informed decisions, right. Testing is   00:08:01.040 --> 00:08:07.440 an important component for governance and AI and,  of course, in all of these other domains, as well. 00:08:07.440 --> 00:08:16.080 So I'll just add the other, kind of, input into  the process for this second round was exploring   00:08:16.880 --> 00:08:25.200 other analogies beyond those that we, kind of,  got feedback on. And one of the early, kind of,   00:08:26.080 --> 00:08:30.320 examples of another domain that would be really  worthwhile to study that came to mind from,   00:08:30.320 --> 00:08:34.400 sort of, just studying the literature was  genome editing. You know, genome editing   00:08:34.400 --> 00:08:37.920 was really interesting through the process of  thinking about other kind of general-purpose   00:08:37.920 --> 00:08:43.240 technologies. We also arrived at nanoscience  and brought those into the conversation. 00:08:43.240 --> 00:08:46.000 SULLIVAN: That's great. I mean,  actually, if you could double-click,   00:08:46.000 --> 00:08:50.160 I mean, you just named a number of  industries. I'd love to just understand   00:08:50.160 --> 00:08:54.000 which of those worlds maybe feels the  closest to what we're wrestling with,   00:08:54.000 --> 00:09:00.065 with AI and maybe which is kind of the farthest  off, and what makes them stand out to you? 00:09:00.065 --> 00:09:02.320 CRAIG DECKARD: Oh, such a good  question. For this second round,   00:09:02.320 --> 00:09:07.440 we actually brought together eight different  domains, right. And I think we actually thought   00:09:07.440 --> 00:09:13.280 we would come out of this conversation with some  bit of clarity around, Oh, if we just, sort of,   00:09:13.280 --> 00:09:20.160 take this approach for this domain or that  domain, we'll sort of have—at least for now—really   00:09:20.160 --> 00:09:27.200 solved part of the puzzle. [LAUGHS] And, you know,  our public policy team the day after the workshop,   00:09:27.200 --> 00:09:32.320 we had a, sort of, follow-on discussion,  and the very first thing that we started   00:09:32.320 --> 00:09:37.040 with in that conversation was like, OK, so  which of these domains? And fascinatingly,   00:09:37.040 --> 00:09:43.280 like, everyone was sort of like, Ahh! [LAUGHS]  None of them are applying perfectly. I mean,   00:09:43.280 --> 00:09:46.720 this is also speaking to the limitations  of analogies that we already acknowledged. 00:09:47.440 --> 00:09:53.360 And also, you know, all of the experts  from across these domains gave us really   00:09:53.360 --> 00:09:59.120 interesting insights into, sort of, the  tradeoffs and the limitations and how they   00:09:59.120 --> 00:10:05.040 were working. None are really applying  perfectly for us. But all of them do   00:10:05.040 --> 00:10:10.880 offer a thread of insight that is really  useful for thinking about testing in AI,   00:10:10.880 --> 00:10:16.960 and there are some different dimensions that  I think are really useful as framing for that. 00:10:16.960 --> 00:10:21.040 I mean, one is just this  horizontal-versus-vertical,   00:10:21.040 --> 00:10:27.040 kind of, difference in domains and, you know,  the horizontal technology like genome editing   00:10:27.040 --> 00:10:37.200 or nanoscience just being inherently different and  seemingly very similar to AI in that you want to   00:10:37.200 --> 00:10:46.080 be able to understand risks in the technology  itself and there is just so much contextual,   00:10:46.080 --> 00:10:52.640 sort of, factor that matters in the application  of those technologies for how the risk manifests   00:10:52.640 --> 00:10:58.400 that you really need to, kind of, do those  two things at once—of understanding the   00:10:58.400 --> 00:11:04.560 technology but then really thinking about risk and  governance in the context of application versus,   00:11:04.560 --> 00:11:11.120 you know, a context like or a domain like civil  aviation or nuclear technology, for example. 00:11:11.120 --> 00:11:15.680 You know, even in the workshop  itself that we hosted late last year,   00:11:15.680 --> 00:11:22.000 where we brought together this second round of  experts, it was really interesting. We actually   00:11:22.000 --> 00:11:27.440 started the conversation by trying to understand  how those different domains defined risks,   00:11:27.440 --> 00:11:32.800 where they were able to set risk thresholds.  That's been such a part of the AI policy   00:11:32.800 --> 00:11:39.520 conversation in the last year. And, you know,  it was really instructive that the more vertical   00:11:39.520 --> 00:11:44.720 domains were able to, sort of, snap to clearer  answers much more quickly.[LAUGHS] But, like,   00:11:44.720 --> 00:11:50.560 the horizontal nanoscience and genome editing were  not because it just depends, right. So anyway,   00:11:50.560 --> 00:11:55.920 the horizontal-vertical dimension seems like a  really important one to draw from and apply to AI. 00:11:55.920 --> 00:12:00.960 The couple of others that I would offer is just,  you know, thinking about the different kinds of   00:12:00.960 --> 00:12:05.840 technologies. You know, obviously, there's some  of the domains that we studied that they're just   00:12:05.840 --> 00:12:11.360 inherently, sort of, like, physical technologies  … a mix of physical and digital or virtual in a   00:12:11.360 --> 00:12:15.680 lot of cases because all of these are, of course,  applying digital technology. But like, you know,   00:12:15.680 --> 00:12:21.120 there is just a difference between something like  an airplane or a medical device or, you know,   00:12:21.120 --> 00:12:27.760 the more kind of virtual or intangible sort of  technologies even, you know, of course, AI and   00:12:27.760 --> 00:12:32.800 some of the other like cyber and genome editing  but also like, you know, financial services having   00:12:32.800 --> 00:12:38.160 some of that quality. And again, I think the thing  that's interesting to us about AI is to think   00:12:38.160 --> 00:12:44.400 about AI and risk evaluation of AI as being, you  know, having a large component of that being about   00:12:44.400 --> 00:12:49.440 the kind of virtual or intangible technology.  And also, you know, there is a future of robotics   00:12:49.440 --> 00:12:55.440 where we might need to think about the, kind of,  physical risk evaluation kind of work, as well. 00:12:55.440 --> 00:13:02.560 And then the final thing I'd maybe say in terms of  thinking about which domains have the lessons for   00:13:02.560 --> 00:13:08.080 AI that are most applicable is just how they've  grappled with these different kind of governance   00:13:08.080 --> 00:13:17.120 questions. Things like how to turn the dial  in terms of being more or less prescriptive on   00:13:17.120 --> 00:13:23.360 risk evaluation approaches, how they think  about the balance of, kind of, pre-market versus   00:13:23.360 --> 00:13:29.120 post-market risk evaluation in testing, and what  the tradeoffs have been there across domains has   00:13:29.120 --> 00:13:34.720 been really interesting to kind of tease out. And  then also thinking about, sort of, who does what? 00:13:34.720 --> 00:13:39.840 So, you know, in each of these different domains,  it was interesting to hear about, like, you know,   00:13:39.840 --> 00:13:48.480 the role of industry, the role of governments,  the role of third-party experts in designing   00:13:48.480 --> 00:13:54.240 evaluations and developing standards and  actually doing the work, and, kind of,   00:13:54.240 --> 00:14:00.480 having the pull through of what it means for risk  and governance decisions. There were, again, there   00:14:00.480 --> 00:14:05.360 was a variety of, sort of, approaches across these  domains that I think were interesting for AI. 00:14:05.360 --> 00:14:09.120 SULLIVAN: You mentioned that there's  a number of different stakeholders to   00:14:09.120 --> 00:14:12.480 be considering across the board  as we're thinking about policy,   00:14:12.480 --> 00:14:17.200 as we're thinking about regulation. Where  can we collaborate more across industry?   00:14:17.200 --> 00:14:22.240 Is it academia? Regulators? Just,  how can we move the needle faster? 00:14:22.240 --> 00:14:28.000 CRAIG DECKARD: I think all of the above  [LAUGHTER] is needed. But it's also really   00:14:28.000 --> 00:14:35.200 important to have all of that, kind of, expertise  brought together, you know, and I think, you know,   00:14:35.200 --> 00:14:44.400 one of the things that we certainly heard from  multiple of the domains, if not all of them, was   00:14:44.400 --> 00:14:50.960 that same actual interest and need and the same  sort of ongoing work to try to figure that out. 00:14:52.000 --> 00:14:58.640 You know, even where there had been progress in  some of the other domains with bringing together,   00:14:58.640 --> 00:15:05.920 you know, some industry stakeholders  or, you know, industry and government,   00:15:05.920 --> 00:15:10.000 there was still a desire to actually do more  there. Like, if there was some progress in   00:15:10.000 --> 00:15:16.160 industry and government, the need was, And  more kind of cross-jurisdiction government   00:15:16.160 --> 00:15:21.760 conversation, for example. Or some progress on,  you know, within the industry but needing to,   00:15:21.760 --> 00:15:26.720 like, strengthen the partnership with academia,  for example. So, you know, I think it speaks to,   00:15:26.720 --> 00:15:31.200 like, the quality of your question, to be  honest, that, you know, all of these domains   00:15:31.200 --> 00:15:36.240 are actually still grappling with this and still  seeing the need to grow in that direction more. 00:15:36.240 --> 00:15:43.360 What I'd say about AI today is that we have made  good progress with, you know, starting to build   00:15:43.360 --> 00:15:49.360 some industry partnerships. You know, we were  a founding member of the Frontier Model Forum,   00:15:49.360 --> 00:15:54.960 or FMF, which has been a very useful place for  us to work with some peers on really trying   00:15:54.960 --> 00:16:00.000 to bring forward some best practices that  apply across our organizations. You know,   00:16:00.000 --> 00:16:04.640 there are other forums as well, like MLCommons,  where we're working with others in industry and   00:16:04.640 --> 00:16:09.120 broader, sort of, academic and civil society  communities. Partnership on AI is another   00:16:09.120 --> 00:16:14.720 one I think about that, kind of, fits that  mold, as well, in a really positive way. And,   00:16:14.720 --> 00:16:20.880 like, there are a lot of different, sort of,  governance needs to think through and where,   00:16:20.880 --> 00:16:25.280 you know, we can really think about bringing that  expertise together is going to be so important. 00:16:25.280 --> 00:16:29.680 I think about almost, like,  in the near to mid-term,   00:16:29.680 --> 00:16:36.480 like three issues that we need to address in  the AI, kind of, policy and testing context.   00:16:36.480 --> 00:16:41.120 One is just building kind of, like,  a flexible framework that allows us   00:16:41.120 --> 00:16:48.800 to really build trust while we continue  to advance the science and the standards.   00:16:49.680 --> 00:16:54.880 You know, we are going to need to do both at once.  And so we need a flexible framework that enables   00:16:54.880 --> 00:17:00.080 that kind of agility, and advancing the science  and the standards, that is going to be something   00:17:00.080 --> 00:17:07.520 that really demands that kind of cross-discipline  or cross kind of expertise group coming together   00:17:07.520 --> 00:17:12.800 to work on that—researchers, academics, civil  society, governments and, of course, industry. 00:17:12.800 --> 00:17:17.440 And so I think that is, actually, the second  problem is, like, how do we actually build   00:17:19.120 --> 00:17:24.960 the kind of forums and ways of working  together, the public-private partnership   00:17:24.960 --> 00:17:32.400 kind of efforts that allow all of that expertise  to come together and fit together over time,   00:17:32.400 --> 00:17:36.560 right. Because when these are really big,  broad challenges, you kind of have to break   00:17:36.560 --> 00:17:40.720 them down incrementally, make progress on  them, and then bring them back together. 00:17:40.720 --> 00:17:46.960 And so I think about, like, one example that I,  you know, really have been reflecting on lately   00:17:46.960 --> 00:17:51.360 is, you know, in the context of building  standards, like, how do you do that,   00:17:51.360 --> 00:17:58.080 right? Again, standards are going to benefit  from that whole community of expertise. And,   00:17:58.080 --> 00:18:02.480 you know, there are lots of different kinds of  quote-unquote standards, though, right. You kind   00:18:02.480 --> 00:18:08.640 of have the “small s” industry standards. You  have the kind of “big S” international standards,   00:18:08.640 --> 00:18:14.160 for example. And how do you, kind of, leverage  one to accelerate the other, I think, is part of,   00:18:14.160 --> 00:18:20.480 like, how we need to work together within this  ecosystem. And, like, I think what we and others   00:18:20.480 --> 00:18:23.840 have done in an organization like C2PA [Coalition  for Content Provenance and Authenticity], for   00:18:23.840 --> 00:18:30.480 example, where we've really built an industry  specification but then built on that towards an   00:18:30.480 --> 00:18:36.480 international standard effort is one example  that is interesting, right, to point to. 00:18:36.480 --> 00:18:39.520 And then, you know, I actually think that  bridges to the third thing that we need to   00:18:39.520 --> 00:18:47.920 do together within this whole community, which is,  you know, really think again about how we manage   00:18:47.920 --> 00:18:54.720 the breadth of this challenge and opportunity  of AI by thinking about this horizontal-vertical   00:18:54.720 --> 00:19:00.880 problem. And, you know, I think that's where  it's not just the sort of tech industry,   00:19:00.880 --> 00:19:05.120 for example. It's broader industry that's going to  be really applying this technology that needs to   00:19:05.120 --> 00:19:10.000 get involved in the conversation about not just,  sort of, testing AI models, for example, but also   00:19:10.000 --> 00:19:17.662 testing how AI systems or applications are working  in context. And so, yes, so much fun opportunity! 00:19:17.662 --> 00:19:17.670 [MUSIC] 00:19:17.670 --> 00:19:20.880 SULLIVAN: Amanda, this was just  fantastic. You've really set the   00:19:20.880 --> 00:19:26.865 stage for this podcast. And thank you so much  for sharing your time and wisdom with us. 00:19:26.865 --> 00:19:27.960 CRAIG DECKARD: Thank you. 00:19:27.960 --> 00:19:32.000 SULLIVAN: And to our listeners, we're so  glad you joined us for this conversation.   00:19:32.000 --> 00:19:37.600 An exciting lineup of episodes are on the way, and  we can't wait to have you back for the next one. 00:19:50.440 --> 00:19:51.440 [MUSIC   00:19:51.440 --> 00:20:03.600 FADES]