1 00:00:01,967 --> 00:00:03,257 Elham Tabasi, welcome. 2 00:00:03,257 --> 00:00:05,045 Glad to have you on the podcast. 3 00:00:05,262 --> 00:00:06,362 Thanks for having me. 4 00:00:06,362 --> 00:00:07,686 Glad to be here. 5 00:00:08,751 --> 00:00:15,551 But first, why don't you tell us what NIST is, because it's an important agency, but 6 00:00:15,551 --> 00:00:20,771 one that maybe people are not necessarily familiar with, even if they know something 7 00:00:20,771 --> 00:00:23,839 about the US government and AI policy. 8 00:00:24,782 --> 00:00:25,792 Oh, yeah, happy to. 9 00:00:25,792 --> 00:00:28,002 And that's a great starting point. 10 00:00:28,162 --> 00:00:33,142 So NIST, or National Institute of Standards and Technology, is a non 11 00:00:33,142 --> 00:00:34,402 -regulatory agency. 12 00:00:34,402 --> 00:00:36,962 We are under the Department of Commerce. 13 00:00:37,002 --> 00:00:43,122 And NIST was established back in 1902 with a unique mission to advance US innovation 14 00:00:43,122 --> 00:00:45,802 and industrial competitiveness. 15 00:00:46,482 --> 00:00:50,382 Here at NIST, we have a very broad portfolio of research. 16 00:00:50,382 --> 00:00:55,102 from building the most accurate atomic clock to cybersecurity to artificial 17 00:00:55,102 --> 00:01:00,002 intelligence, advanced communications, but most importantly, a long tradition of 18 00:01:00,002 --> 00:01:02,382 cultivating trust in technology. 19 00:01:02,862 --> 00:01:06,621 And we do that by advancing measurement science and standards, measurement science 20 00:01:06,621 --> 00:01:12,682 and standards that makes technology more reliable, secure, private, fair, or in 21 00:01:12,682 --> 00:01:14,132 other words, more trustworthy. 22 00:01:14,132 --> 00:01:17,310 And that's exactly what we have been doing in the space of AI. 23 00:01:18,273 --> 00:01:19,073 Absolutely. 24 00:01:19,073 --> 00:01:23,273 And so you were one of the architects of the NIST AI Risk Management framework, 25 00:01:23,273 --> 00:01:26,973 which came out in January of 2023. 26 00:01:27,593 --> 00:01:31,193 Other than the general goal of trust and technology that you just described, where 27 00:01:31,193 --> 00:01:32,603 did that project come from? 28 00:01:32,603 --> 00:01:34,055 How did that come about? 29 00:01:34,894 --> 00:01:35,934 Yeah. 30 00:01:36,254 --> 00:01:45,534 So AIRMF that was released a little over a year ago, it was directed by congressional 31 00:01:45,534 --> 00:01:46,674 mandate. 32 00:01:46,674 --> 00:01:52,974 It is a voluntary framework for managing the risk of AI in a flexible, structured, 33 00:01:52,974 --> 00:01:55,554 and measurable way. 34 00:01:56,774 --> 00:01:58,990 The whole point is that 35 00:01:58,990 --> 00:02:04,310 We harness the power of AI for good and AI can serve all people in a beneficial, fair 36 00:02:04,310 --> 00:02:06,730 and equitable manner. 37 00:02:07,750 --> 00:02:11,030 These three attributes that I talked about, flexible, structured and 38 00:02:11,030 --> 00:02:18,370 measurable, we wanted to be flexible to be able to keep up the pace with this rapid 39 00:02:18,370 --> 00:02:23,662 change of the technology, but also so that the different organizations... 40 00:02:23,662 --> 00:02:27,262 with different resources be able to implement that. 41 00:02:27,262 --> 00:02:33,162 We didn't want to be a heavy thing that only the big organizations that have AI 42 00:02:33,162 --> 00:02:38,002 governance or a lot of people doing that be able to operationalize it. 43 00:02:38,362 --> 00:02:44,202 Structured in the terms that when we start working on AI RMF, the concept of the 44 00:02:44,202 --> 00:02:49,082 trustworthiness, the question of what constitutes trust were not quite answered 45 00:02:49,082 --> 00:02:52,558 when there has been a lot of really high value 46 00:02:52,558 --> 00:02:59,188 high -level value -based documents that talk about these things. 47 00:02:59,188 --> 00:03:06,638 None of them try to get those principles into sort of the building blocks that the 48 00:03:06,638 --> 00:03:09,398 developers and the employer of the technology can use that. 49 00:03:09,398 --> 00:03:13,818 So working with the community, we try to come up with unpacking the concept of the 50 00:03:13,818 --> 00:03:18,378 trustworthiness, and come up with the components of the AI system that's 51 00:03:18,378 --> 00:03:22,350 trustworthy, which those characteristics are. 52 00:03:22,350 --> 00:03:28,570 valid and reliable, safe, secure and resilient, privacy -enhanced, explainable 53 00:03:28,570 --> 00:03:34,790 and interpretable, fair with harmful bias, managed and transparent and accountable. 54 00:03:34,870 --> 00:03:40,470 Bring the community on a shared understanding around each of them to get 55 00:03:40,470 --> 00:03:46,450 to the next attribute of the measurable to have a good understanding of what we want 56 00:03:46,450 --> 00:03:48,710 to measure to be able to measure them. 57 00:03:48,710 --> 00:03:52,014 So the importance of being able to... 58 00:03:52,014 --> 00:03:58,554 define trustworthiness into technical and socio -technical terms that developers of 59 00:03:58,554 --> 00:04:02,174 the technology, deployers of the technologies, users of the technology can 60 00:04:02,174 --> 00:04:06,134 understand, and then build a community together to come up with the scientific 61 00:04:06,134 --> 00:04:11,094 underpinning for how to measure those because if you're really serious in the 62 00:04:11,094 --> 00:04:14,534 business of improving things, first we have to be able to measure them. 63 00:04:14,534 --> 00:04:17,492 We can't improve things that we cannot measure. 64 00:04:18,543 --> 00:04:21,203 So I want to get back to that question about measurement because it is really 65 00:04:21,203 --> 00:04:21,723 critical. 66 00:04:21,723 --> 00:04:25,703 But first, let me just ask you, that's a big challenge talking about 67 00:04:25,703 --> 00:04:28,083 trustworthiness around AI and all those principles you described. 68 00:04:28,083 --> 00:04:29,363 They make sense. 69 00:04:29,363 --> 00:04:32,763 But how did you and the team get your arms around that? 70 00:04:32,763 --> 00:04:37,103 How to decide what kinds of aspects to focus on and how to operationalize these 71 00:04:37,103 --> 00:04:38,287 broad principles? 72 00:04:39,086 --> 00:04:40,326 Great question. 73 00:04:40,326 --> 00:04:46,526 So how we do our work in this case and also in many other cases is to work 74 00:04:46,526 --> 00:04:49,586 collaboratively with public and private stakeholders. 75 00:04:49,586 --> 00:04:54,666 So basically everything that we do, we start with a formal process of request for 76 00:04:54,666 --> 00:04:55,245 information. 77 00:04:55,245 --> 00:04:58,486 That's where the government put an announcement into a federal registered 78 00:04:58,486 --> 00:05:03,326 notice, announce to the whole public that we're doing this and ask for comments. 79 00:05:03,326 --> 00:05:08,846 Then we work with the community to do rounds of workshops, listening sessions. 80 00:05:08,846 --> 00:05:13,346 to get answers to all of these questions. 81 00:05:13,906 --> 00:05:22,166 And in doing so, we very early on realized that we should definitely listen to a 82 00:05:22,166 --> 00:05:25,586 community with the expertise that are building the technology or designing the 83 00:05:25,586 --> 00:05:29,986 technology, those that have the expertise in computer science, mathematics, 84 00:05:29,986 --> 00:05:34,416 statistician, but also the community that study the impact of the technology, 85 00:05:34,416 --> 00:05:38,286 psychologists, sociologists, philosophers. 86 00:05:38,286 --> 00:05:43,826 but also the community that are going to be in charge of either top down or bottom 87 00:05:43,826 --> 00:05:48,326 up, try to operationalize and implement it and figure out how it basically fits into 88 00:05:48,326 --> 00:05:54,006 the bigger risk management of the organization, so business groups. 89 00:05:54,265 --> 00:06:00,066 And we also did listening sessions with the AI users, one of our best listening 90 00:06:00,066 --> 00:06:05,346 sessions when we went to community colleges, when we went to minority serving 91 00:06:05,346 --> 00:06:07,502 institutions and talked with the students. 92 00:06:07,502 --> 00:06:14,642 So by getting a lot of input from the community, input from a very diverse sets 93 00:06:14,642 --> 00:06:23,561 of backgrounds and expertise, we came to where we are in terms of the 94 00:06:23,561 --> 00:06:26,528 trustworthiness characteristics, but also the whole document. 95 00:06:27,471 --> 00:06:31,511 Doesn't that broad amount of input though make it a challenge to come up with a 96 00:06:31,511 --> 00:06:34,823 consensus about what direction to go, what the principle should be? 97 00:06:35,470 --> 00:06:36,830 It does. 98 00:06:36,850 --> 00:06:40,890 And I think that's part of the reasons that we want to do that, because it is 99 00:06:40,890 --> 00:06:43,830 really important to keep all of the voices included. 100 00:06:43,830 --> 00:06:48,990 And when you do that, and if you really go for the diversity, of course we're going 101 00:06:48,990 --> 00:06:56,000 to hear inputs both sides of the spectrum and trying to figure out the right space. 102 00:06:56,000 --> 00:06:58,970 And I don't want to say middle ground, because there's a lot of the things to 103 00:06:58,970 --> 00:06:59,750 consider. 104 00:06:59,750 --> 00:07:04,590 You want to be scientific, high scientific integrity and scientific validity, right? 105 00:07:04,590 --> 00:07:08,630 So we want to make sure that whatever we put there is technically robust and 106 00:07:08,630 --> 00:07:09,950 empirically validated. 107 00:07:09,950 --> 00:07:16,010 It is operationalizable and implementable today, but also it doesn't get expired in 108 00:07:16,010 --> 00:07:19,990 a very short amount of time. 109 00:07:20,030 --> 00:07:27,110 So we heard when we start this process from everything and everywhere, from the 110 00:07:27,110 --> 00:07:31,350 civil society groups that wanted everything to be absolutely 111 00:07:31,662 --> 00:07:38,842 public and open and audited and a lot of vigorous tests and transparency 112 00:07:38,842 --> 00:07:41,402 requirements, a lot of really good reasons for that. 113 00:07:41,402 --> 00:07:45,122 And we also hear from the technology developers, importance of the IP and 114 00:07:45,122 --> 00:07:48,142 innovations and how those things have to be protected. 115 00:07:49,222 --> 00:07:55,442 But that's the beauty of this, when you give everybody the voice and we are 116 00:07:55,442 --> 00:08:00,622 objective and want to convene. 117 00:08:00,622 --> 00:08:08,382 try to make value for every participant in the discussions to see how those consensus 118 00:08:08,382 --> 00:08:09,882 are being built. 119 00:08:09,882 --> 00:08:15,102 And the other thing is that when we get input from a very vast group of the 120 00:08:15,102 --> 00:08:18,462 community, NIST has the pen in writing this. 121 00:08:18,462 --> 00:08:21,102 So we would kind of make a decision where we have to go. 122 00:08:21,102 --> 00:08:27,142 Of course, we go out and explain and make sure that to the extent possible, we have 123 00:08:27,142 --> 00:08:27,722 the buy -in. 124 00:08:27,722 --> 00:08:29,646 Those buy -in are important because that's 125 00:08:29,646 --> 00:08:34,746 That's where the adoptions happens, but that is part of the process, building the 126 00:08:34,746 --> 00:08:35,724 consensus. 127 00:08:36,879 --> 00:08:43,015 How can we measure the trustworthiness or responsibility of an AI system? 128 00:08:44,334 --> 00:08:47,464 That's a very difficult, you know, that's a technical challenge. 129 00:08:47,464 --> 00:08:50,374 That's a question that the community is looking at this. 130 00:08:50,374 --> 00:08:54,014 It's difficult at many different dimensions. 131 00:08:54,534 --> 00:08:58,914 So our institutions, our work, the group that I have been a part of this, and by 132 00:08:58,914 --> 00:09:02,154 the way, I joined this in June 7, 1999. 133 00:09:02,154 --> 00:09:06,854 So in a few months, it's going to be my 25 years and all my career here, I've been 134 00:09:06,854 --> 00:09:13,102 working on computer vision, machine learning type of efforts. 135 00:09:13,102 --> 00:09:16,062 and a lot of tests and evaluations. 136 00:09:16,062 --> 00:09:20,882 So if some of your listeners are familiar with the face recognition tests or 137 00:09:20,882 --> 00:09:25,022 fingerprint work that we have done, I've been part of that team doing that type of 138 00:09:25,022 --> 00:09:26,022 evaluation. 139 00:09:26,282 --> 00:09:32,342 So with a rich history and experience and understanding of how to do evaluations, 140 00:09:33,222 --> 00:09:38,362 evaluations of the AI system poses a new set of questions and challenges. 141 00:09:38,562 --> 00:09:41,646 One is, of course, what we mean by 142 00:09:41,646 --> 00:09:44,606 safe AI, what we mean by secure AI. 143 00:09:44,686 --> 00:09:52,486 But even if we can come up with some understanding of those, AI systems are not 144 00:09:52,486 --> 00:09:57,226 just about data computing algorithm. 145 00:09:57,226 --> 00:10:01,186 There are complex interactions of the data computing algorithm with the environment 146 00:10:01,186 --> 00:10:04,006 they're operating in, the human that operates that. 147 00:10:04,006 --> 00:10:08,334 So the solutions and the answers are not only technical and technology. 148 00:10:08,334 --> 00:10:12,054 but they all had to be answered through the socio -technical lens. 149 00:10:12,054 --> 00:10:17,534 What that means is that the evaluations that we have done before to get systems 150 00:10:17,534 --> 00:10:22,294 out of their context of use, bring them to the laboratory, run the test data through 151 00:10:22,294 --> 00:10:24,644 them and test them, and those are important. 152 00:10:24,644 --> 00:10:25,464 We have to do them. 153 00:10:25,464 --> 00:10:28,874 They give us some baseline of the capabilities and limitations and failure 154 00:10:28,874 --> 00:10:34,134 modes, but they are not going to be sufficient to surface all the sorts of the 155 00:10:34,134 --> 00:10:36,750 problems, negative consequences that can come. 156 00:10:36,750 --> 00:10:38,410 out of the AI systems. 157 00:10:38,410 --> 00:10:43,370 We are hearing a lot about the red teaming of AI systems with spot checking and 158 00:10:43,370 --> 00:10:50,510 basically trying to see how AI systems can get into undesirable behavior. 159 00:10:51,330 --> 00:10:57,550 But beyond those things, we believe that another set of testing where AI systems 160 00:10:57,550 --> 00:11:04,370 are being tested in their native context of use with the humans, individuals that 161 00:11:04,370 --> 00:11:04,878 are 162 00:11:04,878 --> 00:11:06,518 going to use it, it's really important. 163 00:11:06,518 --> 00:11:07,598 We call them field testing. 164 00:11:07,598 --> 00:11:11,058 We have seen examples of that in sort of, you know, pharmaceutical. 165 00:11:11,058 --> 00:11:12,838 FDA is doing this. 166 00:11:12,838 --> 00:11:20,318 So it's, there are all these difficulties of exactly what we are testing for. 167 00:11:20,318 --> 00:11:24,738 And then the spectrum of the test that we are doing is not going to be enough. 168 00:11:25,138 --> 00:11:27,088 The laboratory testing is not going to be enough. 169 00:11:27,088 --> 00:11:29,598 We have to augment it with other testing. 170 00:11:29,598 --> 00:11:32,454 And then the... 171 00:11:32,622 --> 00:11:36,282 The other thing is that we are not going to be able to test multi -purpose AI 172 00:11:36,282 --> 00:11:40,801 systems for all of the different use cases that are being used. 173 00:11:40,801 --> 00:11:42,602 So how to scale all of these things? 174 00:11:42,602 --> 00:11:44,122 So a difficult problem. 175 00:11:44,122 --> 00:11:48,392 It is something that we have been working on since the release of the AI RMF. 176 00:11:48,392 --> 00:11:52,362 One of the four functions on AI RMF is the measure function. 177 00:11:52,362 --> 00:11:58,122 AI RMF provides, in addition to giving the structured approach to what we mean by 178 00:11:58,122 --> 00:12:00,174 trustworthiness, the meat of that is... 179 00:12:00,174 --> 00:12:06,514 72 recommendations categorized in four functions of map, measure, manage, and 180 00:12:06,514 --> 00:12:12,534 govern to how understand and identify AI risks, how to measure them based on those 181 00:12:12,534 --> 00:12:17,494 measurement results, what type of risk responses and mitigations are to be 182 00:12:17,494 --> 00:12:23,894 implemented, and the govern functions give recommendations basically for the 183 00:12:23,894 --> 00:12:28,110 processes and procedures, all of the embeddings that's needed for the right. 184 00:12:28,110 --> 00:12:30,930 risk management processes. 185 00:12:32,069 --> 00:12:36,970 So after the release of the AIRM, we put a lot of our effort into the measure 186 00:12:36,970 --> 00:12:44,250 functions, which quite aligns and become the foundation work that our latest 187 00:12:44,250 --> 00:12:47,770 assignment as part of the EO is built on top of that. 188 00:12:47,770 --> 00:12:52,750 So a lot of work is happening as part of the task that we have with the EO as part 189 00:12:52,750 --> 00:12:55,690 of the AI Safety Institute. 190 00:12:55,690 --> 00:12:57,006 We are standing up. 191 00:12:57,006 --> 00:13:01,366 and its consortium to work with the community on figuring out the evaluations. 192 00:13:01,366 --> 00:13:06,186 The last thing I will say in this space is that it's a tough problem and we are not 193 00:13:06,186 --> 00:13:08,426 going to get the answers in the first go. 194 00:13:08,426 --> 00:13:13,746 So we have to be on the understand that it's going to be iterative and as we do 195 00:13:13,746 --> 00:13:17,986 more of this research, run evaluations, you know, write guidelines based on those 196 00:13:17,986 --> 00:13:19,866 guidelines, conduct evaluations. 197 00:13:19,866 --> 00:13:24,266 The evaluations will tell us that what was the gaps in the guidelines, the evaluation 198 00:13:24,266 --> 00:13:26,030 can also say the gaps in the 199 00:13:26,030 --> 00:13:31,410 So hopefully we are going to generate a positive feedback loop that the 200 00:13:31,410 --> 00:13:38,330 measurement science is getting strengthened by empirical evidence and 201 00:13:38,330 --> 00:13:40,450 science is going to be built around that. 202 00:13:40,450 --> 00:13:45,990 But also the practice and adoption of the AIs are going to get stronger because of 203 00:13:45,990 --> 00:13:47,308 the test results. 204 00:13:47,824 --> 00:13:51,444 Can you talk about some of the directions that you and the community are going to 205 00:13:51,444 --> 00:13:55,399 try to come up with those benchmarks for some of those issues? 206 00:13:56,142 --> 00:14:03,542 All right, so October 30th, Executive Order 14110, Safe, Secure, and Trustworthy 207 00:14:03,542 --> 00:14:06,022 AI was released. 208 00:14:06,022 --> 00:14:11,102 It's a very comprehensive set of instructions to many of the federal 209 00:14:11,102 --> 00:14:16,262 government agencies on how to promote safe, secure, trustworthy AI. 210 00:14:16,282 --> 00:14:23,790 Section 4 on safety and security tasks several agencies, but predominantly NIST. 211 00:14:23,790 --> 00:14:30,170 on advancing the measurement for and evaluations for safe, secure, trustworthy 212 00:14:30,170 --> 00:14:31,490 AI systems. 213 00:14:31,490 --> 00:14:42,260 In particular, it directs NIS to develop guidelines for evaluations, reteaming, 214 00:14:42,260 --> 00:14:44,350 safety, cybersecurity. 215 00:14:44,350 --> 00:14:49,130 It asks NIS to facilitate development of consensus -based standards. 216 00:14:49,130 --> 00:14:52,654 It also asks NIS to provide testing environment for 217 00:14:52,654 --> 00:14:54,874 evaluations of the AI systems. 218 00:14:54,874 --> 00:14:58,814 These guidelines and infrastructures will be voluntary resources for the community 219 00:14:58,814 --> 00:14:59,754 to use. 220 00:14:59,754 --> 00:15:04,534 So right now we had our heads down drafting the documents, which we are 221 00:15:04,534 --> 00:15:07,134 hoping to go out for public comment. 222 00:15:07,134 --> 00:15:10,874 Based on the public comment, we're hoping to improve it to meet the deadlines of the 223 00:15:10,874 --> 00:15:11,954 EO. 224 00:15:12,394 --> 00:15:19,214 And the EO gives us a 270 days deadline that gets us to the end of the July. 225 00:15:19,214 --> 00:15:20,878 But all of these could work. 226 00:15:20,878 --> 00:15:25,358 All of these actions that the EEO sets in the motion, we think that should not and 227 00:15:25,358 --> 00:15:27,798 cannot stop in end of July. 228 00:15:27,798 --> 00:15:34,978 So we're looking into the AI Safety Institute and its consortium to give a 229 00:15:34,978 --> 00:15:38,898 long lasting approach to all of these efforts in the research and evaluation 230 00:15:38,898 --> 00:15:39,878 builder. 231 00:15:40,118 --> 00:15:44,898 So what we are working with the community is first develop the guidelines that we 232 00:15:44,898 --> 00:15:48,618 have asked through the EEO, working through the consortium with the consortium 233 00:15:48,618 --> 00:15:49,390 members. 234 00:15:49,390 --> 00:15:56,650 when it's out to refine them, improve them, conduct evaluations, and based on 235 00:15:56,650 --> 00:16:00,330 those kind of lived experiences, improve the guidelines. 236 00:16:00,610 --> 00:16:06,070 Eventually, we definitely look like to have sort of standardized test 237 00:16:06,070 --> 00:16:11,670 environments, standardized methods and metrics and methodology for testing so 238 00:16:11,670 --> 00:16:16,690 that the community can use them and help with the scaling of the evaluations that 239 00:16:16,690 --> 00:16:18,580 needed to be done in this space. 240 00:16:19,695 --> 00:16:25,055 Have you seen good examples of either private sector companies or public sector 241 00:16:25,055 --> 00:16:28,743 organizations taking the RMF and operationalizing it? 242 00:16:29,806 --> 00:16:33,606 We have heard a lot and we are really grateful for that. 243 00:16:33,606 --> 00:16:40,726 So when the AIRMF was released January 26, 2023, a lot of different companies 244 00:16:40,726 --> 00:16:45,406 voluntarily and entities, I should say, gave us some sort of testimonies on how 245 00:16:45,406 --> 00:16:48,456 they are planning to use AIRMF. 246 00:16:48,456 --> 00:16:52,146 And that again goes back to building that engagement because when it comes out, 247 00:16:52,146 --> 00:16:54,446 everybody sees themselves in it. 248 00:16:54,806 --> 00:16:57,742 And then since then, we have heard from... 249 00:16:57,742 --> 00:17:02,162 large medium, small size businesses that are using it. 250 00:17:02,682 --> 00:17:07,322 Some of those have given us, we call them use cases, how they're using it 251 00:17:07,322 --> 00:17:07,752 internally. 252 00:17:07,752 --> 00:17:10,222 Some of those are posted on our website. 253 00:17:10,462 --> 00:17:14,942 Some of the other one has been talking to us and letting us know that, particularly 254 00:17:14,942 --> 00:17:19,102 for the larger organizations, how they are aligning their internal processes with AI 255 00:17:19,102 --> 00:17:19,862 or MF. 256 00:17:19,862 --> 00:17:24,270 And we're always happy to work with the smaller size businesses to... 257 00:17:24,270 --> 00:17:26,650 help them with the operationalization. 258 00:17:26,650 --> 00:17:32,690 So one of the things, for example, we did in March of 2023, we put out the 259 00:17:32,690 --> 00:17:34,610 trustworthy AI resource center. 260 00:17:34,610 --> 00:17:40,470 So in addition to the release of the ARRMF, with the goal of keeping ARRMF, we 261 00:17:40,470 --> 00:17:46,450 call it as evergreen as possible, with a revision cycle of three to five years, we 262 00:17:46,450 --> 00:17:51,110 put out a playbook out that basically gives a lot of... 263 00:17:51,726 --> 00:17:56,866 informative references and information and descriptions on exactly what to do for 264 00:17:56,866 --> 00:18:02,726 each of those 72 recommendations, what they can do today, what are the tools and 265 00:18:02,726 --> 00:18:06,946 standards that can be used today for implementations. 266 00:18:07,426 --> 00:18:12,206 So the playbook, we put a revision cycle of every six months so we can keep it up 267 00:18:12,206 --> 00:18:19,086 with the latest changes in the technology as tools and information becomes 268 00:18:19,086 --> 00:18:20,286 available. 269 00:18:20,966 --> 00:18:21,774 We... 270 00:18:21,774 --> 00:18:26,674 We made sure that this is sort of interactive, filterable based on the role 271 00:18:26,674 --> 00:18:31,514 of the AI actor, designer, developer, deployer, the area that they want to get 272 00:18:31,514 --> 00:18:35,714 information out from bias or privacy. 273 00:18:35,714 --> 00:18:44,114 And we're also working with the community to add more tools and support, supportive 274 00:18:44,114 --> 00:18:48,814 documents, informative references, implementations that can help with the 275 00:18:48,814 --> 00:18:49,694 implementation of that. 276 00:18:49,694 --> 00:18:51,470 But across the board, we have seen... 277 00:18:51,470 --> 00:18:52,750 It's being used. 278 00:18:52,750 --> 00:19:00,970 Another interesting or welcoming happening is that at the time of the release of the 279 00:19:00,970 --> 00:19:04,050 AI RMF, we also released Crosswalks. 280 00:19:04,050 --> 00:19:08,690 So AI RMF was not the first document being out, and we want to definitely make sure 281 00:19:08,690 --> 00:19:13,330 that we build based on the knowledge of the community and leverage. 282 00:19:13,330 --> 00:19:18,610 For example, the definition of AI system in the AI RMF is based on the OECD 283 00:19:18,610 --> 00:19:19,770 definition. 284 00:19:20,230 --> 00:19:21,038 So if we... 285 00:19:21,038 --> 00:19:22,558 based on the input of the community. 286 00:19:22,558 --> 00:19:27,338 We improved OECD definition and then OECD took some of our improvement, put it back 287 00:19:27,338 --> 00:19:30,618 into their definitions, definition of AI systems. 288 00:19:32,118 --> 00:19:38,578 So we want to make sure that we are not adding confusion to the market by yet put 289 00:19:38,578 --> 00:19:39,628 another document out. 290 00:19:39,628 --> 00:19:45,798 So we put out crosswalks that says that, for example, how AI RMF relates to OECD AI 291 00:19:45,798 --> 00:19:49,742 recommendations, the draft of the EU AI Act. 292 00:19:49,742 --> 00:19:54,502 few weeks ago, but at that time it was a draft with some other documents that's 293 00:19:54,502 --> 00:20:00,442 coming out of the White House, but also with ISO, the standards that were coming 294 00:20:00,442 --> 00:20:04,182 out of the ISO international standard organizations. 295 00:20:04,302 --> 00:20:10,002 ISO Insights is the technical tag to ISO, US technical tag to ISO. 296 00:20:10,002 --> 00:20:14,042 Basically pick that up and as ISO is coming up with other standards, they 297 00:20:14,042 --> 00:20:17,042 develop crosswalks with AI RMF. 298 00:20:17,142 --> 00:20:19,182 And hopefully all of these things that... 299 00:20:19,182 --> 00:20:25,142 I think is an indication that AI -RMF are being used, so that's why they need those 300 00:20:25,142 --> 00:20:31,862 crosswalks, but also hopefully bringing more semantic interoperability in the 301 00:20:31,862 --> 00:20:32,742 short run. 302 00:20:34,479 --> 00:20:39,619 You mentioned the executive order and the AI Safety Institute that was one of the 303 00:20:39,619 --> 00:20:43,479 things that was created out of that process and you're now taking over as CTO 304 00:20:43,479 --> 00:20:45,839 of that USAI Safety Institute. 305 00:20:45,919 --> 00:20:49,959 How is that different from the work that NIST has done before and what's on the 306 00:20:49,959 --> 00:20:51,847 agenda for the Institute? 307 00:20:52,846 --> 00:21:04,506 Yeah, we look at both the EO and the Safety Institute as basically, that's the 308 00:21:04,506 --> 00:21:11,586 phrase I use, supercharging our efforts in advancing trustworthy responsible AI. 309 00:21:11,726 --> 00:21:19,246 So I think in terms of the goal and vision to advance the science, practice, and 310 00:21:19,246 --> 00:21:20,910 policy of AI systems, 311 00:21:20,910 --> 00:21:25,770 increasing adoptions of the AI is the work that we are doing. 312 00:21:25,770 --> 00:21:34,010 And all of these things are giving us more capacity and strengthens our engagements 313 00:21:34,010 --> 00:21:36,570 to do better and more. 314 00:21:36,590 --> 00:21:41,550 Eventually, we also want to keep an eye on the, or help with the adoption. 315 00:21:41,550 --> 00:21:45,690 So as a measurement science agency, we want to make sure that the right 316 00:21:45,690 --> 00:21:49,934 measurement science tools, standards, guidelines are there. 317 00:21:49,934 --> 00:21:53,094 for anybody that wants to design, develop, or use AI systems. 318 00:21:53,094 --> 00:21:59,534 And hopefully, by doing this, we actually help the adoption rates of AI in science, 319 00:21:59,554 --> 00:22:06,774 in health, climate, agriculture, all of those areas that we think that that's 320 00:22:06,774 --> 00:22:12,756 where AI technology can bring many, many benefits. 321 00:22:13,967 --> 00:22:17,447 What do you think the biggest gap is right now or the biggest area where we haven't 322 00:22:17,447 --> 00:22:21,647 seen enough progress on some of the elements that are going to be necessary to 323 00:22:21,647 --> 00:22:22,567 promote that? 324 00:22:23,886 --> 00:22:28,866 I might be biased coming from a measurement science agency, but I would 325 00:22:28,866 --> 00:22:30,946 say measurement and evaluations. 326 00:22:30,946 --> 00:22:34,606 I think everybody wants to save, secure, trustworthy AI, but we don't know how to 327 00:22:34,606 --> 00:22:34,996 measure them. 328 00:22:34,996 --> 00:22:36,286 We don't have the metrics there. 329 00:22:36,286 --> 00:22:38,886 We don't have the methodologies there. 330 00:22:39,266 --> 00:22:48,706 There are documents, policies they like to talk about, assuring that AI systems are 331 00:22:48,706 --> 00:22:48,946 safe. 332 00:22:48,946 --> 00:22:50,646 You know, we saw it in the EU AI Act. 333 00:22:50,646 --> 00:22:51,950 They want to see... 334 00:22:51,950 --> 00:22:56,570 some sort of assurance or conformity assessment, but those specifications 335 00:22:56,570 --> 00:22:58,790 hasn't been written yet. 336 00:22:58,790 --> 00:23:02,090 So what do we mean by safety? 337 00:23:03,210 --> 00:23:09,110 And policies usually stay at the abstract high level that AI should be safe or non 338 00:23:09,110 --> 00:23:14,270 -discriminatory, but bringing it down, getting everybody on the same page, that's 339 00:23:14,270 --> 00:23:17,050 what it means, and how to assure for them. 340 00:23:17,050 --> 00:23:21,358 I think that's the biggest, that's one of the biggest gaps. 341 00:23:22,670 --> 00:23:30,450 In general, in this AI space, I will say that we know a lot less than what we 342 00:23:30,450 --> 00:23:31,290 should. 343 00:23:31,290 --> 00:23:39,190 So anything that we can do to be able to know more what to expect, how to 344 00:23:39,190 --> 00:23:45,450 characterize the behavior functionality of the systems, these are the good partners. 345 00:23:47,375 --> 00:23:51,375 How much are those really scientific questions though? 346 00:23:51,375 --> 00:23:55,494 And part of what I'm thinking about is you get to some issues like bias, where 347 00:23:55,494 --> 00:23:59,555 there's various different tests about algorithmic fairness, but some of them are 348 00:23:59,555 --> 00:24:00,915 fundamentally incompatible. 349 00:24:00,915 --> 00:24:04,675 And you mentioned the socio -technical aspects of the systems where there may be 350 00:24:04,675 --> 00:24:07,765 things that are not just about yes or no. 351 00:24:07,765 --> 00:24:14,335 So how do you and others in the community work to identify what it is where we can 352 00:24:14,335 --> 00:24:17,575 actually get an answer through the kind of science you're talking about? 353 00:24:17,710 --> 00:24:18,650 Yeah. 354 00:24:18,650 --> 00:24:20,350 Thank you for that question. 355 00:24:20,350 --> 00:24:29,230 So back in March 2022, we put out a document out, special publication 1270. 356 00:24:29,230 --> 00:24:36,770 It has a name towards standards for understanding harmful bias in AI systems. 357 00:24:36,770 --> 00:24:40,550 But these are some of the foundational work that we're doing with the community 358 00:24:40,550 --> 00:24:41,382 that. 359 00:24:41,486 --> 00:24:45,706 Again, as you said, there are tools for bias, but what are really measuring? 360 00:24:45,706 --> 00:24:46,856 Are they measuring the right thing? 361 00:24:46,856 --> 00:24:48,286 What should we really measure? 362 00:24:48,286 --> 00:24:51,846 What is bias across the life cycle of AI anyway? 363 00:24:51,846 --> 00:24:54,496 So that document talks about a lot of these things. 364 00:24:54,496 --> 00:24:58,546 And if I want to summarize, the most important to me, the most important point 365 00:24:58,546 --> 00:25:07,806 of that document is that bias is not just about disparity and demographic disparity. 366 00:25:07,806 --> 00:25:09,582 A lot of us, when you're talking about 367 00:25:09,582 --> 00:25:12,962 bias, the first thing that comes to our mind is demographic disparity. 368 00:25:12,962 --> 00:25:19,422 The document talks about computational and statistical bias as one type of bias, but 369 00:25:19,422 --> 00:25:22,802 there are also other types of bias, systemic biases. 370 00:25:22,802 --> 00:25:28,982 If you go and do a statistically valid data collection of the community and 371 00:25:28,982 --> 00:25:34,622 society today, that's going to be biased and the biases baked into the society will 372 00:25:34,622 --> 00:25:39,214 get into the systems because our training data will carry those biases. 373 00:25:39,214 --> 00:25:43,934 It also talks about another type of bias, which is cognitive biases. 374 00:25:43,934 --> 00:25:49,534 How the different individuals process information given by AI systems or 375 00:25:49,534 --> 00:25:51,374 interact with AI systems. 376 00:25:51,374 --> 00:25:54,394 And based on that, how we react to those is different. 377 00:25:54,394 --> 00:25:56,234 So there's biases there too. 378 00:25:56,234 --> 00:26:02,254 Understanding all of these types of biases, having a wholesome understanding 379 00:26:02,254 --> 00:26:05,678 of now in this topic what the bias is and... 380 00:26:05,678 --> 00:26:11,738 how biases showed itself at the different stages of the AI lifecycle become really 381 00:26:11,738 --> 00:26:16,158 important on other steps of the evaluation. 382 00:26:16,158 --> 00:26:21,378 So the science part of that, again, becomes to having a sort of a wholesome 383 00:26:21,378 --> 00:26:27,578 approach to what could be the sources of the biases, where they're coming from, and 384 00:26:27,578 --> 00:26:32,174 that information can help us to do the measurement. 385 00:26:32,174 --> 00:26:35,774 You asked the question about the socio -technical. 386 00:26:35,774 --> 00:26:38,644 So all of these problems are socio -technical. 387 00:26:38,644 --> 00:26:41,934 So there is not just a technical or technology solution for them. 388 00:26:41,934 --> 00:26:46,474 We just talk about cognitive science biases, for example. 389 00:26:46,934 --> 00:26:54,774 And the fact that one of the components of the responsible AI is human centricity. 390 00:26:54,934 --> 00:26:57,354 So you have the humans there. 391 00:26:57,374 --> 00:27:01,742 And that gets to another challenge that we have with building the science. 392 00:27:01,742 --> 00:27:08,042 for the measurement that a lot of things that we are used to on, for example, 393 00:27:08,042 --> 00:27:15,522 measuring length, mass, electricity, those areas that NIST work on coming up with 394 00:27:15,522 --> 00:27:21,642 standards for measuring them at the turn of the last century, which were critical 395 00:27:21,642 --> 00:27:25,642 for the underlying technologies of that time. 396 00:27:25,782 --> 00:27:27,374 Right now we have... 397 00:27:27,374 --> 00:27:33,174 quantitative metrics to measure them, intropable ways of measuring them. 398 00:27:33,814 --> 00:27:38,794 We are at the distant of technology, we are trying to do the very same thing for 399 00:27:38,794 --> 00:27:40,640 the technology of our time. 400 00:27:42,511 --> 00:27:45,931 As you mentioned, some other jurisdictions, the European Union in 401 00:27:45,931 --> 00:27:49,775 particular, have now a comprehensive legislation. 402 00:27:52,047 --> 00:27:53,087 states. 403 00:27:53,747 --> 00:27:59,147 Talk a little bit about what the function is of the work that NIST does in that 404 00:27:59,147 --> 00:28:03,015 context where there aren't necessarily specialized regulations. 405 00:28:03,694 --> 00:28:07,314 So NIST is a non -regulatory agency. 406 00:28:10,134 --> 00:28:13,464 We don't work with regulators. 407 00:28:13,464 --> 00:28:17,014 None of the things that we put out is for mandatory use. 408 00:28:17,014 --> 00:28:18,694 Everything is for voluntary use. 409 00:28:18,694 --> 00:28:25,174 And even the evaluations that we do while we evaluate for, for example, accuracy or 410 00:28:25,174 --> 00:28:28,430 security or privacy or any of these things. 411 00:28:28,430 --> 00:28:33,130 We don't make any judgment on how safe is safe, how private is private, how accurate 412 00:28:33,130 --> 00:28:37,830 is accurate, because there is no one size fits all and it depends on the context of 413 00:28:37,830 --> 00:28:38,490 use. 414 00:28:38,490 --> 00:28:44,770 What we are trying to do and our work with the policy makers is to provide empirical 415 00:28:44,770 --> 00:28:51,810 evidence to help with development of the policies that are clear and also make sure 416 00:28:51,810 --> 00:28:57,134 that the underlying science and technical tool and standard exists. 417 00:28:57,134 --> 00:29:00,894 to help with the enforcement of the policy. 418 00:29:00,894 --> 00:29:05,594 We just talked about it, that the laws can come and say that AI must be safe, but 419 00:29:05,594 --> 00:29:06,354 safety means. 420 00:29:06,354 --> 00:29:13,384 And we see it what happened with the EU Act as well that after its passing, now, 421 00:29:13,384 --> 00:29:17,854 since then, like European centralization body are tasked with coming up with the 422 00:29:17,854 --> 00:29:23,714 standards for basically checking the conformity assessment to the requirements 423 00:29:23,714 --> 00:29:24,654 in the law. 424 00:29:24,654 --> 00:29:25,070 So... 425 00:29:25,070 --> 00:29:26,830 So we are not regulatory. 426 00:29:26,830 --> 00:29:29,630 Everything that we put out is for voluntary use. 427 00:29:29,630 --> 00:29:37,330 But we give sort of empirical evidence for clear policy making, evidence -based 428 00:29:37,330 --> 00:29:44,010 policy making, and the technical tools and the measurements and evaluations needed 429 00:29:44,010 --> 00:29:48,230 and standards needed for enforcement of that. 430 00:29:48,230 --> 00:29:51,470 I use the word standards several times, and it may be. 431 00:29:51,470 --> 00:29:57,090 helpful if I clear that while standard is our middle name, we are not standard 432 00:29:57,090 --> 00:30:00,670 development organizations such as ISO or IEEE. 433 00:30:00,670 --> 00:30:06,370 There are really good policies in the US in effect and in existence that the job of 434 00:30:06,370 --> 00:30:11,400 the federal government is to help industry develop the standards. 435 00:30:11,400 --> 00:30:13,310 So they are in the driver's seat. 436 00:30:13,310 --> 00:30:19,530 We are supporting their work because, again, we bring those objective technical 437 00:30:19,530 --> 00:30:20,558 contributions. 438 00:30:20,558 --> 00:30:22,718 in the development of the standards. 439 00:30:23,118 --> 00:30:29,298 Back in 2019, we had a task by an executive order to write a plan for 440 00:30:29,298 --> 00:30:33,238 federal government engagement in development of AI standards. 441 00:30:33,238 --> 00:30:40,438 With this current EL, we have a task to develop a plan for global engagement to 442 00:30:40,438 --> 00:30:43,698 promote consensus industry standards. 443 00:30:44,578 --> 00:30:45,358 And... 444 00:30:45,358 --> 00:30:49,398 So we work a lot with the community, both industry and standard development 445 00:30:49,398 --> 00:30:53,938 organizations on promotion and facilitate development of the good standards, but we 446 00:30:53,938 --> 00:30:56,520 are not developing standards. 447 00:30:57,199 --> 00:31:02,939 What is your sense about the current state of global cooperation and coordination on 448 00:31:02,939 --> 00:31:04,039 AI response? 449 00:31:05,390 --> 00:31:11,110 So a lot of really, a lot of conversations has been happening for the past many years 450 00:31:11,110 --> 00:31:14,610 that I have been involved and a lot of really good work is happening. 451 00:31:14,750 --> 00:31:21,990 OECD has, OECD AI Governance Party is doing a lot of good work on bringing 452 00:31:21,990 --> 00:31:25,850 communities together but also provide a lot of technical clarity for policy 453 00:31:25,850 --> 00:31:29,490 meeting through a lot of the work that they're doing, a lot of documents that 454 00:31:29,490 --> 00:31:31,370 they're putting out. 455 00:31:31,490 --> 00:31:33,370 We saw. 456 00:31:33,582 --> 00:31:40,142 a really good progress through the G7 kind of code of conduct that try to bring more 457 00:31:40,142 --> 00:31:49,542 sort of a global interoperability on sort of voluntary commitment that happened in 458 00:31:49,542 --> 00:31:50,842 the US. 459 00:31:51,082 --> 00:31:56,402 Bottom line is that global conversations and global interoperability for the AI 460 00:31:56,402 --> 00:31:58,682 space is really important. 461 00:31:58,682 --> 00:32:02,542 It is for any type of standards, but particularly in the space. 462 00:32:02,542 --> 00:32:10,362 The technology is obviously very global and the users are across the globe. 463 00:32:10,362 --> 00:32:17,102 So we are thinking about really maximizing the benefits of AI while minimizing its 464 00:32:17,102 --> 00:32:21,562 negative consequences, harnessing the power of AI for good. 465 00:32:21,622 --> 00:32:30,638 We need more global engagement, both at the research side to build. 466 00:32:30,638 --> 00:32:34,138 the scientific underpinning on the standard side. 467 00:32:34,138 --> 00:32:41,078 So we all are kind of using the same standards and also on the implementation 468 00:32:41,078 --> 00:32:42,818 and operationalization side. 469 00:32:42,818 --> 00:32:46,338 So we can all benefit from this technology. 470 00:32:46,338 --> 00:32:51,138 So a lot of really good work is happening and a lot more can be done, I guess. 471 00:32:51,855 --> 00:32:52,285 Absolutely. 472 00:32:52,285 --> 00:32:54,435 Well, that's a really great place to wind up. 473 00:32:54,435 --> 00:32:56,863 Thank you so much for a really fascinating conversation. 474 00:32:57,262 --> 00:32:58,232 Thanks for having me. 475 00:32:58,232 --> 00:32:59,618 Enjoy the conversation.