Washington AI Network Podcast Ep. 19 Tammy Haddad Interview with Elham Tabassi Recorded at the Special Competitive Studies Project’s (SCSP) Ai Expo For National Competitiveness May 7, 2024, Washington, Dc

Tammy Haddad: [00:00:00] Welcome to the Washington A. I. Network podcast. I'm Tammy Haddad, founder of the Washington A. I. Network, where we bring together official Washington, D. C. Insiders, and A. I. Experts who are challenging, debating, and just trying to figure out the rules of the road for artificial intelligence. This special episode features a keynote interview with Elham Tabassi, Chief AI Advisor at NIST and Chief Technology Officer for the new U. S. AI Safety Institute. The conversation you're about to hear took place at the Special Competitive Studies Project's inaugural AI Expo for National Competitiveness on Tuesday, May 7th in Washington, D. C. Join us as Elham Tebassi brings us up to date on the US government's AI efforts and answers audience questions. I hope you enjoy the conversation. [00:01:00] Thanks for listening. Tammy Haddad: Welcome everyone to the A. I. Expo for National Competitiveness. Let's thank Illy and all our friends at the Special Competitive Studies Project for putting this together. We're so glad you're here. And it's such an honor to talk to Elham Tabassi, who you guys know is the chief A. I. advisor for NIST. It's the most powerful, transformative technology of our time. You've played such a critical role for so many years. Welcome, welcome, welcome, and I'm going to start by taking you back to that day when ChatGPT came out. First of all, was that a surprise to you? Elham Tabassi: First, thank you very much for having me, delighted to be here, and it's great to see many friends in the audience. So, ChatGPT. When they came out, a year before that, we had seen [00:02:00] release of several similar softwares that can take, for example, the software such as Dolly that can take prompts as text and generate images that very much look like the real world images. So we were kind of familiar with the technology and about that time, we are also working on finalizing the AI Risk Management Framework. So we paid a lot of attention to where the technology is going. Particularly with the eye towards making sure that AI risk management framework is can be as evergreen as possible, knowing that the technology is moving really fast. We know that it's not going to be, revision is going to be needed, but we didn't want to have a very short shelf life. So A.I.R.M.F. Came out in January in November. Chat GPT three went out in February. Chat GPT four came out and then basically took the oxygen off every conversations on every room. So what we did immediately after that put a group of experts together, start reaching out to experts outside of NIST, interviewing [00:03:00] people, asking the question that is the guidelines AIRMF that we put out in January is still relevant or immediate changes is needed. And what we heard from the community is that, no, it is relevant, but we also need to learn risks that are unique to this technology, to large language models, to generative AI models and think about kind of risk management for the risks that are unique to generative AI or uniquely exasperated by generative AI. So we quickly put a public working group together to look at those type of risks. We're working with them over the summer until the executive order came out and basically give a more urgency and speed to the work that we were doing. Tammy Haddad: And how do you pick who's in that group, the working group of that researchers, private industry, government, who is that? Elham Tabassi: At NIST, we have a very talented pool of experts and scientists and all of them with very much committed and with a love for public service, shout out [00:04:00] to public service. But we also know that we don't have all of the information, so we reach out to the broader community. For the public working group, we basically it was open to any individuals that wanted to sign in on and be part of this being part of the government, there are some sort of a kind of security that make sure that certain individuals from certain countries are not part of the work, but our work is open, transparent, uh, and kind of reachable and accessible to the broader community for them to use. Elham Tabassi: So for that public working group, it was everyone, but that is also a hallmark of the way we do the work. We, again, have a lot of really good experts, but we also know that we don't have all of the answers. So we reach out to the community through many formal mechanisms, such as a request for information. Elham Tabassi: We run workshops. We go to the meetings. We try to get all of the inputs. And when we reach out to the community, we make sure that we reach out to the experts, a very broad and [00:05:00] inclusive and diverse set of thoughts and backgrounds and expertise. Of course we want to hear from technology developers that's usually encompass expertise such as computer science, mathematicians, statisticians, engineers, but also the community that study the impact of technology. And that's the psychology, sociologists, cognitive scientists, because AI systems are more than just data compute and algorithm. A complex interactions of data compute an algorithm with the environment with the human. So it's important to have the socio technical lens. So That's how we do it, and in our all of work, we make sure that we have all of this community involved but also The experts in standard development so that we can give the right input to the Research community but also policymakers. Tammy Haddad: Six months ago the president announced His executive order, which literally realigns government and people working in AI, basically saying everyone's got to have someone in AI. How has that changed what you're [00:06:00] doing and how do you support those efforts? Elham Tabassi: So executive order, as you said, is a very comprehensive set of instructions to federal government to be basically an example for advancing trustworthy and responsible use of AI. I'm told that this is the longest executive order in the history of executive orders, but definitely in terms of it's breadth and depth and the number of the instructions and tasks to different agencies. Tammy Haddad: How many people know how long the executive order is? Tammy Haddad: 111 pages, I think it is, right? Elham Tabassi: That's right. That's right. the Federal Register notice, by cutting out the margins and making the font smaller, they made it to 87 pages. But this, yes, the one that you look at it is that long. So, we got several tasks as part of the executive order, and the tasks that we got really built on the foundational work that we had been doing on understanding trustworthy implementation and adoption of responsible AI. The work that we have been doing around A. I. Risk [00:07:00] management. And I like to say that those direct tasks on the timeline that we got through the EO really supercharged our effort to work with the community stronger and faster and better to release guidelines for test and evaluations for secure software development for synthetic content authentications, for global engagement in promoting technical standards, and I'm very pleased to say that last week on Monday, April 29, we put four of the draft out for public comment. They are out for public comment to June 2nd. So to all of you that are already looking at it, thank you for all for the rest of you. Please check out our website. We really like to hear from you guys, and we are on track to deliver all of them by the deadline of the year, which is end of July. Tammy Haddad: And what's the website? Anyone have done it yet? Public comments? Okay, now's your chance. You're hearing it directly from Elham. Elham Tabassi: So, send a note to me, but if you [00:08:00] just go and Google NIST AIEO, it just take you directly to the website that has all of the information. Thank you. Tammy Haddad: Six months have passed. Can you talk about how the technology itself has changed in that period of time? Elham Tabassi: Yeah. So the technology, you know, the pace of change and also adoption of this technology is nothing like we had seen before. The speed and scale of this technology is just one of the many challenges that we have, and this pace is, frankly, much faster than any evaluation standard work or even policy can keep up with that. So that all means that we have to work faster and work with the community harder and also know that as all of these things, uh, changes, and that's really true for any scientific endeavor, we learn by doing. So let's just work together, let's just figure out the challenges, come up with the solutions. And knowing that we need to iterate, we are not going to get the perfect answer. So we get some answers, we [00:09:00] learn something and let's build on that and do more. Tammy Haddad: It sounds like an exhaustive process. How do you keep it all organized? I mean, are you on, I don't know, Excel sheets, Google Docs, you know, do you like 10 researchers say this, five say that? How does that work and how is it collated? Elham Tabassi: All of the above and more. But I think a lot of us, I think it's the sense of the mission and a lot of us really believe in what we can do to help this particular, sort of the crossroad of the technology, science and policy. And really we, many of us really feel a moral obligations to make sure that AI can be, the power of AI can be harnessed for good while its negative consequences and harms can be minimized. So I think this compelling mission and on the role that we are playing, but also seeing the impact of the work that we do. We put AIRMF out in January and in weeks after that, we were hearing that people are implementing it and hearing good things about this. Elham Tabassi: So these are all really strong reinforcements [00:10:00] for us to keep going. And of course, the work of the whole community. So grateful for all that all the support. Tammy Haddad: So managing the risks of AI? What's the process for that? You guys really started it years ago when you put together the NIST framework. But where are you now on that? Elham Tabassi: Yeah, I want to say a few things before directly answer your question. So we often hear that risk and risk or safety can be in contradiction innovation. But really what we are doing and have been doing all along. Is figuring out how to manage the risk and negative consequence of AI systems to unlock innovation to harness the power of AI for good. So really that's what we're doing. This is the journey we are. And if you look at the definition of risk in the AIRMF that, as you said, we did it years ago. The definition of risk is a composite measure of the probability of an event happening and the consequences of that event. [00:11:00] The consequence of an event can be positive, negative, or neutral. That makes, as strange as it sounds, risk can be positive, negative, or neutral. All this means that risk management becomes about maximizing the positive while minimizing the negative consequences at the same time. So when you're doing the risk management, the risk of not using AI, the risk of not using technology, become one of the risks that needs to be understood and managed. So, I just want to mention that all of these things is really to unlock innovations, to bring the community to find, to work on the challenges. And our job is that through evaluations through other work that we do, put some empirical knowledge and evidence on the limits and capabilities of the technology, quantify the risk and impacts that there are so the community better know where to put the efforts on advancing research, advancing innovations, and tackling the challenges with risk management. Tammy Haddad: So you're saying that innovation actually helps you with risk [00:12:00] management? Elham Tabassi: Risk management is really for innovation, yes. Yes, these are really, you know, the better we know the failures of the technology, that's where the more innovations can happen. And we have done it with other technologies before at NIST. If you look at, for example, on the evaluations that we have been running for biometrics and face recognition, the first series of the evaluations that we did, face recognition algorithms fail if the person was not looking directly into the camera, that problem is solved, of course, because of the technology improved, but quantifying the failure modes, quantifying the risks, quantifying what are the problem with the technology really gets the developers, scientists, researchers back on the drawing board back on the white board to come with better ideas and innovation. Tammy Haddad: Let's turn and talk about authentication. There's so many different points of view. Where are you coming from on these issues? Elham Tabassi: Thank you. So this is important to say that at NIST we don't have really [00:13:00] preconceived answers. We do the science, we work with the community, and we let the empirical evidence and science lead us to the answers. So in this particular topic, which is a very important topic and a lot of really smart people are looking at that. We actually do have a task as part of the executive order that starts by just do a landscape survey of what are the techniques, measures, tools for synthetic content authentications. And, as many of the reports on the work that we do at NIST, we like to actually start with terminology and taxonomy at the lower level because it's really important to bring the whole community on a shared understanding about the topics and the word that we use. It's often the case that we use the same term, meaning different things. So the report is out. You can go look at it and, we look at it and bring the community around the discussions about digital content transparency. So you know, you all have heard watermark, you had seen those kind of draft that goes across the page. All of those things are [00:14:00] examples of ways for the watermark or actually trying to say something about the documents and the content of that documents. So the report that you will see is just start talking about digital content proven, digital content transparency. And in that context, talk about the role of the metadata tracking or provenance data tracking. That's really basically recording and tracking the origin and history of a piece of content. This by itself is not going solve the problem of transparency, but that information is really important to basically give some information about where this piece of content came about, if it was human generated, AI generated, or, manipulated by AI. And then there's also another track of synthetic content detection. So if there is a content, then our algorithm can look at it. Where a human can look at it and make a judgment that it was generated by human or by an algorithm. Tammy Haddad: Will there ever be a time that a regular person, not a scientist, will be able [00:15:00] to look at something and know if it's synthetic or original content? Elham Tabassi: Yeah, that's a great question. And, our hope is that provide that type of transparency to get there. Of course, there are several technical challenges to solve that. There's also a lot of questions around digital literacy and how to actually share that information, assuming that we have a technical solution to get to how to actually make sure that the content is human generated or AI generated or manipulated, but how to present that and how to communicate that with humans. So all of these things are on the table to discuss. And I stay hopeful that. Yes, the science and the scientists and researchers and working the whole community, we will get the solutions. Tammy Haddad: Do you think that AI will ever be in schools, that kids will learn from a young age what AI is, what's synthetic, what's real? Is there any sense of that now or anyone who's working on it? Elham Tabassi: So I think, my children are grown [00:16:00] ups, I don't know, but, I think people are using it and children are using it and... Tammy Haddad: but I mean the education part... Elham Tabassi: you know, my phone, any technology is that these are powerful tool, and by right safeguards and right use cases and right set of kind of how to use them, but also how to build them safe. Yes. I always say this, that we want technologies that's easy to do the right thing, difficult to do the wrong thing, and easy to recover if somebody inadvertently did the wrong thing. Perfect example of that is the three pin plug, electricity plug that we had. It took some time with a lot of, you know, maybe even fire or something that we get there. But if we look at it this way, and for the AI also approach it this way, that instead of putting or thinking about the safeguards that we want to put after the technology has been developed, um, Around the use and through the policy and regulations. Try to fix them. Yes, that's one way of doing this. But also, let's think about how we can actually build the technology. [00:17:00] That's all of these questions and challenges can be tackled as part of the design and development of that. So, I stay hopeful that yes, AI would be as seamless in everyday life as possible. And, along the way, we have also solved many of those challenges either through technical measures, socio technical measures, better educations, and, again as I really want to say that we get AI to work for us and improve our lives. Tammy Haddad: I want to ask you about the NIST Gen AI Challenge. Elham Tabassi: Oh, thank you. Yeah, thank you. Thanks for asking that and giving me the opportunity to brag about this. So, one of the things that we do at NIST is run evaluations to, again, get a sort of a sense of limits and capabilities of technology. And we have done it for many different type of technology. As part of this broader evaluations, another type of activities and efforts that we run is what we call challenge problems. And the goal of that is really to encourage and inspire researchers and technologists and engineers to [00:18:00] come up with solutions. So the Gen AI challenge that was also announced last Monday has two tracks. One track invites submissions of algorithms that can generate synthetic content that can look very real. So the first track that we are doing is text to text. It gets a text and generates text, or it gets a series of documents and generates summarizations. And we're gonna go and look at them and evaluate them to see how good the algorithm can generate those synthetic texts. Tammy Haddad: Are these new companies or researchers or individuals or is this like OpenAI, NVIDIA, whatever? Elham Tabassi: Our test is open to everybody, and we are hoping to get a very broad participation. But the Gen AI has another track there too and that's a discriminator that gets a piece of text and try to make a judgment that this was generated by human or algorithm. So by looking at both of them and getting the best outcomes from the generators to the discriminators. Or [00:19:00] applying the best discriminators on the results of the generators. We're trying to basically do a sort of a blue team red team to improve both of those things. So, participation is open to everybody. The only thing is that they have to be able to implement the API that we have. But if I can judge from the past evaluations that we run, industry participates, academia participates, small and medium sized companies participate. And that's really our job. Kind of convene the community around problems and challenges that relate to all of them and try to advance them. Tammy Haddad: Are you concerned about China or other bad actors participating in this program? Elham Tabassi: I do believe in open science, and I do believe in that knowledge is always really a great thing. So it's better for us to know what they are doing and what their capabilities are rather than us being in dark. But as I said, as part of the participations, the entities that sign up for our test go through some vetting process, regular government vetting process. [00:20:00] And if there are foreign nationals, they go through extra vetting process that's done by experts at NIST or outside of the NIST to make sure that national security is considerations is all preserved. Tammy Haddad: And how big is your team? This sounds like a heavy lift. Don't you guys think? It sounds like a lot. Elham Tabassi: So our team is growing on. We like to grow more. I always say that we are really limited by the number of the hours in a day and the number of the people that are working on this. And there's only one degree of freedom in that equations and the number of the people. But we also increase our capacities by working with the whole community. So when we put the document out for public comment and we hear from hundreds of entities and organizations outside this is that really expanding our capacity and recently also many of you know that we established or stood up a consortium as part of the Safety Institute on the number of the entities that are part of the consortium are [00:21:00] around 270 and increasing. So that gives us another way of really working with. Tammy Haddad: Who were the 270. Elham Tabassi: Basically, I want to say almost any actors in the A. I. All of the big names of industry are there. We also have smaller businesses that are making tools or working in this area. We have universities, we have civil society, we have experts from the verticals, you know, health care, finance, all of them. The name of all of those are actually posted on our website. We really truly believe in being open and transparent. So everything is on the website. Tammy Haddad: Okay, we have a question? Audience Member: You talked about putting a watermark on things that have been created, let's say, synthetically or in some way that they're not appearing, you know, naturally or genuinely. What happens in the case of of the upcoming presidential election when we [00:22:00] have, foreign, countries like Russia, for example, and others, putting AI developed images and other things out there that contain misinformation and are misleading and what can this do about it and is that something you're working on? Elham Tabassi: Thank you for the question, and that's a really important topic. I want to say a few things first that NIST is a non regulatory agency and our job is a really work with the community to develop guidelines, run evaluations, advanced measurement science and standards, measurement science and standard that make technology more reliable, more robust. We don't do any certifications. And again, we are not regulatory. So in the context of the synthetic content and digital content provenance, we're really working with the community to, A, understand what's happening right now, what's the knowledge space looks like right now, and then after that, work with the community on developing [00:23:00] possible guidelines for those reasons. So development of the guidelines is not something that we are currently doing, but your question is a lot broader than that. Your question gets into misinformation, disinformations, and distributions of those type of information, which is several steps away from the work that we are doing. Those are really important conversations, but our work doesn't get into misinformation and disinformation. So, as I said, the, for example, content provenance can give information, you know, provide in transparency about origin and history of a content but, by itself, Doesn't quite ensure or assure trustworthiness. Audience Member: Where in the federal government, would you say is focusing closely on the misinformation, disinformation issues that we're all going to be facing? Tammy Haddad: That's that's really CISA that's that's doing that. Jen Easterly and her team at CISA. Any other areas? Elham Tabassi: Yeah, I would. Yeah, I would say that there are, there is work [00:24:00] across the government agencies that are working on this. Definitely, CISA is one of them more on the operational side. CISA has the responsibility of basically looking after elections being fair and robust. But our job right now is more around, again, the tech, everything that we do is around technical guidance and socio technical guidance, but we are several steps before. Tammy Haddad: Is there another question? Audience Member: You mentioned that there are positive and negative risks. Are there one or two of each of those that you'd be able to share with us from kind of the, the scientific data perspective? Elham Tabassi: Yeah. Audience Member: I mean, there, there's some obvious ones, content, but just from your perspective, are there one or two that you could share at each? Elham Tabassi: Thank you. So in the AIRMF, we actually come up with trustworthiness characteristics. So it's sort of the flip side of the risks and it talk about the characteristics of the system that you want to see to make them trustworthy. So if you wanna talk about risk, some of those things becomes for example, harmful bias that AI systems can exasperate the [00:25:00] inequities in the society and all of this. And if I want to say a positive risk of AI systems is broader access to, for example, a lot of resources. I also want to say that as part of the executive order tasking that we had, one of the documents, one of the four documents talks about the 12 risks that are sort of unique or uniquely exasperated by generative AI. So, hallucinations, giving violent or toxic recommendations, all of those things are being listed as part of that. But other things are, for example, things such as data privacy, information security, information integrity, these are all the risks. Tammy Haddad: Has there been any progress on hallucinations? I've read some recent stories in the last couple weeks that the hallucinations are in a decline. I mean, if that's something you can really address. Elham Tabassi: That's a good news if that's being declined, and I think that's what happens when the, when we have more information is about where the systems fails, and that was that's sort of positive feedback that I talked that developers use this things go back to the drawing [00:26:00] board. We haven't done the, at least we like to have a real empirical data to talk about it, and we haven't done that test yet, but yeah. Audience Member: Can I ask a follow up to that? Tammy Haddad: Please go ahead. Audience Member: So from a generative AI perspective, part of the challenge is that Federal agencies have very clearly, legally reviewed content that is available on their websites. Whereas a Gen AI may actually interpret that and then explain that in plain language to or in a different language even. Is there any risk that you see or that you're tracking in that alteration of approved language? Elham Tabassi: If your question is about LLMs getting a text and then provide a summary of that, and if the summary is accurate, hopefully the Gen AI challenge is going to look at that. So, stay in touch. Thank you. Tammy Haddad: Okay, one final question. What about the plan for global engagement? Because it's hard enough to get [00:27:00] everyone in the United States to be on the same page on anything, let alone AI. So, how are you working on the global engagement side? Elham Tabassi: Yeah, that's a great question and thank you for that. So, another task that we had as part of the EO was developing a global engagement for promotion of AI standards. AI standards is becoming more and more a topic of interest too, and I keep saying it. I go to the standard meetings since 2005 and everybody rolled their eyes because it was, it is, it can be a very boring topic, a lot of discussions. But it really warms my heart. That now standard is a really important topic and some area on AI standard ecosystem is where that can really inform AI actors on how to enforce and Implement a lot of regulations. So for example, we have seen you AI are passing regulations or any other regulations. You know, they say that the the text of the law stays as very abstract level of AI must be non discriminatory, AI must be safe, they don't get into explaining or saying what [00:28:00] they mean by non discriminatory, what they mean by safe. Or even more importantly, how can we test and assure that it is safe or non discriminatory? And in the absence of having an ecosystem of standards that address these questions and provide tools on how to look at these problems, these questions can be adjudicated in courts and that's not the way we wanted to go. So in the development, in the global plan for engagement in standards, we talk about activities before, during and after standardizations with a lot of emphasis on AI is a global technology. Data doesn't know border technology, doesn't know border. And, we need to work with the partners around the world to advance one set of interoperable standards, one set off rule of the roads for AI safety that helps with more interoperability across the borders that can improve trades that can improve the science, and ultimately more economic growth and prosperity for everyone. Tammy Haddad: [00:29:00] That is fantastic. Thank you, Elham. Thank you for listening to the Washington AI Network podcast. Be sure to subscribe and join the conversation. The Washington AI Network is a bipartisan forum bringing together the top leaders and industry experts to discuss the biggest opportunities and the greatest challenges around AI. The Washington AI Network podcast is produced and recorded by Haddad Media. Thanks for listening. Tammy Haddad: Okay, I want everyone in here to stand up. Elham, come on, we're going to stand up. And I want you guys to all Come on, stand up, get your phones up. Let's move, it's an AI Expo, you have to move fast. Okay, and I want you to put your phones up and take a picture of Elham, come on. Yeah, yeah, yeah, just stay right there because she started off Okay, and now take a picture, everyone together. One, two, three. [00:30:00] Okay, now everyone turn around and get yourself in the picture with Elham. So you can say you were at the AI Expo with the Chief Technology Officer of NIST. Thank you Elham so much. Much appreciate your time.