From Action Plan to Mission Impact: Humans Still Required

In this episode of INSA's From Action Plan to Mission Impact: Key Takeaways from America's AI Action Plan podcast, hosts Chitra Sivanandam and Dr. Yevgeniy Sirotin sit down with Dr. Missy Cummings, Director of the Mason Autonomy and Robotics Center at George Mason University, to examine why workforce readiness is critical to successful AI adoption. Dr. Cummings discusses the growing gap between AI capabilities and AI literacy, the risks of overreliance on large language models, and the importance of understanding when AI is the right tool for the mission. They discuss continuous reskilling, leadership responsibilities, and practical strategies for building an AI-ready workforce capable of applying these technologies responsibly in national security environments.

Transcript

Yevgeniy 0:05

Welcome to the All Source Podcast from the Intelligence and National Security Alliance. My name is Yevgeny Sorotin.

Chitra 0:11

And I'm Chitra Savanadham. This episode continues our series from Action Plan to Mission Impact, where we examine how the AI Action Plan is being implemented across the national security community.

Yevgeniy 0:23

Today's episode is Humans Still Required. The AI Action Plan makes it clear that workforce readiness is a cross-cutting priority and decisive factor whether an U.S. can actually capitalize on AI.

Chitra 0:37

And this means moving beyond hiring specialists to preparing every operator and leader to effectively use and trust these systems in mission environments.

Yevgeniy 0:46

So today we're going to be looking closely at where those gaps exist today and how AI literacy is becoming a baseline requirement for every role, not just an elective.

Chitra 0:56

Thanks, Yevgeny. And uh joining us today to break down the hard challenges of adoption and workforce pipelines is Dr. Missy Cummings, Director of the Mason Autonomy and Robotics Center and professor at George Mason University. Missy, thank you for being here. Thanks for having me. So America's AI Action Plan is built on a couple pillars: innovation, infrastructure, and diplomacy. And it identified promoting American workers as a cross-cutting policy that underpins them all. Why do you think workforce readiness is the decisive factor as to whether the US can really capitalize on AI or not?

SPEAKER_02 1:34

If we think that AI is an amazing new capability, which it can be, depending on how you use it, it's really important that people understand when to use it, how to use it, but also when not to use it, right? So AI is a tool in a toolbox. And I think that people's AI literacy skills were severely lacking before large language models and frontier models showed up. And I think this new capability has only accelerated that gap between what people think they know and what they know.

Chitra 2:17

Yeah, actually you're right. The more there is, the bigger that gap is.

Yevgeniy 2:22

Yeah. And I the point of follow-up that's folk knowledge, right? So there's some research out there that I've been looking at where, you know, in the absence of like really strong formal education, sometimes people develop folk notions of how the systems are working. How do you think that shapes people's ability to leverage those technologies correctly in terms of understanding where they might go wrong and things like that?

SPEAKER_02 2:49

I think folk knowledge, when we're talking about it what we might think are unskilled people, people who may not have a lot of education in this space, I think that they have a cause and effect understanding of systems, which is just generally how humans reason when they're presented with a new system. I am less concerned about the folk knowledge that is coming out of people's understanding of AI as opposed to the wishful thinking that is and the magical thinking that is coming out of people with PhDs who should know better, but I think are desperate to sell a product. And in a, and I think sometimes if they talk about a technology, their CEOs want to believe that a technology can do what they're saying, even though they've had all the formal training to understand that it shouldn't be do, it's not doing what they're saying it is, but they want to believe. And I do think that we actually have alongside AI, which is a really useful tool in the right situations, we have a religion of AI that's growing up just alongside it. And I that's dangerous because if people don't understand how AI really works, then we don't put the money in the right places to make sure that we fill in the gaps of where it is actually having problems.

Chitra 4:19

Yeah, that's interesting because I um I spent a lot of time with um with emerging tech companies and and startups. And I think there's this notion around there that probably has some decent, you know, founding for it, um that you inherit less liability if you don't know all the dirty details than if you do, right? Um and I think it's like maybe it stems from like um I think there's some things related to like cyber risk as an example, where if you knew there was a hole, um you the the the amount you have to pony up if that's exposed is different than if you didn't know it existed, right? And I fear if that understanding and that kind of really kind of pushing towards like understanding like where am I exposed, what is that what does it actually do? What is it not doing? Um, where do I have to limit it? Like maybe, maybe it all stems from that same in com same kind of commonplace understanding of I'm better off like not knowing the unknown as opposed to like digging deep into trying to like learn more. And I don't know, it's just an interesting theory around why there is a lot of that, right? Why we don't see as much of that, like, you know, hungry learning for the unknown unknowns, right?

SPEAKER_02 5:35

I don't think it's a theory. I think it's solidly embedded in practice. And I have seen many companies do this. I occasionally do expert witness consulting, but I also work for companies as a consultant. And across the board when it comes to AI, less testing is better. I have heard it specifically said we are not going to engage in this set of tests because we don't want to know what the outcome is in case we get sued. So I think that's here and now.

Yevgeniy 6:06

So we're we've we've had uh an interesting episode on AI testing, our previous episode, where here I'm curious to know, you know, as it comes to workforce and getting ready to actually interact with these systems, because they're develop the developments are coming fast and a lot of organizations sort of are adopting these systems. What is it that sort of we need to make sure our workers understand about these systems to be able to leverage them in a responsible way and to sort of have our workforce ready to get some of the benefits that are being offered, because these systems do offer certain things and while sort of avoiding some of these risks that you've identified. Trevor Burrus, Jr.

Chitra 6:47

Trying to close the gap.

Yevgeniy 6:48

Yeah.

SPEAKER_02 6:48

Yeah, I I think that it's almost impossible to disambiguate and pull apart testing with the human experience, because even if you don't do formal testing, I'm gonna tell you every single one of your employees is going to actually do testing, whether you think it's that way or not. We do a lot, like for example, this semester, I'm teaching a class on how to engineer AI systems. So, how can you use a large language model to help you design and build regular systems or agenc AI, for example? And inevitably, when the students are using AI to help build AI, the issue of hallucinations comes up. And they will find it, they will dig. And I think this is a problem with companies trying to avoid what is the elephant in the room. If you don't embrace the fact that these technologies are imperfect and you don't put in the learning suggestions, guidelines, whatever to go along with it, the people are going to find these problems. And then they are going to not only distrust the technology, but also distrust whatever part of the organization is doing the training and or putting forth the requirements for whatever system they're working on.

Chitra 8:13

So that's good. And that's I feel like there's a really strong, um, I guess hopeful perspective on like the the developers and um their kind of acumen on being able to think this through and maybe maybe less so on the leadership, but some some some stuff there. What about the the average user that's using this stuff? Like what do they have to do in terms of the adoption and use of it in the workforce? And you know, where do you see like the bigger gaps that exist on the national security front in terms of the folks that this stuff is coming down to support?

SPEAKER_02 8:46

I think that what I see in the workforce, and when I say the workforce, I'm talking about college-educated people working in and around the government. They are on a mandate to use these technologies. This is not just a government thing. For example, my university, George Mason, has mandated that we're going to use these technologies. Which, you know, I think it's very important for an educational institution to be able to teach about these technologies with these technologies, because I'm gonna tell you the number one problem that we're having in universities, it's not just George Mason, this is across the nation. My grades in my classes, which in the past year, where generative AI has really come on the scene, have gotten demonstrably worse, even though the class is being taught very similar to time in the past. And what is happening is that people, this goes back to the faith-based feelings about AI. They want to believe that AI can fill in the gaps. They overtrust AI. You know, students anywhere, and this is not just undergraduates. I mean, this is all the way through the PhD ranks. I think you we see a lot more of the magical thinking, thinking that the AI is going to be able to do the things that it has promised from whomever that it can. And so I've actually completely gutted the way I teach now because of generative AI. Indeed. In a way, we're only going retro. We only have exams old school, like with the blue books. Oh my gosh. And like there's no technology. And uh, but and these are for technical classes. And so it's not a bad thing necessarily, because it's making me think harder and making me better as a professor. But I mean, that's not sustainable. If the bulk of students, and again, this isn't just a George Mason problem, it's a nationwide problem. If people want to become over-reliant so quickly on the technology and they're not able to do the critical thinking, then we're going to have real problems. And that similar over trust in these technologies is echoed across all companies that I talk with and work.

Yevgeniy 11:08

When we talk about this kind of automation bias, I think, at scale with these LLMs. So the AI in itself is not necessarily now a new technology. It's these these new large foundation models are kind of new. Um we've had in other domains, AI systems that perform superior to human, say perceptual performance in many cases. Can you talk about what's really changed at the core? You know, are is it just a matter of m accessibility that these are chatbots available to everybody? Is it the expansion to like basically all domains because these foundation models can do such a wide variety of tasks? What's really changed in your mind regarding these AI systems that are now available to us and how has that sort of impacted the workforce readiness of people to interact with these this AI as opposed to like say the AI that existed prior to I don't know, 2022?

SPEAKER_02 12:06

Aaron Powell If you've ever heard me talk before, you know I'm kind of the classic AI curmudgeon, and I'm going to lean right into that and start to fight with you. You're like, oh I would say, first of all, foundation models have not sh been shown to be everything all the time.

SPEAKER_00 12:23

True.

SPEAKER_02 12:24

They indeed, the research is increasingly pointing to the fact that more narrow models are going to be the way that we perform in the future, especially if you have specialized knowledge. So you can't have, you know, the gargantuan models that are trained on the whole internet also have a lot of garbage in them. And for uh and I we've been been doing research in my lab. If if you want to generate code and you're using a large language model that has access to everything on the internet, I mean, we can just set aside the cybersecurity issues, which are huge. It also turns out there's a lot of garbage that it's picking up. And so the larger models are not necessarily better for specialized domains. But if we want to set that aside and say, okay, well, what about more broadly, office work? Okay, that's a kind of a general task. It may be that some large language models are very good for, you know, I do see a lot of the number one thing that people tell me that they like to use large language models for in companies or to write their mission statements and all the documents that go around that. Yeah. I mean, that's perfect. It's a great application. And I personally use uh AI a lot to help me summarize documents that I don't have to necessarily worry about the fact-based content, that it doesn't have to be, you know, there's not numbers in it that have to be right, but just themes that have to be right. So I mean, I think that that's that is good, but I think where the problem is, is that people are being told by companies who want to make money off of these models that they can do all of the things that they say they can do. And for example, where we see students really having trouble trying to use large language models to do mathematical analyses, you know, it's a language model, not a calculator. And, you know, some of the language models have APIs that tie them into mathematical models that, and that's great. But I dare say that both in university settings and in company settings, if you were to ask somebody, is the large language model you're working on, does it have a mathematical tie-in, or is it, are you just asking it by itself and it and it ha has no idea? Um but you know, they're sycophants, so they're not gonna tell you they have no idea. Most people, even pretty savvy people I know in the government, don't know what the capabilities of their whatever tool that they're using are.

Chitra 15:15

Yeah. That's an interesting thing because I think that that might be like we were talking about this um in a previous episode about like not having the right language or um a tight enough language based on how we describe things. And I feel like with AI being like almost too synonymous these days with LLMs, um, it's gotten worse because then you can see how across the workforce the gap gets even worse and worse and worse because they all think they're talking about the same thing, but they're actually all talking about completely different things. Right. And then we don't have a way to say, especially as we go into more agentic frameworks, like, hey, I actually have like so many different things actually interoperating with each other. But if I just bin it all as AI things, right, I I've almost dumbed down the lexicon to a point where you're creating more of a knowledge gap, right? And and the general workforce doesn't appreciate the difference.

SPEAKER_02 16:06

This past year, I have said the phrase Occam's razor more times in my entire life than uh just in in just a few short months. It because I I am teaching a class on how to engineer systems, and we forget that the best system is the simplest system. And because I'm a curmudgeon, uh, it goes with the territory for me to remind everybody that GoFi, good old-fashioned AI that is rules-based, also works pretty well in most situations. And it's easy to test and you know what you're doing. So I feel like because it's a newfangled tool, and this is, you know, we can even separate out large language models. I mean, it's just the it's just the new fad, right? And people are gonna overuse it and use it in ways that they shouldn't be using it. And it just, you know, look, we actually had some pretty good tools before frontier models became a thing. Maybe we shouldn't use it. And it's okay that you don't use it, but I do think that that's kind of where we're losing the thread is that AI is not just one thing. And you really need to understand the whole field of AI tools. AI is a huge umbrella. And if you think AI is all foundational models, I think that students especially don't appreciate that that's likely going to get them fired faster. Misapplication of a large language model at a time and a place where you shouldn't use it, that's that carries real threat of losing a job later in the workplace.

Yevgeniy 17:39

Aaron Powell So it sounds like what one of the things going back to workforce readiness is it sounds like what you're saying is that one of the things that the workforce needs to understand are the limitations to the performance that they can expect from these models in different domains.

SPEAKER_02 17:58

Aaron Powell Different domains and different layers of complexity. You know, using a tool to summarize something can be very helpful. However, if you're trying to use it, and this is where my students get into trouble, is we will do labs and they have to fly drones around, for example, and they have to gather data. And then they shove all this data into a large language model and have it summarize that data, and it's a hundred percent not correct. And I think the bigger issue to me is A, I teach it to them. I I've told them right before the lab, whatever you do, do not put this in a large language model because it's not going to give you the right answers. It's the first thing they do. They will still do it. Okay. I, you know, I'm just that's the thing, every I get it. And every manager needs to understand that that is going to happen. But even more alarming than that is they don't check their work. And I think that this is the bigger issue across so many settings. Students will turn in something to me. I cannot tell you the number of times that they will turn in assignments to me. And somewhere in whatever the assignment it says, I'm sorry, I can't do that. I'm a large language model. And if you can't if that happens, it's an automatic F in on whatever assignment you get. And um and it happens all the time. And I tell the students at the beginning of the semester that same story, and they still do it. They still do it.

Chitra 19:28

I mean, uh so this brings up a really interesting point. So uh in the AI action plan, um, they talk about kind of worker first in terms of the the AI mission readiness and um what that means in terms of like broad literacy. And I think it's the same thing. Like we have this, these same kind of anecdotes and run rules on like do this, don't do this. Like these are the the gotchas that we've seen now over the last couple of years, and we still see them. So how how do we kind of overcome that hurdle? And for the national security environment at least, figure out like what does this worker first strategy with AI augmenting everybody and everybody being aggressively pushed into like use it and use it and use it. And if they're behind on their literacy and they still use it, like there's implications for national security. So how do how do we kind of get past that and what do we do about the workforce?

SPEAKER_02 20:20

Aaron Powell We could be here all day trying to lay out the roadmap. I would say any company at a first pass needs to understand what the reskilling roadmap should look like. Because you have I'm gonna take out junior employees, put them aside for just a second and say you've got your middle and higher level employees that are only understanding AI from what they're hearing in the news, read in Wired magazine, you know, and hearing the AI companies push, which is of course very optimistic and not at all realistic. So companies have to first put together a reskilling strategy. So I think we've got to fix, fix the messaging at the C-suites, truly reskill your middle managers, and then companies, organizations, governments, uh NGOs, when they're bringing in new people, I I think that they really need to do a good job of giving feedback to the universities where they're getting these people from. Because universities are of all of these institutions, we're the slowest to change. And so, but what will drive real change in a university is when a company comes back and says, your employees do not understand AI. Your employees use it when it's not appropriate. They don't know how to double check, they're not literate, you know, any kind of fee. So I when I travel around the world and I talk to various groups, I'm like, please, please, please give universities feed. Back because that is actually the best way to change the whole pipeline. Right. Because we, if you'll get the right when they're junior, then it's not such a big deal. But that also means that companies have to have continuous reskilling. No company likes that. No company wants to do that. They hate it. They just want to have one and done, and that's it. But companies need to get it in their heads that they need an AI training unit and it is permanent because the capabilities are going to morph and change. And maybe generative AI gets lesser depending on what field you're in. But then computer vision might get, you know, have big changes. It's not going away. But if you don't get it through your head that it's always going to be needed, then I think, you know, it's going to be a I think we will start to see the companies who actually understand the need for reskilling, continuous reskilling, they're the companies that are going to rise to the top.

Chitra 23:00

I actually really like that approach because we were talking about how much things are changing and um and how like I really think the pressures on these the CIOs of the world, right? To say, like, we can't think about this in the classic IT refresh cycle. Like we don't have the luxury to say, let me think I'm I'm making a d a procurement decision for my my large enterprise, and um, this is what the run rules are, and two years from now we'll decide if we move to something different. Like everything's changing on a month-to-month basis. And reskilling in this environment, I don't think, is an annual exercise, right? It's it probably does need to be something that's continuous, that is really tied to this other care and feeding and probably closely aligned with how the CIOs buy, because they have to be, they have to be willing and interested to experiment and push their workforce and teach their workforce and make sure that they are actually like, you know, upgrading their acumen as the new things are coming out so that you can figure out what to do and what not to do. And I do think there's a gap, but it's kind of it's it's an interesting way to approach it where we have to maybe in our organizations think about that AI reskilling as like a prominent piece of it. Permanent piece of it. A permanent piece of it for sure.

unknown 24:13

Yeah.

Yevgeniy 24:14

Following up on that thread, I mean how much of this is going to be addressed through formal training programs, like some kind of you know, uh uh a boot camp or some kind of a you know certificate course or something along those lines? And how much of that is just gonna be experience driven? Because you know, we're still obviously this technology is in some ways still brand new. And it to certain organizations it's just now being introduced at scale and made available to all employees and things like that. How much of this is gonna get sort of resolved over time just as these individuals start interacting with these systems and experiencing these limitations, maybe making some mistakes? What will be the balance?

SPEAKER_02 24:59

You know, the answer I'm about to give you is so unsatisfactory. It's going to be both. You know, so I mean it really depends on the company and how many employees you have and what their background is. But for a typical government organization, for example, especially if no one really knows AI at all, then you need kind of intense boot camps. But then think about it like embedded journalists, you know, how like we have embedded journalists during the war. I mean, you you need to have trainers who are kind of embedded with different parts of the organization, especially as maybe a new AI capability rolls out, then one team gets it. Let's say one team gets a whole new computer vision system to help them analyze imagery. That's uh my daughter worked for the National Geospatial Institute. So, you know, that's a lot of what goes on in places like that. And you can imagine they get new tools. Maybe sometimes the tools work well in uncluttered environments and they don't work well. But, you know, maybe there are different approaches that you can take to help people improve. Um, but you're not gonna know that until you have an embedded trainer with them to truly understand how people are using it. Look, people are always gonna use a technology in a very naturalistic way. Not all the technologies that are coming out have great human computer interfaces. Some are very wonky, some have great human interfaces. So you don't have time to retrain everyone on, okay, you do this button and then do this button because people have different workflows. But if you have embedded trainers who can understand how one group is using it may be different than another group, then that's gonna actually be much more helpful. One of the things that we really need to get away from are how Bill in the cubicle next to me, he knows his workflow. And then he sees me struggling with it and he teaches me his workflow, which is actually completely wrong. Um, but now, now I take Bill's mistakes, I propagate the mistakes, and then there is a lot of this like, oh, let me sh, I know how to do this. I learned, I just taught myself how to do this, right?

Yevgeniy 27:13

Well, the folk knowledge, right?

SPEAKER_02 27:14

That's exactly right. I mean, and and look, this is nothing new. This is not new to AI. When I was a F-18 pilot, we all learned very quickly from each other how to do this thing called an internal controlled approach where we could actually set up the bombing system to help us do landings on runways, and we were not supposed to be used using this. You know, all the disclaimers do not use these systems to do this task, and yet we did it anyway. And it would be okay most of the time. But if you tried to use this on a carrier on a pitching deck, I mean, you'd kill yourself. But I mean, that's that's not what folk knowledge told you, right? So you do a company needs to be very aware of how tools like this that can be adopted in a million different ways, even across different workflows, even on the same workflow, you need people who are embedded in those units to kind of see what's happening and help try to redirect people to more, especially in safety critical environments, to more safe and predictable outcomes.

Chitra 28:19

Yeah. So some of this that you're describing is, you know, familiarity and training and kind of understanding how to um how to do things the right way. And some of what you're describing is really like hygiene development, right? So maybe it's really when we think about the workforce, we have to figure out like how we temper, like, yeah, when we're under pressure, you know, do it and and and stick the landing and make your make make it um work for what you need it to work for, but understand where where it might fail and understand where those risks are and like and have the hygiene to know like it, hey, I'm I my gut tells me I shouldn't do this and in this circumstance I need to get some extra eyes on it or do this other thing to to double check my homework and get through that.

SPEAKER_02 29:05

I think that's I also think it means a lot of workload on the managers because then they need a special skill set to recognize when AI products may or may not be achieving what their objectives are. And so I do think that the hardest part of all of this are going to be on the people who are supervising both the people and have to make sure that the products are coming out in a way that are correct and that you don't have to send back over and over and over again to get done correctly. But that is why the middle managers need to work very closely with these trainers, because a middle manager, I mean, that would be all you did uh if you didn't have some help in that area.

Yevgeniy 29:55

Aaron Powell So it sounds like to get to mission impact, to take these technologies sort of and and drive organizational excellence with them, these these trainers or maybe some kind of AI champions need to be there, gathering all the best practices across the organization, understanding the limitations of the technology, and then developing some kind of shaping materials and guidelines to help managers and specific teams really get the most out of the tech while you know while avoiding the risks. It sounds like that's where the direction is kind of going in terms of actionable steps that the organizations can take.

SPEAKER_02 30:33

So I'm gonna fight with you again, a little bit more fighting. I would I think all of that was correct, except for the phrase AI champion. I would love people to take a step back because that that is a very positively loaded phrase. And I would change it to AI realist, which is to say sometimes you want to be a champion, and sometimes you need to recalibrate people's trust, which is to say, don't use it in this way, stop doing it for this tool. So uh, you know, an AI shepherd, maybe. Um, but but but we need to make sure that people understand, look, this is all about calibrating expectations, particularly as you go up the chain of command. AI is the right tool sometimes. And depending on the type of AI, we need to make that clear. Sometimes you need to take a step back and think about Occam's razor and think about, okay, well, maybe I don't need to use a nuclear bomb to put this little nail in. Maybe just a basic old tool in whatever way we use more. Maybe that's okay. Uh, but we just need to make sure that people are very clear on it's good. We call this function allocation. Sometimes AI, particular form of AI, is good in this function, but sometimes it is not good in this other function.

Chitra 32:02

It's funny, the the thing that came to my head when uh when you were talking was like uh being a pull pill pusher. And it's like we definitely don't want to be so uh biased that we forget that there might be a better solution and you don't need to take that pill to solve that problem, right? And it's not the the magic that you know heals us all. So so that got me thinking, like, okay, well then how do you if I if I was to like really like push on this like pharmaceutical kind of analogy, how do we like up our acumen and understanding of like when do I need to go down that path and and when does the uh the holistic approach or the the kind of old school AI approach work? And um it makes me think about the the AI action plan has kind of emphasizes some of these like technical schools and and partnerships with these more like formalized programs and practical training. Like, is that the answer to all this stuff? Like, is it more of that and certifications and working through um the more technical talent pipeline, or is it something else related to the folks that are not necessarily in that STEM technical realm that we have to focus on because it's that wider mass that needs to really like up the anteon the acumen and understand the risks and not just pop the pill, right?

SPEAKER_02 33:19

Again, another unsatisfactory answer, all of them, right? I mean, we have to do everything from really start thinking about how we train people in primary and secondary school. What should kids know? What should they not know? When should when it's the right place to introduce those tools? You really do need to make sure that we don't replace critical thinking. That is so important. And we need to start figuring out when is the right timing to make that happen. It is very possible that we start to replace human-led education with AI-led education. You know, I think there's a successful argument that if you're living in a third world nation and you don't have any other resources, maybe that's okay. I am kind of concerned that the Ivy League schools become the only place where you're taught by a human. And so and the the reality is human-led education is still going to be supreme, the best quality. How you use AI to augment that is still an open question. But I think that there is a stampede to have more AI-led education. That's we could that's like a whole nother episode. We should send other people here to talk about that. But I think you're gonna need a lot of that in primary, secondary, college education, like I talked about previously. But then you have, you know, and what we do at George Mason is we provide certificates. This is a mini master's at the graduate degree level. Okay, not everybody can do that. All right. So then we come out, organizations pay us, we come out and try to train the trainers coming thing. Okay, that's only gonna take you so far. Uh, and it's really gonna depending on the size of the company. Like, look, small companies are not gonna be able to have these embedded AI trainers in every aspect, right? So they're gonna have to figure out what works for them, some training here. Uh, maybe we take everybody to a conference there to kind of get them refreshed. Um, bigger companies are gonna have more resources. There is not going to be a one size fits all. You know, I can't give you, although as I say it, it sounds really good, like the Dr. Missy's uh plan for how you should have AI education. And then I could be incredibly rich for doing that. But that's not possible because as soon as I we stop this podcast, I'm gonna pick up my phone and there's gonna be 20 new articles that said there's all those new things in AI, and you just can't keep track of it. That's actually why you have to have, even for small companies, at least one person who is sitting on top of that so that they understand what's really coming down the pike. How does my company maybe need to leverage this or not leverage this? And how can I communicate that across the company to skill sets that all people understand?

Chitra 36:16

I mean, how how does the national security and just even the broader like federal um government apparatus think about it? We're so used to like the the analog to all this is LCATs and certifications and these formal things that you can check a box to to say, like, hey, I know you know a thing because I see this particular set of words written down, right? And what you're describing seems like it's not necessarily like anything like that, right? So how do how do we how do we not create a check the box and figure out how do we really like, you know, create the advantage and and give the necessary training and pipeline development and reskilling that's needed.

SPEAKER_02 36:58

I think the realistic answer to that is that you're gonna have to partition who your people are by the levels of safety criticality that their efforts touch. So, for example, if we're having people use AI and weapons, those people should have to have a necessarily provable higher level of skill than somebody who is working on document preparation, maybe requirements generation, where lots of, I mean, it's just gonna be a different level of skill that you're going to need. I hesitate to say that the government should start thinking about new kinds of certification in this space because the government is always at the bottom of the totem pole in terms of uh getting the time and effort to help develop, you know, they're they're just lagging behind the companies who are leaping ahead in AI. I do think that, and this is actually becoming big in a bipartisan way across across Congress, is we need to start thinking what does safety critical mean? What is a level of consequence? And I teach AI and hazard analyses, which are super important. Like, you know, if I'm using AI to go to shop on Amazon, is that such a big deal? Probably not. But if I'm using AI to pick targets, uh, it's different. And we need to make sure that, and and this is actually one thing that I see that we deb this government, I work with the United Nations, all governments need to do is to step back and say, you know what? This one, LLM, good for helping me generate systems engineering documents, probably not so good for generating a target list. And, you know, we need to start instead of maybe classifying the person, starting to classify the task and what kind of skill levels that we want for a particular task to be done.

Chitra 38:58

Yeah. I think that like threads in nicely even with what we were talking about in a previous episode on like the the testing philosophy with uh Mark Munsell. I mean, I think um I think you're spot on, and we m we might have to rewrite the way we were approaching this like across the board because this is a paradigm shift. And I think the the old models on not just workforce training, but then you're you're you are giving me the thought on like, yeah, classification as well. They're all hand in hand, right? Because it's the people-centric model that we used to have that is now a little bit flipped on its head.

Yevgeniy 39:32

Yeah. And it seems like, you know, we keep coming back in this episode, it looks like, to some of the themes from the previous episode in understanding the performance of these systems and that really informing how you team them up with humans to perform specific tasks and maybe not just globally thinking that we're gonna do a single thing for all tasks. Um, wrapping up, you know, as we look at this AI action plan, there's there's definitely a push to to increase the adoption of these systems because I think in certain use cases they're definitely you know in enablers of uh and I'm thinking, you know, prototyping, code generation, things like that. Maybe there are good applications that we can all agree are gonna benefit from that, like you mentioned, summarization. Um but what's the one thing that leaders in these you know different national security agencies and organizations have to keep in mind as they think about the human element of this transformation? So what is the thing that I you want to leave folks with as they you know conclude this episode?

SPEAKER_02 40:38

AI is a tool in a toolbox. And if you just like a pilot, like I was a carrier pilot. An Air Force pilot can't do what a carrier pilot does. I would like to point out though that all Navy pilots can do what Air Force pilots do. I just want to say that on the side. But like fighting words. Uh so you would never ha have somebody who had one kind of capabilities, skills, and capabilities. It just doesn't always work. Even Navy fighter pilots don't necessarily make good surgeons, for example. They have to go to medical school to get a new set of skills. I hope so. Right, exactly. So I that's what I think is like when we start to think about these models, you've gotta make sure you know what the right capabilities are, the form and function. Is this really what I need for this function? Um, and do I have the humans, do they have the right background to really fully understand that?

Chitra 41:42

No, I like that because um instead of like the one size, I mean, obviously it's not one size fits all for AI development and usage, so we shouldn't take that same approach on training either in workforce readiness, right? The mission requires us to be able to adapt our training and reskilling based on what you're doing and why you need to do it.

Yevgeniy 42:05

Absolutely. Missy, I want to thank you for joining us and breaking this down as we think about how to prepare our workforce to work alongside AI in these mission environments.

Chitra 42:16

And thank you to our listeners for turning into In says all source podcast series from Action Plan to Mission Impact.

Yevgeniy 42:24

Our next episode is Allies, Adversaries, and AI. And we'll be returning our focus to the global stage, examining how the AI Action Plan shapes international collaboration, competition, and security.

Chitra 42:37

And joining us for that conversation will be uh Chip Usher, the Senior Director for Intelligence at the Special Competitive Studies Project.

Yevgeniy 42:44

We'll dig into AI diplomacy, trusted partnerships, supply chains, exploring how the U.S. and its allies can compare, compete, and lead the rapidly shifting global landscape.

Chitra 42:56

And until then, I'm Chitra Svanandham.

Yevgeniy 42:58

And I'm Yevgeny Sarotin.

Chitra 43:00

And thanks for listening.