Global Affairs Expert Webinar: Trust Infrastructure and Artificial Intelligence

Vinh X. Nguyen, senior fellow for artificial intelligence at the Council on Foreign Relations, leads a conversation on trust infrastructure and artificial intelligence.

These webinars provide an opportunity for college and university educators and students to discuss global issues with CFR fellows, Foreign Affairs authors, and other leading experts. To register for future invitations, please complete this form or email [email protected] with your name, title, and academic affiliation. 

Speaker
Vinh X. Nguyen
Senior Fellow for Artificial Intelligence
Council on Foreign Relations

Presider
Irina A. Faskianos
Vice President, National Program and Outreach
Council on Foreign Relations

Transcript

FASKIANOS: Welcome to today’s session of the Fall 2026 Global Affairs Expert Webinar series. I’m Irina Faskianos, vice president of the National Program and Outreach here at CFR. This webinar series provides a forum for students and educators to interact with CFR experts and other thought leaders and participate in foreign policy discussions. Thank you for joining us.

Today’s discussion is on the record, and the video and transcript will be available on CFR.org/education if you would like to share the materials with your colleagues or classmates. And, as always, CFR takes no institutional positions on matters of policy.

We’re delighted to have Vinh Nguyen with us to discuss trust infrastructure and artificial intelligence. Mr. Nguyen is a senior fellow for artificial intelligence at CFR. At seventeen, he was recruited into the National Security Agency’s elite Stokes program. He went on to become the youngest employee in the agency’s history to be promoted to the senior executive ranks, and later served as the NSA’s first chief responsible AI officer and as chief data scientist for operations. He has also served in the National Intelligence Council as the intelligence community’s most senior cyber analyst advising the Director of National Intelligence on cyber threats and their geopolitical implications, and he’s served under four presidents and is the recipient of the intelligence community’s highest honor, the National Intelligence Distinguished Service Medal. He has a number of affiliations in addition to CFR—including senior technical advisor to Anthropic.

So, Vinh, thanks very much for being with us. There’s nobody that we would want to have on more to have this conversation about trust infrastructure. So maybe you could begin by telling us what this means in the context of AI and how the challenge of building and maintaining that trust has evolved as AI systems, as we have seen, have become more powerful and more deeply integrated into every part of society.

NGUYEN: Yeah, thank you, Irina, and thank you for joining us today. I’m really excited to see and hear your questions.

You know, as someone who grew up in the 1990s and building the digital infrastructure that we have today, I learned many, many, many lessons that we can build a safe, secure, and trustworthy AI future. This is—you know, yes, it’s unprecedented. It’s new. It’s exciting. It is quite uncertain. It’s—you know, we are very anxious about the current environment, especially with the development of AI. But I want to give you a few key lessons that I learned, you know, as we built the infrastructure.

Number one was that we actually failed to learn that when we are building these digital infrastructure and AI, like, when we are not thinking about security and safety and building the trust and the confidence around it, we will have to—we will need it and we have to bolt it on later, and bolting it on later does not work. And that’s what we ended up with a lot of the cybersecurity challenges that we have today in our critical infrastructure—you know, the fraud campaigns that you have to deal with and all of the challenges that we face today. So lesson one is do not bolt it on; build it from the beginning, right?

The second is that we learned how—like, we learned how to manage a system that is dynamic and AI is quite dynamic, right? It’s probabilistic, it’s opaque, and we thought, oh, these are, like, new stuff. Yes, they are new when combined together, but we learned that we can deploy these systems. When it’s probabilistic, we learned that we have to monitor. We have to really understand and monitor. We have to measure, and we have to assume that there will be incidents and we will have to respond and recover from setback, and so we have to be resilient and think through on how to do that.

And so the lessons from the 1990s and the digital infrastructure buildout are useful, but the geopolitics today, the competition—the global environment today is quite challenging. The technology is probabilistic, like I said before. It’s opaque. It can do many things autonomously.

At the same time, we have to, you know, figure out how to navigate a very competitive environment in a global order, and if the country believes that AI is going to increase and enhance our prosperity and our security, then we really have to figure out how to be competitive in the realm of AI, and they have to do it in a way that is trustworthy because we have to sell these products, right?

We have to use the products, and no one will trust us if we just tell them, like, you know, just use it, right? So I think—I think the global environment will demand both a—capabilities, increase enhancement, something cheaper and more accessible, but at the same time, we have to speed up the security and safety along with it. It doesn’t have to, you know, minimize safety and security to make gains in AI.

And that leads me to thinking much more long term about what I call trust infrastructure. These are the little things that we build over time that we never considered.

I’ll give you a very interesting example in the early 2000s. If you—if you recall or may not recall, for a long time we did not dare to put our credit card numbers onto the internet. Like, everyone was warned do not do it. Hackers will steal it and steal your credit card. And no one—so how can commerce—e-commerce operate?

Well, it operates because the browsers and the companies came together and said, hey, we need to build certificate authorities and we need to make sure that our—there’s encryption when we are purchasing products.

And so you have these protocols and new ways called SSL (Secure Sockets Layer) and, you know, it’s teaching us, you know, if there’s a key you can put your credit cards in. If there’s no key, don’t put it in.

But today, if you use Chrome or Safari or all these browsers the default is encryption, right? It’s encrypted. So you actually don’t have to think about it. Those are little pieces—technical pieces—that we learned on how to build the technical infrastructure.

We can do the same today for us. But there are structural trust infrastructure, things like incident response, right? When things fail, companies are unwilling and less likely to publicly go and say, yes, our products failed.

But you—you know, like aviation, we need to make sure that you protect the people who pull, like, the flag and say, hey, there’s, like, something going on here and I want to report it, and report it in a way that there’s no consequences, right?

So it’s not going to be blowback against your reporting of failure, and that’s what we have in aviation today, right, and we have these safety systems built to do so.

And so we need a way to make sure that we enhance our incident reporting structurally so we can learn from our mistakes. It’s a new technology. We have—we have not mastered the technologies. We have to figure out how to do the technical infrastructure, you know, like the little SSL, right, or making sure that there are traces when we are rolling out agentic AI so we know who is accountable and who is behind the activities.

Otherwise, you know, we don’t have accountability, right? So there are little things that seem to be so minor but they are consequential for us to build a long-term trust infrastructure in the long run.

So, you know, at the core, the basic technical things that we can do in terms of visibility, in terms of accountability, in terms of privacy and identities at the structural level. There’s a lot of things that we can do to make this a visible infrastructure, right?

Trust is, like, very opaque, you know, ambiguous things, but actually in security trust is really building the infrastructure of confidence for us to actually trust in the system, trust in the people building the system, trust in the people who are deploying the system, and trust in the science of the systems.

And right now we don’t have any of that, so we have to build all the ecosystem and infrastructure at the core, and you all have a chance to actually voice, demand, articulate, help build this new infrastructure, architectural, right, structural, as well as technical into the future for AI.

And if we can do that, we will be able to increase the confidence in the development and the deployments of AI, and then we’ll learn that we can harness the technologies and manage and mitigate the risk that are manifesting across these systems today.

And so I just want to give you, like, a quick run but I really—the point here is to have a conversation with you, to hear your questions, and to hear your hopes and fears. I don’t have all the answers, but I’m here for you.

FASKIANOS: Wonderful. Thank you, Vinh. That was great.

Let’s go to all of you for your questions. Please click the “raise hand” icon on your screen to ask a question on an iPad or tablet. Click the more button to access the raise hand feature, and when you’re called upon please accept the unmute prompt and state your name and affiliation. Please be concise. And we will also take written questions in the Q&A icon and if you can write your question—include your affiliation in when you’re writing your question that would be great.

So I’m going to go first to Susan Aaronson. And if you can unmute yourself. There you go.

Q: Thank you. I’m Susan Aaronson from GW Elliott School and I’m also co-PI of the NSF-NIST Trustworthy AI Institute where we do a lot of work on trust.

And two things that we found—and I just wonder what you think about this—is the United States recently has stopped using the word trust or trustworthy. They use the word trusted. I have no idea what that means. And we found that meaning—gist—is really important to public understanding of AI. How can we communicate to the current U.S. government the importance of trustworthiness and trust? Like, trustworthiness is a perception versus trust as a form of social capital. I’m really struggling and would love to hear your thoughts.

NGUYEN: Susan, thank you for your question, and I am a big fan of your colleague Zoe as well. So thank you for asking. (Laughter.) No, this is not planned. Susan just popped up.

So the way I think about it, you know, is for this government they care about national security, economic competition. So you have to talk about, like, hard power, you know, and for a long time—I’ll give you a really interesting example.

In the military when you ask them, like, you know, General, do you want your weapons to be assured so that it’s not—you know, secure and assured so that the enemies cannot, like, take over the weapons and misuse it? And they’re, like, yes, absolutely.

Do you want the confidence and trust in the system? Yes, right?

And then if you say, well, you know, do you want AI security and AI safety in your system, they’re, like, I’m not sure what that is, right?

So you have to use the right words here, and I—the way I think about it is that trust can be very ambiguous and opaque, and it’s hard to really understand. But what it is, is really building the structure and the infrastructure to build a confidence in the system that you have, which is trust, right?

And so the way that I talk about it with, you know, the government is really about assurances—confidence and assurances. Do you want to assure that, you know, the weapon systems or the national security systems or this piece of software behave reliably?

Do you want it to be secure so it’s—you know, so it’s not attacked by an outsider or an insider? Do you want it to be safe so that it doesn’t create harm and hazards? Do you want—right? And then they will say yes, yes, yes, yes.

And so it’s really about the type of things that we care. So I think we should talk about, like—in specific terms like reliability, security, safety, privacy or confidentiality. So converted, right? Confidentiality, and, like, basically build the—ensure an infrastructure, independence, evaluation, incident response to ensure that we have those characteristics and that we trust and we have the confidence in them.

And so I think if we talk in those terms, national security, governments, they do understand these terms and so I highly recommend, right? We don’t have to be stickler for, like, a word. If the environment is changing, we really have to communicate with the people within the context that they operate in.

So, hopefully, that’s helpful, Susan.

FASKIANOS: Great. Thank you.

I’m going to go next to a written question from Heidi Currin, who’s a graduate student at The Citadel: Could you talk a little bit about your experience working at the NSA and the kind of work you did with AI and national security?

NGUYEN: Yeah. Heidi, thank you. I think you should join NSA and the intelligence community. They’re great people.

I think, you know, when, like—when people think about these agencies, they’re like, oh, they’re, like, faceless bureaucracy.

No, they’re actually run by people like you and me, and they are kind—you know, very kind people and they’re patriotic and they want to do the right things, and so I highly recommend to do so.

I would say at NSA the goal is—at least for NSA is to ensure that we prevent, you know, intelligence surprises by providing the best information for our decision makers and our policymakers, and at the same time, help to protect and secure our national security systems.

These are classified systems—you know, the weapon systems. They could be ranging from nuclear command control to the infrastructure that we rely on for war fighting. And so the work is very focused on the mission and you get to—you know, basically, they assign you problems and you work on the problems with teams, super technical, and you do what you can, right? You collaborate with the private sector, with academics, with everyone who can help solve the problem.

The bulk of the problems are, unfortunately, classified, and so it’s fun. And, you know, like, if you’re watching James Bond—it may not be the same as James Bond, but it’s actually very close to it, right? Like, you go and you’re, like, oh, you know, like, I’m actually operating and analyzing an adversary’s communication to prevent, you know, planning for a terrorist attack, to cartels, to, you know, building missiles that can harm us, or building space weapons and cyber weapons, right?

So there’s a variety of things—you know, learning about the intent of leaders. So it is really ranging from, you know, using your linguistic capabilities, your foreign affairs capabilities, your regional domain expertise, your technical expertise.

So it’s—you know, it’s for—really for a lot of people, and the key is teamwork and bringing the different skillsets together to solve the problem. But the problems are usually really fun, and if you do it for, like, twenty-plus years like I do, it’s, like, after a while a bit of a routine.

But when you actually break out of it, you’re, like, oh my gosh, you know, I was—you know, I get authorized to do such a thing that would be considered a—you know, a criminal act in a foreign country.

So I think it’s a really fun place and it’s a very close community. So, hopefully, you get to join, and happy to follow up.

FASKIANOS: Thank you, Vinh.

Just a question. We’ve seen a lot of cuts to the—to different agencies in DC—federal agencies. Has the NSA had cuts as well or are they operating at the same level?

NGUYEN: Yeah. I think, you know, based on last year, I think everyone had to go through the DOGE cuts and so, but the agencies are rehiring. I think right now, I mean, depending on the agency. It depends on the workforce planning and where they’re aiming for.

So I cannot speak to every single individual agency because I think they all have individual plans and timeline to do so. But I highly recommend, you know, you to just check the NSA.gov or the intelligence.gov to see availabilities. I haven’t checked lately, but there should be availabilities.

But they do work on really wonderful problems if—you know, and it’s truly public service.

FASKIANOS: Wonderful.

Let’s go to a raised hand from Mojúbàolú Olúfunké Okome.

Q: Yeah. Thank you very much. I’m Mojúbàolú Olúfunké Okome and I teach political science at Brooklyn College.

So I have three issues. One is trust, the other is participation, and then international cooperation. So trust—how do you build trust when there’s so much misinformation and disinformation and a lot of fear about what AI is going to do to us?

Then there’s also participation. You encouraged all of us to join. I mean, so I’m not in intelligence and I know a majority of human beings are not. How do ordinary people participate in the governance in a proactive way?

Then international cooperation. This is such a huge and disruptive technology, and it has many facets. It’s impossible to manage without substantive international cooperation and, you know, the lack of trust is not just a national issue. It is international. How do we build international cooperation for effective governance?

NGUYEN: Yeah. Professor Okome, thank you for the question.

I would say that trust is really built between people, individuals to individuals. What I learned through this change is that, yes, there are interesting AI capabilities, but the speed of progress is really based upon the trust that we build between us, between people, and trust is broken everywhere.

Institutional trust is quite low and trust in America is low as well, right? And so my take is that we have to—you know, we cannot just be, you know, all dismayed and be pessimistic. I’m usually very optimistic that when we are put in this situation it’s up to us individually to do so.

And I do agree that today the international voices and the collaboration and the access to Frontier advanced AI is not where it should be. I know that the UN just announced the upcoming—the next year’s global dialogue on AI, and my colleague Kat Duffy and Tony Oweke are very, very involved in that and we really want to bring the voices, you know, to the table.

But geopolitically it is challenging. I think we—I think we all individually should demand more from our governments and corporations on what is needed, because if we don’t voice that, you know—and we should continue to voice it, right?

It could—sounds helpless, but I think we should—we should continue to do so and find opportunities where we can bring our colleagues, you know, across the Global South in areas that are underserved to join us, to demand more of the AI companies, of the hyperscalers, of, you know, of corporations who are building this.

But we have to be very specific on what we want, right, because sometimes it’s so noisy we are not clear on what we exactly want and I think it dilutes our power. So wherever you can—we can find ways to say, hey, for—like, for example, if you were to ask me how do we trust in the international community for Frontier AI, I would say, one, make sure it’s accessible across multiple languages and not just Chinese, English, maybe European languages and maybe some other languages, right?

Like, can it be much more? Because I think the capacity for the linguistic—I mean, it’s a language model and if it is not accessible for a larger group of people it will be—it will be a missed opportunity.

The second one is ensuring that when we’re deploying the system there will be opportunities and pathways for recourse and accountability, because lacking any accountability in the use of AI or the decision made on you by AI and that you don’t have any accountability, you cannot pinpoint back—if you don’t have a way to have due process, then I think we—it’s a missed opportunity. Actually, it’s a huge disadvantage when we basically forgot who we are as people, right?

These are American and Western democratic values, but we have to show the world that we can—we can do that, and lacking any of that I find it really hard for engagement.

And lastly, I think confidentiality and privacy is really key. Given that Frontier advanced AI are using large amounts of data and eventually personal data, things that you put into the system that you may or may not consented for the companies to use will take away, you know, your data.

And so, I think we need to really emphasize on confidentiality and privacy and respect, you know, our rights and respect who we are and provide a level of agency for all of us. And so I think building this will require all of us to do so.

To me, I don’t really care what other people or governments have to say. I know my values, I know what I have to do, and I know that I have to build trust one person at a time. But that is really better than nothing.

And so it’s not—it’s going to be challenging, but I think we have to start one person at a time and understand that we have to include and bring everyone along to really harness and benefit from this technology.

So, hopefully, that’s—that can—that helps along the path. But thank you for your question.

FASKIANOS: Thank you, Vinh.

We’re going to go next to a written question from students at Southern Maine Community College who are wondering about the Hugging Face incident: If AI can go beyond the bounds set forth by its programmers, how can we build this trust environment?

NGUYEN: Yeah.

FASKIANOS: So maybe you could just recap briefly the Hugging Face incident for anybody who doesn’t know what happened and then go from there.

NGUYEN: Yeah. Yes.

For those who are not tracking the OpenAI Hugging Face incident, what happened was that OpenAI was doing a testing evaluation of its new model, of a set of models, and the AI system were limited in what it can do. But it found a way to—you know, basically, it’s designed to complete a certain task and meet a certain benchmark and the AI couldn’t do it and so wonder where else it can actually cheat on that benchmark.

And then it hypothesized a Hugging Face incident—(laughs)—for it to cheat. And so it, you know, broke out of its sandbox, and broke some vulnerabilities, and found ways to, like, get out, and was able to reach into Hugging Face, which is a company that is providing models—open-weights models and leaderboards for people to download as, you know, helping the open-source community to use AI. And the OpenAI agents were able to break in and be able to find the data that it needed and open. And then Hugging Face found out about it and OpenAI found out about it at the same time and were able to figure out that, oh, my good lord, like, the sandboxes that we built to contain the AI had failed.

And so how can you trust a system that doesn’t follow your objectives, right, or follow the bounds that you are setting?

Great question. Right now, we just don’t have the right—we are just not progressive enough in terms of our scientific understanding in what the few would be calling AI alignment or interpretability of AI for us to make progress.

And so what we can do is advance the scientific understanding of alignments and the engineering on how to align the behaviors of the AI—you know, make it learn so that it doesn’t, like, do things like this. But you can see that’s very hard, right?

It’s like you have a six-year-old child with, like, a lot of abilities and a lot of knowledge and you basically let the six-year-old go right into the playground and just say, do things. And then it left the playground and you didn’t know what happened, and that’s what really happened.

But right now it’s not like the one child. You just unleashed 10,000 six-year-olds right into the playground. That’s where we really are. You know, six-year-old with, you know, great genius—(laughs)—who are geniuses, right?

And so that is really, really difficult. But I think I am hopeful that we can advance the sciences of alignment. We can advance the sciences and engineering for security sandboxes, meaning that, you know, contain—(off mic)—

FASKIANOS: Vinh, did you—Vinh, did you just hit—OK. Vinh? OK.

NGUYEN: —(off mic)—action, but you can do low consequential actions, right, and taking the low consequential task and you’ll be OK with it. But I think instead of, like, say, trust or not trust it’s more like, OK, there are actions and tasks that we should not trust AI to do and there’s tasks that we are OK with.

That’s your personal choice and enterprises’ choices as well, right? And there is appetite. But I think—I don’t think you have to trust it or not trust it. You can still—you know, you can still have, you know, doubt about the system but allow it to actually complete task that you are OK if things go wrong.

So it doesn’t have to be, like, not use or use it.

FASKIANOS: Thank you.

Going next to Greg Treverton. Greg, if you can unmute yourself. There we go.

Q: OK, great. Thank you. It’s great to listen to you, Vinh. I had the good fortune of—Greg Treverton, University of Southern California. I also had the great good fortune to work with Vinh, so it’s a pleasure to see him in a new context.

My question is a little different. You’re a—intelligence community, as you’ve said, has taken its shots in the last several years including us personally, right? But I wonder, as you look across the community, what worries you the most, particularly in terms of how difficult or how long it will take to fix some of the damage that’s been done?

NGUYEN: Yeah. Greg, it’s great to hear you.

Q: Likewise.

NGUYEN: Yes, and—yes, I think there are damages, right? And my take is that it will take a generation to rebuild the trust of the workforce and the national security workforce—you know, to build and regain the trust between policymakers and—and, you know, the analysts and the engineers on the ground.

Trust is easily broken. It is very hard to build, and—but I believe that for our country to be prosperous and secure in the long run we have to start somewhere, and right now I think the environment is challenging for sure when intelligence, analytic result, and assessment may be challenged, may be questioned, even if you are using the best tradecraft, the best ways that you can produce.

I think this will pass. We will learn from the experience, and my hope is that the community can learn and the policymaking community can learn how to value each other, how to show respect and be respectful, see the value what the intelligence community can bring, and hopeful that the people will recommit, you know, to public service, because if people are not doing it no one will.

And so, unfortunately, we are where we are, but I’m really hopeful that we can rebuild trust, and we just have to open our hearts and start somewhere at this point. But yes, our trust is broken. It’s broken everywhere, like I mentioned. Now we just have to rebuild the trust between the policymaking community and the intelligence community.

FASKIANOS: Thank you.

I’m going to go next to Eric Schmidt, who is with his students at Kentucky Wesleyan College: You seem to argue that trust infrastructure could be an American strategic advantage. However, China already has an extensive digital architecture of tools which are particularly useful for centralized domestic deployment, whereas the American trust infrastructure is relatively fragmented and subject to democratic accountability. Who actually has the advantage? Is there a risk that China establishes a dominant operational architecture before we do?

NGUYEN: Yeah, that’s really a great question. I would argue that it’s—like, for us as the United States and the West, we learned that we don’t build trust because we told someone to trust us, and that’s what any authoritarian government will say, right, is that we set a standard and the government set it and, therefore, you know, this is trustworthy.

The way that I think about trust in AI is very similar to two sectors. One is the financial sector and the other one is aviation.

In the financial sector, if you think about the early days of financial assets, people may not trust the municipal bond from Oklahoma in 1900 because you don’t know what happened, right? And so eventually you have different institutions, different companies, different teams popping up, right? Be it Moody’s and, you know, other shops popping up to rate and—evaluate and rate these assets.

To build a trust ecosystem, you actually have to have independent third-party evaluation of the thing, right? So if you are building—so today we have, say, electrical products, right—like any device electrical.

If you actually flip over your device, any device, you will see a shield from United—sorry—from Underwriter Labs, UL. And if there’s a—you know, there’s a UL shield, you should be assured and confident that it doesn’t, like, blow up in your face when this gets heated, right?

And then you look at some of the local news, you know, people bought products and, like, the products, like, exploded. Usually, they have not met the UL standards.

And so UL is not a government facility, right? Moody’s is not a government entity. We learned as a people, as a democracy, that we trust when there is independent evaluation and benchmarking of the thing, and in this case, it’s independent evaluation benchmarking of the AI capabilities.

Now, independence for evaluators today is quite limited because funding and access still are tied to the companies. And so for us as a nation, as an ecosystem, we really have to make sure that we really invest and ensure that our evaluation benchmarking ecosystem is independent, because if it’s independent from money, from—you know, from pressure, from access, then people will at least trust it because you have a third party to vouch for that.

So that’s, like, one way that we will be different from how China would run because the Chinese AI labs or the Chinese government will force it through, if it wants to, certify through some government entity, right? So third-party evaluation of a Chinese product is still, you know—I mean, do you trust a Shenzhen independent evaluator of AI or any product?

Maybe, but maybe not, because you know that Shenzhen company may be under undue pressure from the Chinese government. We just don’t know, right? But the trust is quite limited.

The other one is actually incident response and incident reporting. It’s actually very easy to ignore or to, like, not report failures publicly. If we add incident response and reporting into the mix and people can publicly report on incidents when things fail, then people—it’s more trustworthy.

So it’s really our collective efforts that ensure democratic trust. It is not just like some government entity blasted with, like, a logo and suddenly is trusted. May work in 1950 or 1960s. Probably not—doesn’t work today.

And so, I think you can—you can be competitive, and I think that is our strategic inherent advantage if we do it right. If we don’t do it right and have the government, you know, bless this AI is really good, then no one will trust them again. And it seems like the Chinese model, which we are not competitive to be an authoritarian regime. The Chinese are very good at that.

So that’s how I think about it. Over.

FASKIANOS: Thank you.

I’m going to go next to Joseph Bower, whose raised hand. There you go.

Q: Thank you very much for your comment.

Today, we got in our emails here at Harvard a note coming out of the Berkman Center on the issues that you’ve been talking about, and they use the analogy a thousand years ago in Europe, England, wherever, dogs were used as tools, not—there were some domesticated dogs. And if we think of going from sort of Pomeranians all the way to pit bulls, a key thing was they actually used foxes but they killed all the wolves because the wolves could not be trained to be trusted.

And then they—in their reasoning, they then come and give an example in New York City that had a website which ended up giving really bad advice, things that were illegal, that would—the chatbot had just developed.

And the—as I understand, the essence of what they’re saying is when there is accountability then we begin to have a way of proceeding.

Now, Underwriter Lab was created by GE (General Electric) to get hold of the wolves that it was competing with and establish standards that would protect companies that cared. But in the end of the day, if there’s no accountability, if there’s no brand—and we’ve just had the U.S. government say, you do it yourselves, how—why should we trust?

We’re seeing breakdowns constantly. If we don’t have the link and between the platform and the damage, then what do we do? Why trust?

NGUYEN: Yeah. Joseph, that’s a really good question on accountability, right, is that when it will come—like when you really think about this, I think we over-index on the Frontier AI companies and capabilities because we talk about it, how great it is at math, coding, and da, da, da.

We are not really debating or talking about how we actually deploy the technologies and how—and where is it appropriate and where it will get you the return on investment and managing your liabilities.

So I think we should really aim at and demand from those who are deploying AI and hold them accountable because then they have to hold the system accountable in some ways, technically, structurally, or whatnot.

And so it’s really up to us, right? And, you know, we—I think we should do it earlier than waiting for court cases because I think the court cases are coming in many environments where AI was used on consequential decision such as hiring, lending, healthcare decisions, and even, you know, dynamic pricing and price fixing. And so those are, like, really damaging us.

Right now, I think the national conversation is so much on capabilities and managing the risk, but I think we should also continue to talk about those who are deploying the technologies, you know, be it a financial service or healthcare service.

They have to follow the various regulations for the financial sector, the healthcare sector, right—critical infrastructure. So we should, like, demand more in terms of, like, accountability, meaning that, one, whatever—you know, they have to be transparent and open, telling us when AI is making decisions that is impacting us right now.

We don’t know, right? Like, you’re putting resumes into, you know, like, the ecosystem and the number-one question is, like, can I bypass the AI, right? I mean, like, maybe we should ask—we should not ask the question, should we—how can we bypass the AI. We should demand the companies not to use AI because we want an openness on how AI is being used to vet us, right, for our hiring. And so I think we should speak up on those.

I think enterprises, for those who may not have been tracking, as part of CFR’s elite AI effort, which is really, you know, kind of aiming at AI deployment as foreign policy—as foreign policy is that the deployment decisions being made by enterprises and by people are actually making foreign policy decisions along the path because the government has relinquished and pushed its decision and authorities onto those who are deploying AI.

So if you are a large financial corporation and you’re deploying financial services across the globe, you are making foreign policy decisions because you are telling people who have access to them, who can use them, who have recourse, how accountable the system is. You are really embedding American values, or not, into the system that you’re deploying.

And so, I think we should work with those to make sure that they understand that they also should be accountable and ensure that American values, you know, manifest through that.

But right now, the ecosystem that we have is that the administration had already, you know, relinquished—you know, pushed self-regulation on the companies and at the same time hold them accountable, right. But they want to regulate them and so you’re in, like, the middle part.

So I think we’re in temporary engagements, right, or temporary arrangements right now between the companies and the administration, and so they are expecting the companies to set the safety standard and meet the safety and security standards or else—basically, or else, or we will send the Department of Justice against you or we will put export control on something, right?

And so there’s a lot of, you know, kind of like instability, but I think where we are right now is do—like, you know, based on the White House accord that the companies signed, that the president, you know, championed, that they should be responsible to self-regulate and we shall watch you to do so, and that’s really where we are, you know, as of, like, you know, a week or two ago.

And so I think we should continue to demand accountability for the use of AI and shift the conversation on how AI is impacting us by the companies, because they have to tell the Frontier AI companies, the hyperscalers, to shore up their security, their safety, their reliability for us to accept that.

I think the people have a voice. We just need to speak up constantly.

FASKIANOS: Vinh, I think you’ve answered the—there was a similar question from Heather Gregg, an assistant professor at Naval Postgraduate School, about how the U.S. government should be regulating or managing AI companies.

And then I—since the U.S. government has pushed it back on the companies, in your view, in an ideal world should this be happening at the federal level? Should the U.S. government say these are the—or is it rightly placed that the companies should be—it should be more—what’s the word? Fragmented? Distilled?

NGUYEN: Yeah. Yeah, I—my take is that the government is in a tough spot because the president, you know, already promises, you know, in terms of less—fewer undue regulations that hamper innovation in companies, right?

And at the same time, it’s a dangerous capability that has national security and economic security implications. So they’re kind of, like, in the middle. And so the way I think about this is that on the surface there is a voluntary commitment and accord between the companies and the government.

But in practice, I think it’s quite mandatory because no one wants to upset or frustrate or run in—or, you know, make the president upset or perceive, you know, wrongly of the company’s doing.

At the same time, the threat of sending the Department of Justice, sending the Department of Commerce for export control, and for other different measures is a—you know, a threat. And so my take at this point is on the surface, it’s voluntary commitment. But I think in practice, the companies have to navigate this very closely and carefully and cautiously, and therefore it’s a mandatory regime.

And so, you know, like, the White House can publicly say our executive order, you know, that is expecting the companies to do pre-deployment testing for cybersecurity risk is voluntary. You have thirty days to do it. It’s voluntary.

But in practice, it’s quite mandatory because if the White House doesn’t let them and approve or greenlight the release of a certain model, they will be in trouble, right? And so it’s voluntary until it’s not, and right now I think the companies are OK—my sense—with the level of stability because they want stability. They want, like, some clarity and stability.

So they’re getting a level of clarity and stability. But my take is that it’s a temporary arrangement until something breaks out. It could be an incident. It could be something that is quite disruptive that is interfering in our financial stability. That is when they will have to rethink this.

But right now, I think we’re kind of like status quo. We’re OK with the arrangements on the surface and in practice. We’ll just have to wait for an incident and that will change.

FASKIANOS: All right.

I’m going to take the next question from Farid Murad—raised hand. If you can unmute yourself and tell us where you’re from.

Q: Hi. Can you hear me?

FASKIANOS: Yes.

Q: Yes. Thank you for picking me.

So I wanted to ask a question regarding the piece about accountability. So to kind of introduce myself, I’m a graduate student at Florida International University studying global affairs and security, and I do research under the Jack D. Gordon Institute Intelligence fellowship.

So my areas of studies are usually more of the humanitarian aspects of war and drone warfare and stuff like that, and I’ve become more interested in AI over the past year or so that I’ve been in school, and I want to—I’m currently trying to figure out a solution or a policy suggestion for research that I’m doing on possibly regulating AI itself similar—in the same fashion that weapons of mass destruction were regulated around the world, so as far as who can use it, how it’s used—in the sense of AI, like, how it’s used, what servers are they going to be on, if any.

My question essentially is do you think it’s possible in any shape or fashion to regulate AI with how difficult—because the way I’m thinking of it is that it almost—would be almost impossible because of how narrowly impossible it is to regulate the internet, for instance, right? So seeing how they’re kind of similar to each other.

If the U.S. or world powers were to convene to create policies to regulate AI, would that be possible? Because AI is being used right now, as you know, for training terrorist organizations, funneling money, and so forth, stuff like that. So kind of curious.

NGUYEN: Yeah. Farid, thank you so much for the question.

My take is maybe in 2022 and 2023 we had a chance to at least manage it. Not anymore. It is not possible at this point—technically not possible to restrain, constrain, limit the movements, the diffusion of the technologies—the technologies out there.

You can gate it. You can limit the access. You can limit the compute. But if you have a model that is capable to do whatever you want, right—it could be funding and enhancing terrorists or cartel operations—it’s just, like, out of the box.

And so right now, the gates, the fences, the limitation, you know, on exports are, to me, temporary. The technology is diffused.

What we can do is that we can provide a version of the technology that is secure and safe. And so you cannot eliminate a hundred percent, but you can incentivize people to use a secure and safe version and do it well and run it in a secure way 80 percent of the time, and then you have to put the resources to minimize, mitigate, neutralize the 20 percent misuse of the technology.

So I’m, you know, practical here and realistic here. Governments can, you know, come together tomorrow with a treaty. OK. Everyone will say, we shall not, right?

How are you going to implement that? How can you, like, limit that? If you look at our export control, our entity listing, they can go a long way, but it doesn’t really prevent China from acquiring banned, you know, export-controlled equipments, right?

And so it’s not going to work. But what you can do and be practical in the policymaking space and in the national security space is that you can say, OK, bad environment. So incentivize everyone to use the good parts so we can cut down the risk. All the little parts we just have to work together to squeeze that down, and prioritize the threats, right?

Are we concerned about terrorist use? Are we concerned about cartels? Are we concerned about cyber criminals conducting fraudulent operation at scale? What do we really care? What is damaging to society? And we just have to be realistic, cutting those things down one by one.

Alternatively, we can just debate all day long and we haven’t really solved any of the problem. And so I’m just being realistic on where we are today. But thank you so much for the question.

Back to you, Irina.

FASKIANOS: I’m going to take one last question from Allison Hughes, who’s an undergrad student at Culver-Stockton College: Are critical thinking skills in our youth in danger as we continue to rely more heavily on AI for things like schoolwork?

NGUYEN: Absolutely.

You know, I grew up in Vietnam. We do math in our heads. (Laughs.) We don’t use a calculator. And when I immigrated here to Boston, I was surprised that people cannot do math, you know, for some reason, right?

That was quite simple. I think these are muscles, right—reasoning, thinking, interacting, integrating, synthesizing.

These are human activities that I think we should cherish and we should protect, but it’s really our personal choice, right? But I think, overall, if we—if we don’t use the skills that we have, we will lose the skills that we have.

And so my take is quite easy is that you’ve got to practice the writing, the thinking, the reasoning, the synthesizing, the connection, or otherwise you lose that and that you have to rebuild it when you need it.

And so across the board, yes, I think if we are heavily reliant on it, we will basically relinquish our agency to it, and that’s all we really have is our free will and our agency, and so without agency, we have no liberty and no freedom.

And so I think we have to think deeply about what we do and what we don’t do, but we shouldn’t be, like, allergic to the AI, right? Like, we should be OK, be open, learn about it and set expectation for us and for what we do.

And so it is going to be difficult. It will be challenging for sure in the next few years, but I am very hopeful that we can overcome this and harness this technology like we have always done.

FASKIANOS: Wonderful.

Vinh Nguyen, thank you very much for doing this for us.

And I encourage all of you to visit CFR.org for the work that Vinh and Kat Duffy and others are doing on the LEAD AI initiative. You can find more—Vinh referenced it, but more on our website. So we have a lot going on in the AI space so I hope that you will take advantage of that.

Our next Global Affairs Expert Webinar will be on Wednesday, October 28, at 1:00 p.m. Eastern Time with James Lindsay, who is here as our distinguished senior fellow in U.S. foreign policy on what the midterms mean for U.S. foreign policy.

And, again, going back to our website, I encourage you to go to CFR.org, ForeignAffairs.com, and ThinkGlobalHealth for research and analysis on global issues, CFR.org/education for teaching and learning resources that are free, and also to learn more about CFR paid internships for students and fellowships for professors at CFR.org/career-opportunities.

So again, thank you, Vinh, and thanks to all of you for being with us. Enjoy the rest of your day.

(END)