From Digital Assets to AI | Emerging Tech, Financial Stability & US Financial System (ft. Lisa Cook)

Thank you for that kind introduction, Neale. Being back on Stanford's campus is always an honor and conjures up great memories. I spent several formative years here. First as a student in the AEA summer program, which prepares students to pursue graduate study in economics, and then as a national fellow at the Hoover Institution. To say these stints at Stanford were transformative would be an understatement. The summer program prepared me and set me on a new intellectual and career journey, and my three years here as a postdoc set out an entirely new line of inquiry and research. In fact, I started my research on patents and innovation or the economics of innovation here and benefited greatly from my interaction with economists here at SIEPR, the economics department, the business school, the law school, Hoover—and this included Ken Arrow, Tim Bresnahan, Jeremy Bulow, Milton Friedman, Avner Grief, Mitch Polinsky, Paul Romer, and Gavin Wright. For my decades spent in the Bay Area, here and at Berkeley—the school that shall not be named—I witnessed how seriously new ideas are taken, examined, implemented, and spread. It is always invigorating to return to such a center of innovation. I applaud SIEPR for holding this event to discuss AI and its power to influence the trajectory of the economy and transform the financial system. I know many in this room are grappling with how to harness this technology’s obvious multi-dimensional promise while being mindful of important risks. Having adopted machine learning in the AEA summer program in 2018 when I was director, and having used it in my own research before coming to the Fed, I arrived at the board of governors in 2022 raising questions about and urging the study and adoption of AI. So rest assured, policymakers at the Federal Reserve are also deeply engaged. Today I will start by offering my latest economic outlook with a focus on implications for AI for both sides of our dual mandate of maximum employment and price stability. Then, consistent with my long-standing support for responsible innovation, I will address the benefits AI could deliver for the financial system before addressing some of the risks and vulnerabilities the technology presents to financial stability. I will conclude by sharing how the Fed itself is embracing the power of AI to help ensure the financial system remains sound and resilient. To set the stage, I will begin with my economic outlook. Allow me to begin with inflation. Inflation is clearly moving in the wrong direction. Based on the latest data, it is estimated that the PCE price index rose 3.8% over the 12 months ending in April. That reading is well above our 2% target. The recent rise in gasoline prices due to the conflict in Iran was the primary driver. But even when one excludes the volatile food and energy components, core PCE inflation is estimated to have risen by 3.3% over the last 12 months ending in April, its highest reading since 2023. Inflation has been pushed up by shocks that should, in theory, be temporary and short-lived. Tariffs should also result in a one-time shift upwards in the price level, and the effects of tariffs on inflation should begin to abate soon. The path of energy prices is tied to the ongoing conflict, the results of which are highly uncertain. Still, most forecasters and market participants, as reflected in oil futures, expect that oil and gasoline prices should decline to some extent by the end of the year. Nonetheless, even temporary and short-lived shocks should influence inflation over the medium term. Firms may embed these shocks into their pricing decisions, and workers may incorporate them into wage negotiations. Moreover, yet another shock to prices could be layered on from the heightened investment demand due to AI. To date, companies have announced more than $1.5 trillion dollars in data center plans, only a small portion of which have been realized. Those figures suggest that substantial AI-related investment remains in the pipeline from the data centers alone. Effects of this demand on prices are apparent: Prices have risen significantly for chips, other high-tech equipment, and software; wages in specialty trades and construction have picked up notably; electricity and water prices have each increased about 5% over the past year. Further, in the coming years, firms may expand along the intensive margin, but they may also expand along the extensive margin and undertake new AI-related capital expenditure such as in robotics. In contrast to inflation, the labor market appears to be largely stable. The unemployment rate, 4.3% in April, has remained unchanged, on net, since last summer. The rate is in line with estimates of the natural rate of unemployment, suggesting that the supply and demand of labor are roughly balanced. Despite some high-profile announcements of layoffs, initial claims for unemployment insurance remain low and stable. However, I view the downside risks to the labor market as being elevated. One factor is heightened uncertainty about output due to the conflict in the Middle East. A softening in demand could lead to a softening in the labor market. Uncertainty may also weigh on firms’ hiring plans, which may be one reason for the current low-hire environment. Furthermore, I have been and will continue to be highly attentive to AI developments and how they will affect the labor market. We could be approaching the most significant reorganization of work in generations. Even if in the long run new jobs are created, I'm aware that the timing and cost of benefits of AI may differ. Specifically, AI-related job loss could precede job gains. Although we do not have conclusive evidence of this occurring yet, it may still be on the horizon, and increased churn in the labor market could be anticipated. Businesses are adopting AI at an increasing rate, but many have not yet used it to change the way they organize work. Indeed, the vast majority of small business respondents in the Federal Reserve's 2025 Small Business Credit Survey say that their labor costs have not changed as a result of AI. Yet many businesses I hear from expect that AI will lead them to fundamentally change their business practices in the future. Finally, I will turn to economic growth. Here, I am optimistic. Over the past year, GDP growth has remained robust. Labor productivity growth has exceeded its pre-pandemic average. I do not need to report this in the middle of Silicon Valley, but business creation has remained high. Having done research on innovation and its macroeconomic effects for the 20 years before I came to the Fed, I believe AI is a technology like no other I have seen in my lifetime. But as a student of Paul Romer and endogenous growth more generally, I have been waiting for this moment when the post-WWII investment in knowledge—in the knowledge economy—would increase the arrival rate of ideas. As firms incorporate AI more systematically into their production processes, I expect that AI will further boost productivity growth, contributing to my expectation that GDP will grow robustly in the near to medium term. What does this mean for monetary policy? I see elevated risks to both sides of our mandate. And from a risk-management perspective, I currently believe that the right course of action is to hold rates steady. However, I want to be clear about my risk assessment. The risks remain tilted toward higher inflation. In my baseline forecast, disinflation should resume in upcoming months without having to raise rates. Similarly, I expect the labor market will remain stable without having to lower rates. After five years of above target inflation, I am particularly attuned to the risk that elevated inflation will become embedded in price- and wage-setting behavior. As such, I am prepared to raise rates if the expected disinflation does not appear in a timely manner. Likewise, I will continue to monitor labor market developments as well and would be prepared to adjust my policy stance downward should the labor market deteriorate. I will now turn to the theme of this conference: AI's effect on innovation, resilience, and risk in the financial system. I am excited to discuss this topic for at least two reasons. First, as an economist who began studying the economics of innovation in earnest on this campus some time ago, I see great benefit that could come from the financial system from AI. Second, I serve as chair of the Federal Reserve Board's financial stability committee. And so, as a policymaker, figuring out how to encourage innovation while ensuring the risks are contained and the system remains resilient, is a major preoccupation. Overall, I would like to stress that I believe in experimentation. This approach thrives in Silicon Valley and we are embracing it at the Fed as well. That is why I cofounded the Emerging Tech Economic Research Network, the system-wide effort to share AI research and results of AI experiments, and also why I have been encouraged to see staff at the Fed looking for ways to adopt AI technology in new and imaginative ways over the past several years. Not every effort will be a success, and as I learned from my training here and at Berkeley, that is okay. We are seeing results from this experimentation-driven mindset which I will talk about shortly after I discuss the broader benefits of AI-driven innovation in the financial system. I am optimistic about AI's promise to boost productivity and increase the arrival rate of ideas, which will support growth, lead to the creation of new firms that will disseminate new ideas and produce new jobs, and put downward pressure on inflation. Within the financial sector specifically, I'm excited about the benefits from AI we are beginning to see. The financial sector is adapting to the current generation of AI tools and increasing its adoption, initially in highly manual or resource-intensive areas. This transition includes in compliance functions, call centers and back office operations. Generating novel analytics has also become faster and more flexible. Using AI as a coding tool is helping the financial sector tackle age-old problems such as updating legacy code and integrating systems. Next-generation models should more broadly adopt and integrate into client- and market-facing applications. Large technology and financial services firms—those who provide the hardware, software, and systems that underly much of the global economy—use advanced AI tools to scan for potential cyber vulnerabilities that could be exploited. Further, AI adoption offers many opportunities for our financial system to be improved. These tools could allow firms to improve access to credit, allocate capital more efficiently, and speed processes. For example, AI could enable firms to accomplish the following: Develop new and better products that are more customized to individuals, broadening access to sophisticated financial products; provide retail investors with the tools necessary to identify trends and emerging risks earlier; and, leverage the benefits of efficiency gains to allow more capital to lending—to allocate more capital to lending and investment—which should lead to more economic activity and growth, as I mentioned earlier. Broadly, I see AI as stimulating economic growth which, all else equal, should support financial stability. However, as a policymaker, I understand that innovation can lead to increased risk if not monitored appropriately. I think about this likelihood both through the lens of AI's interactions with long-standing vulnerabilities, and of the risk a hypothetical AI shock would present to the system. AI might introduce vulnerabilities to the financial system through a number of channels. One of the most commonly cited is the increased prevalence of AI-driven algorithmic trading. Traditional algorithms are fast, simple rules operating at nanosecond frequencies, but they are relatively rigid and hard-coded. Gen-AI and machine learning add self-learning based on historical experience, adaptation based on current market conditions, and analysis of unstructured data, such as text. Policymakers and academics have noted that, increasingly, AI-driven algo trading may generate financial-stability risks such as more correlated trading, endogenous model collusion, potential market manipulation, and greater market concentration. Another potential risk comes from the probability that AI may displace or disrupt entire sectors. For example, concerns about AI disruption risk have affected speculative-grade bonds in the technology sector, where spreads have increased as our financial stability report noted earlier this month. These trends reflect AI disruption concerns in the software industry and arose after a large AI firm introduced products aimed at that sector. Concerns about credit exposure to software also contributed to the wave of redemptions that have put significant pressure on both traded and non-traded perpetual business development companies in recent months. Another emerging trend that may have implications for financial stability relates to the fact that firms are increasingly tapping debt markets to finance the capital investments related to AI infrastructure. Many of the hyperscaler firms have executed large investment-grade bond deals in recent months to fund AI capital expenditures. In addition, smaller data center developers are raising debt from private debt funds, as well as asset-backed credit markets, to fund their investments. While many of the largest investors are also strong borrowers, the increasing use of leverage to finance investments and an emerging technology carries risk, and a sustained boom in debt issuance could eventually represent a financial-stability concern. I will note that even under very ambitious investment and debt-issuance projections, we would be unlikely to return to peak leverage levels observed before the global financial crisis. And of course, we cannot talk about risk without discussing cyber risk. Recent advances in the ability of LLMs and agentic AI systems to detect, exploit, and create new vulnerabilities have introduced new challenges in safeguarding system security for financial institutions, infrastructure, and third-party service providers. Very powerful AI tools such as Anthropic’s Mythos Preview model have demonstrated the ability to detect previously undetectable vulnerabilities in software applications that support important and widely used computer systems. Non-malicious cyber events such as software malfunctions—we call them updates on three-day weekends—have also caused disruptions to the provision of financial services. AI can make developing software (particularly writing code) faster and easier. However, by contributing to the rapid proliferation of code, the aggressive use of AI may indirectly strain current security review processes. The ultimate implications of AI for cyber security remain unclear. Advanced AI coding agents can be used to enhance the security of many important computer systems to prevent future AI-related cyberattacks. It remains possible that AI makes financial institutions more resilient regarding cyberattack vulnerabilities. Just like the financial firms and other entities across the economy, the Fed is also looking to responsibly deploy AI to advance our mission and to improve our own work. To be clear, as I said at the NBER Summer Institute last year, the Federal Open Market Committee is not using AI in developing or setting policy. But many parts of the Fed system are using AI for a variety of other tasks, particularly in the area of financial stability, and we already see tangible benefits. By using AI ourselves, we can improve our analysis of the financial sector and are better able to highlight vulnerabilities—whether they are new ones introduced by AI or old ones that we may have missed. The use of AI can make us better at our job with enhanced monitoring and improved analysis. I would like to share two specific ways we are using AI to advance our critical mission of monitoring financial stability. First, newly formed teams of experts within the Division of Financial Stability are analyzing technological risks to financial stability. These collaborative groups assess how cyber, AI, and quantum computing create both vulnerabilities and opportunities. For example, economists Anne Lundgaard Hansen at the Richmond Fed and Seung Jung Lee at the Board have investigated the effect of generative AI adoption on financial stability through laboratory-style experiments using LLMs. Their research on herd behavior and investment decisions in a stylized lab setting found that AI agents make more rational decisions than humans. The research suggests that agents are more likely to make decisions based on data and analysis rather than simply following general market trends. This outcome could potentially lead to fewer asset bubbles arising from animal spirits. These innovative teams have also developed practical tools for our mission. One team designed a method to construct a small, cost-efficient AI model that can classify a large amount of text just as accurately as a larger model using a technique called “active knowledge distillation.” The method achieves up to an 80% reduction in computation costs while maintaining accuracy. This efficiency matters. Financial stability analysis increasingly requires processing vast amounts of unstructured text data, including regulatory filings, earnings calls, and news articles. Another interesting project applied natural language processing to decades of Beige Book data, finding that even when controlling for traditional metrics, the sentiment in these anecdotal compilations provides meaningful explanatory power in forecasting recessions. Second, the staff from the Board and all 12 Reserve Banks recently participated in an agentic AI sprint. This event encouraged experimentation and explored what agentic AI could do for financial-stability analysis. It was great to see all the AI systems that could reason through problems, decide which analytical approaches to use, and complete complex tasks with minimal human intervention. A valuable insight we found in one of the projects was that agentic AI systems can be more systematic in identifying network-based risks than our standard approaches. This outcome was not because we do not understand their theoretical importance, but because in many cases we lack the capacity to comprehensively analyze complex empirical structural patterns of networks at scale. This type of systematic capability translates into potentially meaningful efficiency gains for financial-stability work. For example, other prototypes demonstrated that they could select, run, and analyze many financial–stability-relevant scenarios that would be prohibitively time-consuming using traditional methods. This development enables the kind of thorough analysis that humans would struggle to complete in a reasonable time frame. However, and this is critical, systematic coverage without accuracy would be worse than a selective approach. The most promising approaches build verification into the system architecture itself. These approaches have multiple agents confer before reaching a consensus and include mechanisms that force the agents to consider contrarian perspectives. This process in turn can be cross-checked by researchers. If that sounds familiar at a place like Stanford, it should. The same types of methods that have yielded breakthrough thinking from humans for centuries are what I just described for AI agents. The totality of our experience with AI leads us to the conclusion that alongside experimentation, strong governance and risk management must be our foundation. The most promising approaches augment human judgment with AI capabilities while building verification into the architecture itself. The urgency is real. AI is advancing rapidly and financial institutions are adopting these technologies apace. As policymakers, we must understand these systems through hands-on experience. By building our own AI capabilities, we gain invaluable insights into both the promise and risks these technologies bring to the financial system. With appropriate governance frameworks, autonomous intelligence can potentially expand our analytical capabilities. These tools could enhance our capacity to identify and respond to evolving threats. But we proceed with both optimism and caution as warranted at this moment of a technological inflection point. Thank you again for the opportunity to return to Stanford and to speak at this timely and important event. I look forward to our discussion.

So, first I know that chairman Warsh has deep ties to the Stanford community. So I think that experience will hold him in good stead and we thank you for sharing him with us. So I can't speak for him but I can tell you what I have observed when I joined the Fed. The human capital there is so important. The staff are deeply committed to doing research that is based in data, based in analysis, and they are deeply committed as public servants. And I think that he will find this as he returns to the Fed—he probably saw it when he was a Fed governor. But I think this is going to be, I think, the most special constant that remains at the Fed when he arrives. And I think that he will try to keep this spirit alive, and I hope so. And that's how we rolled at the Fed. I love it.

So, as I said in my speech, one thing I'm paying close attention to is the PCE index and especially the core PCE index. If you're seeing the highest levels than we've seen for much of my time at the Fed then that's alarming, and certainly I was saying that this is moving in the wrong direction. One of the things that I'm very attuned to is inflation expectations and what you don't want to see is inflation expectations getting incorporated into the decisions of firms and into the wage behavior of workers. So I'm watching that very carefully. I'm also watching the price of oil. You know, I want to make sure that this is going down and I'm also watching if that starts going in the wrong direction, if it takes a long time, this can be problematic, not just in the short run. I mean, as I was saying, from the oil futures data the expectation from market participants is that ultimately oil prices will go down but there's so many risks associated with that—that prediction is highly uncertain and the timing is highly uncertain. With respect to the labor market, what I typically look at are three series: the unemployment rate, the labor force participation rate, and UI claims—unemployment insurance claims. And if there is a deterioration in the labor market, those three should pick that up. So, as I was saying before, I see some downside risk to the labor market. So, I'm watching those very carefully.

Those are the major ones that I use, too. And, but I use a variety of surveys. So the New York Fed has a great survey of different horizons for inflation expectations, and certainly I look at the Michigan Survey of Consumer Sentiment. And I mean, I guess one thing that you know well in your time in Washington, you have to monitor a variety of series—because one thing that you, Erik, many others might know: There are flaws in every data series. So, you want to have a dashboard so that if one series is flashing and the others are not, then you're going to pay attention to the totality of data. So, that's how I think about inflation expectations, but surveys and market-based measures are typically what I use engaging inflationary expectations.

So the most difficult part of understanding AI's implications for both sides of the dual mandate is understanding the timing and magnitude of these effects and those are just highly uncertain. And I was speaking to this dynamic inconsistency problem where you might have the productivity gains before you have the job losses or you might have the inflation before you have the productivity gains—we don't know where all of this will settle. So we're monitoring it very carefully; we're monitoring various series very carefully. And on the labor market side, I guess the first way I think about it is there's structural adjustment in the labor market as AI becomes good at certain tasks and people who are doing more of those tasks change the nature of the work or maybe even are changing jobs. And then there's sort of cyclical factors that historically sometimes we see greater investment in labor saving technology coming out of downturns. And it strikes me the cyclical factors are more front and center for the Fed—the structural factors maybe less so but I guess I'm curious as to your perspective: How do we think about the different types of impacts, their importance, the relevance of them for the Fed's policy response function?

And our personal lives, like I want a quarter of that taken away, too. Totally. When is AI going to do the dishes?

So you know, in the financial-stability framework we're monitoring all kinds of market developments all the time and this is just—AI-related valuations are just—one type of asset that we're following. So we follow an entire panoply of them, from household financial well-being to financial assets. I mean, we follow the entire spectrum. So I think that one of the big issues that we brought up in the financial stability report was this request—this heightened request for redemptions—that are associated with private credit, especially BDC's who are under a bit of pressure to exceed their redemption levels, their stated redemption levels. But I think that what we found in the FSR is that this is not a large problem right now, that they are structured to manage these risks. And I think that a lot more banks who are involved, the BDC's themselves, investors, are doing their due diligence now and reading their contracts; these funds were supposed to be locked up for some time and that's where the safety is supposed to come from. So, I think given that, there's a lot more introspection but we follow a whole host of financial-stability shocks and vulnerabilities. Shocks we can't do much about; they come from nowhere and they can amplify stress and we look at vulnerabilities in the sense that they are well-known financial risks and we want to make sure that those are managed so that the financial system stays resilient. So our risk-management perspective, either if we're talking about financial stability the same is true in supervision and regulation, we want to make sure that financial firms—the firms we oversee, those participating in the financial system—are managing their material risks no matter what the source of those material risks might be.

Oh, I think that's a great question and one that I get a lot. Now, I've made some parents and some students mad, so don't, you know, hold on to your seat—everybody still has to know something. Everybody still has to know something. When I am experimenting, not at the Fed but at home with different AI models, I have to check them each time. If I didn't know something, I wouldn't know which one would win in my horse races. So, you're still going to have to know things. And you're still going to have to be a problem solver. So, you're still going to have to have the skills. Now, I'm always promoting, I'm always a hype person for economics. So, that's not going to change. That is one way to obtain these problem solving skills. But anything that results in problem solving, I would say mathematics, statistics, economics, philosophy, deep language learning, phonetics. I think those will be very useful in this environment. And I think anything that helps with problem solving is going to be good. What we're beginning to see in the economics of education is learning loss associated with early adoption of AI. So I think that's where we have to be careful. So, don't over-rely on AI as a tool for learning—it is not a substitute. Somebody's going to have to know something. And what, when we're talking to firms, to our contacts, they're saying that the new jobs—some of the new jobs being created, some of the new tasks being created—is managers who are overseeing what AI has produced. So that means that if you can do that, if you can, if you have the knowledge base to be able to do that, then you should be in good shape. But I think there's no way around it that you're still going to have to know problem solving; you're still going to have to fact check. We didn't stop learning basic arithmetic just 'cuz we got a calculator.