Keynote Remarks by Mary C. Daly on AI and the Economy

Hey, good afternoon. Say thank you for that kind introduction and thank you to the Silicon Valley Leadership Group and San Jose State University for having me here. Even with the turnout today on a rainy day, it's just a delight to be here and I look forward to a great conversation. So today I want to talk about a topic of the day, AI. Now standing here in Silicon Valley, many of you and I talked to I think half of them this morning at a round table may feel quite excited about AI, thinking it's full of promise and possibility. But for others, it may feel more worrisome, a potential disruptor to lives, livelihoods, even what it means to be human. I would offer this is not especially surprising. People also had mixed feelings about electricity and the automobile with some fearing them as dangerous, supernatural, even evil, while others imagined their potential and quickly adopted the novelty of driving or turning on lights in their homes. Irrespective of where people stood on those two technologies though, almost everyone wondered how they would change the role of humans in the world. But in all of the technologies, knowledge gives us power to separate facts from fears, speculation from reality, and worst case scenarios from more likely outcomes. And so this is my topic for today. What we know about AI, what remains uncertain, and how monetary policy should respond in both the near term and in the longer run. Before I go further, if you haven't been to a Federal Reserve official talk, I need to remind you that the views I expressed today are my own and do not necessarily reflect anyone else on the FOMC or in the Federal Reserve system. So, I'd like to start with a simple maxim. Transformations take time. Let's think about electrification. Sitting here with lights and computers and fully electric vehicles, it's easy to imagine that it just happened. It got invented and there it was. But the truth is it took nearly 100 years to move from Faraday's generation of electric current in 1830 to electricity boosting productivity growth and transforming the economy making it what we know today. That's a hundred years. And the evolution is instructive. So I'd like to show you this chart. The figure shows the timeline of electric innovation and productivity growth in the United States. You can see Faraday's discovery early on. Many things had to happen to turn Faraday's discoveries, however, into practical applications. Innovators had to develop the light bulb, the electric motor, and the electric unit drive. Builders had to lay out an electrical grid, install transmission lines, and ensure sufficient access for households and businesses to make electricity a worthwhile investment, something more than just the early adopters would be interested in. And this was only the beginning. To create sustained gains in productivity growth associated with a general purpose technology, the very essence of work had to change. Production had to be reassigned. Factory floors re-engineered and workforces retrained. Firms had to shed the constraints of the previous steam powered world and the thinking and reimagine business in an era of electricity. And the scale of change was simply enormous and it took time. The question is should we expect a similar long period of access, adoption, learning and transformation for AI? Well, the answer is maybe, but we're already well into it. The origins of AI date all the way back to the 1940s and 50s, and it is actually evolved much like electricity did, as you can see in this next figure. This figure shows the evolution of electricity on the top and the evolution of AI on the bottom with major periods of development for both technologies highlighted. Now if we focus on just the AI part the bottom of the chart you can see that it began with foundational research on neural networks and thinking machines and around that same time it was first recognized as a field of study at a camp at the Dartmouth campus. Then came the experiments use cases and eventually applications like applied machine learning and robotic processing. And these use cases and experiments demonstrated AI's value, but fell short of creating sustained gains in productivity growth, ones that just keep generating again and again and again. Then in 2022, Chat GPT and other large language models with natural language processing became available to anyone with a computer or a smartphone. Their accessibility and humanlike interface encouraged individuals to experiment and businesses to try and harness their capabilities. The arrival of LLMs for the general population is a major change in AI technology and access. And the outcome that it produced has been quite dramatic. We heard this in our business and community outreach and we experienced it with our own employees. AI went from science and sci-fi to everyday work and conversation. I think at this point we know that this surge moment for AI was more than simple enthusiasm. Excitement has turned into real investments. Businesses of all sizes are working to leverage AI. You know, as you heard in the introduction, I travel around the 12th district. If you're not familiar with Federal Reserve regions, that's the nine states in the West, all the inner mountain states, the coastal states, Alaska and Hawaii. And as I travel around, I talk to firms, large and small. And what we're hearing is that they're using AI for very practical purposes, consumer research, back office operations, sales, product development, and they're all using them right now to save time and money. If you go back to the economics literature, you find that research case studies show similar results. Cost savings when firms use AI to automate parts of their process in everything from call centers, software development, and financial management. We talked about that this morning in the round table, but even marketing and healthcare as the figure shows. And this is where those evolution highlights matter. Commercial applications and larger scale adoption is really critical if AI is going to produce the type of systematic increases in productivity growth that electricity has. So the question that we face today is is this going to be enough? Is Gen AI a sufficient catalyst to change the nature of production, work and business? And the truth is, if people are really being honest, no one knows for sure. People have beliefs, but no one really knows for sure. It's easy to see the possibilities, and I just spent an hour with people who see the possibilities, experience them every day. But it's sometimes harder to know when and how they will exactly evolve, especially as you walk through thinking about the rest of the nation and the rest of the world. So far, most macro studies out of the economics literature looking at productivity growth find only limited evidence of a significant AI effect. And even firms that say it's useful find have shown little evidence of transformative gains, the type that we got out of electricity. Importantly, some of this could just be timing. AI adoption and use are still evolving. So, it may be just too soon to see these things happen in the aggregate measures. It's also true that the technology is changing very rapidly. So firms could simply be focused on acquiring and learning the new tools like Agentic AI rather than rethinking their entire business model, waiting to see what comes next before they change everything. But it could also be that we're simply not there yet. That while Gen AI and related applications are useful, they are not the innovation that spurs broad-based reorganization of the entire economy. You know, I find it instructive to look at the financial sector. They were early adopters to some of the AI related products and several firms have talked publicly about using AI in the loan application process. Uses range from initial document review to checking the final application. While automating these steps in thinking about those silos saves time and money for the company, it still falls short of transforming the overall process. They're basically still taking one thing and passing it to another group which often is run by a human and so it hasn't transformed everything. So that is really analogous to taking the steam powered engine replacing with an electric engine and then leaving the factory floor completely unchanged. Good progress but not transformative. So that's the real question. Then what will it take? And I'd like to offer that it's something you likely don't expect. The factors that made electricity a general-purpose technology were not solely the technologies. They were the ideas that came with it. Innovative firms using imagination and creativity that started fresh and built a world shaped by electricity with electricity in mind rather than leverage electricity in a steam powered world. And I think the same path is likely to hold for AI. Technology will enable, but ideas will determine when it transforms. And this brings me to monetary policy and how we should think about AI as we work to achieve full employment and price stability, our congressionally mandated goals. Fortunately, this is not the first time the FOMC has had to contend with manage policy in a time of transformation or at least transition. In fact, I began my career at the Fed during a similar period, not with the formation of electricity. Let's just be clear. But, it was back in the mid 1990s, the beginning of the computer and internet revolution. Businesses were ramping up investment in IT equipment and software to take advantage of these emerging technologies, but there was little evidence in the official measures of productivity growth that this was having any impact. At the same time, the FOMC policy makers were concerned about elevated inflation, which was printing above target and a tightening labor market. And standard macro models and monetary policy thinking suggested that that interest rates should rise to ward off overheating the economy and the labor market and producing a more significant runup in inflation. But then Chairman Greenspan offered a different view, a different perspective. He actually distrusted the official productivity numbers, finding them at odds with the surge in IT investment he was seeing and we were all seeing, right? They had two things that didn't match. So he posited that the burgeoning computer revolution while not yet visible to us in the aggregate measures would spur productivity growth allowing the economy to grow faster without putting upward pressure on prices. In this scenario the economy would not need to be bridled back and interest rates would not need to increase. So he found evidence for this possibility in the disaggregated data, the micro data, including what he was hearing from executives, people who ran businesses. He found that wholesale and retail firms using inventory management systems to reduce warehouse stockpiling were making transformative gains. Trucking companies leveraging GPS to reduce deadhead hauling across the nation, not having truck beds empty from one location to another. Manufacturing firms using computer assisted designs to do mass customization and reduce waste. And we actually asked the same exact questions in the San Francisco Fed. It was one of my early jobs going out in the field. We talked to firms and we walked factory floors. We looked for evidence that computers and software were making a difference. And we looked everywhere, not just in firms that were technology firms, but in fact in firms that were not. And we found numerous examples. But here's the real thing that I saw and we all saw back then. We discovered something bigger, more significant, if you will. We saw businesses asking questions about how they could fundamentally alter the way they produce things, goods, services, do their business. And their focus became factory design, business process, using computers and software as a foundation. Basically putting up a whiteboard and saying what would the world look like if I'd always had computers and software solutions. And that started to do what electricity had done? Maybe not at the magnitude or scale, but it began to deliver sustained gains in output revenue and longer run profitability. Sustained productivity gains came and they mattered for those companies. So policy makers took this information which was evolving over a longer period than just a few weeks or months but policy makers took this information and they debated you can go back to the transcripts and see this whether and when these micro examples could lead to the type of economy-wide productivity growth that Chairman Greenspan considered. In the end, the FOMC remained more patient on policy than our models and our thinking would have called for, and the roaring 90s followed with strong labor market and sustained growth. So, what do these lessons from the 90s teach us about policy today? Well, first that we won't find all the answers in the aggregate data on productivity, the labor market, and inflation. Seeing developments before they fully emerge requires digging deeper, relying on disaggregated information that foreshadows transformation. You've got to look almost for the green shoots of what's possible. And you got to do that in the micro data, not the aggregate data. Second, there is no matrix, no written down handbook that tells you exactly what data to follow at a moment in time. You know, Greenspan's innovation to my mind wasn't looking at more data. It was looking at the right data for the moment he faced. Finding inconsistencies in what he saw and then working to resolve them, bringing as many facts to the table as he could assemble. The third lesson is that talking to businesses matters. Businesses invest, experiment and learn long before we see things in the aggregate information. Incorporating this data into and seeing into our business of policymaking and asking ourselves whether it supports or negates what we expect is essential to making appropriate policy. In the end, I think we now know and history suggests that it wasn't just good luck that allowed Greenspan and the committee to navigate the 90s. It was also good practice. The willingness to listen to businesses, hear their ideas, use data and evidence to test them, and then invite others to do the same. There sort of this constant process of being curious and testing. Being curious and testing. So turning back to today and thinking about policy now, none of what we learned in the 90s or in previous periods of history is a playbook for managing the economy and AI, but it is a foundation, a reminder that monetary policy is a forward-looking business. We have to look ahead to make good decisions at the moment. And to get it right, we must be grounded in what we see, but also open-minded to what is possible. The willingness to confront what we know and what we don't know is essential to making appropriate policy that's endurable policy that serves all Americans. Thank you very much. I look forward to a great conversation.

We don't know. I mean, that's the part that makes it hard is in productivity numbers, especially when they're happening in what I would think of as real time, it's very challenging to assess or draw it back to exactly what the factors are that have shaped it. You know, in fact, people still don't agree on what happened in the 90s all the time if you look at research. So, it's just something to keep in mind. So then what you do what any good economist or person any industry person would do is they'd say what am I seeing what am I seeing and so right now while we can't find it in the macro studies that would do very sophisticated empirical econometrics and ask the questions how much of this is AI we still can see that there's something going on there the question isn't is it happening the question is how long will it persist and so clearly something's happening in the economy. But if you make a series to go back to your question about productivity, if you make a series of one-time adjustments, so say you automate a production line or you use AI to help in loan application process, you save money once. You don't save money forever. And I mean, you keep saving that money, but you don't get growth out of that. You don't get productivity growth. You get one-time adjustments to the level of productivity of your employees or your process. So what we're looking for is a technology to give us consistently good changes in productivity so that all industries at scale get better. Industries figure out new ways to generate revenue, new ways to do product design, new ideas to come and shape the economy. That's the thing that has a sustained productivity growth part. So it's undeniable productivity growth has gone up. What's not as clear is how long will that last.

It is absolutely possible and something we have to interrogate. I mean right now as you know too well inflation's still above our target our 2% target and price levels been high for much higher for a long time and people are feeling stretched by the high inflation that they see and now oftentimes people say well now AI is going to take hurt the labor market and so now I'm in double doom as people say but I think ultimately the way you think many people think about AI is the investment part of any technology can actually boost demand for goods, services and people and can then raise the pressure on inflation. But then the productivity part comes and that can be that's a disinflationary part. You can see this is all about the timing. And so what we end up investigating is not just the models but asking questions. Are the buildout of data centers raising prices for construction workers? Are the buildout of data centers raising prices for metals and other things that go into them the raw materials are the productivity gains and then on the other side of that are the productivity gains already affecting the cost structure of firms do they see that and even if a series of one-off adjustments can actually change the cost structure and if you look at profit margins when prices haven't been rising as rapidly as they once were and firms are saying they don't have as much power to pass through you would think that they're doing something to help margin protection. And so I think this is there's something going on here. Whether we want to link it all back to AI and then use that as a demonstrated proof that we're in a transformative state, I think that's a little bit too far. But certainly something's happening. And thinking about the productivity growth is exactly what, you know, we did back in the 1990s. We saw evidence firms were being more productive. We were interrogating how long that would last. And interestingly, the 1990s when I said it was the roaring 90s that followed, it was good growth, but it was also a good labor market, a really strong labor market. And so those two things went together because ultimately we had this conversation in the round table and one of the participants made a great point. It's true. Economics is this is how economics works is if an employee using AI gets much more productive, you hire more of them, right? Not fewer of them. So, you know, the economy grows faster, the product development goes faster and demand gets stronger.

Well, I think we have to separate what we're talking about into now, next, later. So, let's think about now. Right now, we have more demand than we have supply for energy, for electricity. Right? If you talk to, we regularly have CEO roundtables with the big power companies across the 12th district. You can look throughout the nation. Demand for power is higher than the supply of power. And there's a lot of reasons why supply has fallen behind. One is demand's just gone up rapidly. But another part is that they've got an aging infrastructure. They have to get those things built out. It's a highly regulated industry. So the infrastructure doesn't just come on like a light switch. Then you have there were supply chain issues that made it hard to get the transformers and other things. So all of this just adds to the imbalance of demand versus supply. But the remedy for that isn't to take away demand, it's to increase supply. So when any CEO of a power company says we can solve this problem by adding more supply. Absolutely. But that's a next and later. And so what you said my constituents what consumers and businesses are saying is I'm worried my electricity prices are going to rise and they've already been going up. And the CEOs of power companies are saying, but if we just keep building, that will go down. And both are true.

You know, it's hard to say. Energy could be disinflationary if we get to a point where supply is greater than demand. Right now, I'm just looking for supply to equal demand. And that would be a big benefit to consumers because it would mean that we wouldn't keep seeing inflationary pressure coming out of the energy sector. The other thing I wanted to ask you through the lens of constituents of the 12th district is one reason you might focus on productivity is there is a direct read through to GDP growth and other data sets that you can look at but the anxiety in the real world is well a job can make an AI tool can make me more productive or it can displace me all together. Where do you see that tension in the economy right now? So one of the things that is true is that the labor market has slowed but it's slowed for a whole variety of reasons and much like when you said well productivity has

I think we always want to be a little humble about the correlations we see and ascribing causality to them. So I wanted to temper your enthusiasm for thinking all the productivity growth is AI. It might be, but it could just be general cost management in a slowing economy or to manage tariff costs, etc.

So on the labor market, the labor market is slowing. It's slowing in industries that are directly telling us that they're using AI, and it's slowing in industries that aren't. So it's one of the things that I just want to be cautious about. When we talk to workers, we talk to communities all the time. What's true is that technologies are a really interesting thing. No technology ever reduces net employment. Not in the history of technologies, but it does change what that employment looks like.

And so there's a period of replacement. Right now it's replacement of tasks. So if your job has certain tasks in it, AI can do those for you. And the next part is augmentation. So you have replace, augment, and create. What's interesting about AI is that unlike, say, electricity when the candle lighters or the lamp lighters or the candle makers got displaced before the US completely became electrified, this is going more quickly.

If you go to a firm, I was on a panel at the Reagan National Library Economic Forum with Patrick Collison from Stripe. And interestingly, he said, 'I am hiring more coders than I'm laying off, but I am laying off coders whose technology skills didn't advance or they weren't the right workers.' And you're seeing this, right? You're seeing businesses reskill themselves to meet the AI moment, and that's going to cause worker anxiety.

And right now, worker anxiety is high. We're in a low firing, low hiring environment. That's already going to make people feel vulnerable. If you haven't found a job and you're newly minted out of college, you just think, 'I was supposed to get a job before I graduated. Now I still don't have one.' That's very worrisome to people. And then if you're thinking, 'Well, I might lose my job, but I don't know how long it will take to get another one,' then you're worried about that.

So, I think it's natural for the sentiment to feel nervous, but it's not the same thing as AI is taking all the jobs because what we're really seeing is AI is replacing tasks, augmenting workers. When we talk to firms, most of those firms are saying, 'I'm augmenting my workforce.' If you're in big manufacturing firms, they don't have enough workers that do skilled labor. And so, they're looking to augment their workforce.

And then we're also seeing jobs created. It's interesting. I gave a talk on this in 2023 and I used prompt engineers as the jobs they were creating, but those jobs are now being replaced. And so, you're... But AI...

It's a warning. You could think of it that way or it's an indicator. So, let's take the warning. The warning is you can't keep up. I would say let's use it as an indicator. It's an indicator that the technology is evolving really fast and workforces need to skill in durable skills. And durable skills are: be AI ready, be able to use AI to lift yourself in the educational space. You know, use the technologies that are out there to build your skills up because you can learn a lot fast if you train yourself to look at AI and say, 'Give me a lesson on how to write...' I've been thinking about this, how to write a smart contract from end to end. What sort of software would I need? How would I do it? What would the code look like? How would I test the code? How would I know it's right before I execute on this smart contract? And so you can do these things in an evening.

And then it's just about being able to do that. So I think that's the message for workers, and I would have taught my young self this same thing: if you put off technology because you're afraid of it, then you won't be in the first place of trying to use the technology to further your own abilities.

Okay, I'm ready.

Well, you know, we don't right now use AI in our monetary policy work, but we do use it in research. As researchers, if you go to any academic institution, you're going to see researchers using AI to see what they can do better on coding and other things, but also data and analytics. What do you see? The place we are there is AI doesn't give you answers to problems. It helps you get to the discovery perspective. So if I use AI as a researcher to look at a bunch of data, I still have to test my hypothesis. I have to go in with a hypothesis. What am I trying to answer? So that's the human person. And so that's why it's not particularly well tooled right now to replace our forecasters and our thinkers, our scholars who think.