participate in today's discussions in different ways, and I think it just demonstrates the depth of commitment and interest about these issues here at the board. So I'm going to focus today on artificial intelligence, a very important but small subset of the issues that David raised. Our focus in this conference is financial inclusion, and something that will likely have great consequences for financial inclusion and our economy more broadly in the years ahead is artificial intelligence, and in particular generative artificial intelligence.
As I've explored in a number of speeches over the past several years, AI has the potential to transform lives and the US economy, possibly empowering workers to be more productive, with lower- and middle-income workers perhaps benefiting the most. But it is also the case that AI may instead exacerbate inequality, eliminating some lower- and middle-income jobs while boosting the income and wealth of higher-income individuals. Since we don't know which of these futures will come about, it is useful to use potential scenarios, as I've done previously with respect to AI and the broader economy.
Every major technological advance has had profound effects on labor markets and the economy. Many workers have suffered from these technological changes, while many other workers have seen new opportunities emerge. In the long run, technological advances tend to broadly raise living standards by creating more jobs than they destroy and increasing productivity. But transitions and outcomes can vary, and in the period following the mass adoption of a general-purpose technology—such as electricity, the telephone, and internet-enabled personal computers—the number of people dislocated and the extent of the harm they may suffer can be large and persistent.
Balancing these scales of costs and benefits involves examining whether the benefits are broadly shared or concentrated. Past experience has shown that technological leaps forward can raise living standards. But when the benefits are concentrated among relatively few people, technology can widen inequalities of income and wealth, especially during the transition period. Widespread adoption of the internet raised the productive capacity of our economy and broadly raised living standards, but it also likely exacerbated inequality because it benefited certain information-intensive jobs like accounting much more than other jobs such as construction workers.
The policy challenge is therefore not simply to observe the development and deployment of AI, but also as a society to consider policies related to AI and its effects on education, job training, and workforce development, competition, tax policy, and other areas that may allow the gains from AI to be shared across workers, households, and communities rather than accruing to a small group of firms and investors. Of course, these policies are not within the remit of the Federal Reserve, but rather for other policymakers to consider and decide.
The question I would like to consider today is whether AI will likely help narrow inequalities of income and wealth, supporting advances in financial inclusion, or widen those inequalities, undermining the recent gains in financial inclusion that we rightly celebrate today. Let's start with an understanding of income and wealth inequality.
To understand income inequality, it helps to break income into its components. The largest component is labor income. Disparities in labor income across individuals reflect the supply and demand for their skills, their productivity, and their time spent at work—all things that could be affected by artificial intelligence. Income also includes earnings from capital and investment, and thus includes the concentration of ownership in firms such as AI companies.
In 2024, the highest-earning fifth of US households earned 52% of all income, and the bottom 20% earned only 3%. In 2024, the United States was the sixth most unequal of the countries in the G20.
Wealth inequality is a function of the distribution of ownership of assets, land, goods, businesses, intellectual property, and other investment assets. The bottom one-half of US households hold less than 3% of wealth. The top one-tenth hold 59%. And the top one-tenth of 1% hold 15% of that wealth. When returns from investment are reinvested, wealth naturally compounds. As a result, those who already own appreciating assets often see their wealth grow much faster than households that rely primarily on wages, widening the gap between the haves and the have-nots.
A central question is whether AI will expand opportunity by giving more people access to valuable skills and productive work, or whether it will reinforce advantages that are already concentrated among a smaller population. Inequality matters not just for workers today, but is also closely connected with an important aspect of the American dream: the expectation that in the future, our children will be able to make better lives for themselves, including through rising living standards.
Let me start with the possible ways in which AI could widen inequality. First, AI could lead to labor displacement. Something that is the top of mind for most people, especially younger workers, is the concern that AI could drastically reduce the demand for them. AI might disproportionately affect new entrants to the labor market.
According to a well-known paper by Claudia Goldin and Larry Katz, in some previous technological waves, the benefits have tended to improve outcomes for more skilled and more educated workers. But in this scenario, AI could harm not only less skilled workers but also younger college-educated workers whose skills are more easily replicated by AI than in prior technological waves. Moreover, workers who use AI more intensely might gain the most, and workers who use the most advanced and expensive AI models might win out over those who use baseline models.
In the Federal Reserve's most recent Survey of Household Economics and Decisionmaking, 43% of workers with a graduate degree reported using AI in the previous month, compared with only 10% of workers with a high school degree or less. The survey further found that workers who used AI were more likely to say that it would improve their careers rather than replace their jobs. Measures of exposure to generative AI also suggest that higher-educated and higher-paid workers are much more exposed to that technology.
The consequences of these facts are not yet clear. If AI mostly substitutes for labor, then not being exposed to AI should be positive for those workers. However, if AI augments existing jobs, then workers not exposed to AI would be left behind. There is substantial uncertainty about how the labor market will evolve. As of right now, there has been little evidence of economy-wide job displacements from AI. Yet, there is also some evidence that AI may have made job entry harder for young workers in some job categories.
Given both the history of major technological advances and how early we are in the timeline of AI adoption, it is important to consider the full range of possible future effects, including the potential for more widespread labor displacement.
Another potential concern is concentration. A high degree of market concentration has important implications for individuals' economic outcomes. We don't know how the market will evolve. At one extreme, competition and distributed innovation could lead to AI becoming a cheap and ubiquitous commodity. In this scenario, access is democratized and gains are widely shared. Startups and smaller businesses would have access to state-of-the-art AI resources and could continue their role as key sources of innovative ideas, goods, and services as well as important engines for job creation for the US economy.
But according to a 2025 paper by Anton Korinek and Haivipul, an important factor is that AI has some characteristics that have in the past reinforced concentrations of market power. Like other high-tech innovations, because AI depends on access to data, model improvements, and computing power, it benefits from economies of scale and scope. Greater data, model improvements, and computing power yield vastly greater intelligence and capabilities.
The high return for this advantage helps explain the huge investments and concentration of AI investments in giant firms, which is why they're referred to as hyperscalers. AI has another feature that seems to be driving ever-greater concentrations of market power: the fact that AI itself is a powerful tool to train and accelerate development of new AI models. That is, AI improves its own research and development. While computing technology has always tended to support the market dominance of industry leaders, the extent of the advantage possessed by AI industry leaders may prove to be unprecedented.
As a result of these forces, it is possible that a small number of AI firms may dominate the market and investment returns may accrue primarily to owners of AI. In this potential future, wealth generation and to some extent income generation for those workers who can benefit from access to AI resources could increase. But less access to ever-improving AI resources for most other firms and their employees would mean slower productivity growth for them, and a steadily widening gap between them and the firms and their employees with more access.
As I said at the outset, I'm not predicting this particular outcome, just exploring it as a possible scenario. Let me now turn to scenarios in which AI could reduce inequality. One scenario is broader access to capability building, resulting in broader productivity gains. Just as the printing press democratized knowledge and the internet democratized information, AI may democratize capability itself.
By giving millions of people access to tutoring, coaching, writing assistance, programming support, and problem-solving guidance, AI could enable individuals to develop skills that were once reserved for those with exceptional education, wealth, or mentorship. In the future, AI could be a rising tide that lifts all boats, rather than widening the inequality we've seen in recent decades.
A growing body of research suggests that AI can augment worker productivity without necessarily replacing workers, with especially large gains for less experienced workers, allowing individuals to perform a broader array of tasks and increasing overall economic output. In one experiment, college-educated professionals completed a range of assignments—writing short reports or analyses—and then carried out a second round of assignments with the help of AI.
AI reduced the average time to complete assignments by 40% and improved the quality of the results by 18%—a combination that constitutes a significant productivity gain. The biggest improvements in that study were among those who did the worst on the unaided assignment, narrowing the productivity gap.
Acquiring skills or building expertise and experience in any occupation is hard. It demands effort and discipline, often requiring years of effort, and depends on education, the capability to acquire and retain knowledge, and the judgment to use knowledge and other resources wisely. AI has the potential to expand expertise, shortening the time and reducing the work needed to build skills or directly providing those skills themselves, raising productivity the most for lower-skilled, less-educated workers.
AI could also lower the barriers to entrepreneurship for those with a good idea who may lack certain skills, such as finance or accounting, to develop and implement that idea. The democratization of coding capabilities enabled by AI models is also an example of how AI could level the playing field for a critical input for many businesses. While it's too early to know how helpful AI might be to entrepreneurship, the rapid adoption of AI among small businesses is a strong signal of its potential value.
Another way AI could help reduce inequality is through the creation of new jobs—ones we haven't even imagined yet—as we've seen with other general-purpose technologies. Consider that by some estimates there are 12 million full-time social media influencers earning a living today in the United States, something that was unimaginable a decade ago—or at least unimaginable to me.
Research suggests that major technological advances lead to this type of job creation, and the bigger the advance, the greater the impact on labor markets. For those who tend to see job dislocation from technological advances as a zero-sum game, it is worth remembering what economists call the lump of labor fallacy. The labor market is not zero-sum.
For example, spreadsheet software such as Excel replaced the lower-skilled aspects of basic accounting jobs, which raised the productivity of accountants. That is, instead of substituting for a person, technology augmented that person's ability to do their job. Instead of eliminating a field, technological advances redefined what was possible. AI may be able to do that across the economy. If so, AI could help lower inequality by creating new, more productive, and higher-paying jobs.
What I have laid out thus far, roughly speaking, is the worst case and the best case scenarios for how AI may affect inequality, and there could be many scenarios in between. It's impossible to predict now which of these versions of the future is more likely. But what I'd like to do next is talk through a couple of examples of factors that could shape that outcome.
The first is education, job training, and workforce development. In the same way that computer skills are essential to many jobs today, proficiency with AI may well become a necessary skill for the jobs of the future. Workers will need to prompt AI, integrate AI into workflows, oversee coding agents, and manage multiple AI agents. They will need to exercise judgment of AI inputs and outputs, verify AI results, and expand the frontier of human knowledge.
AI's amazing facility in writing computer code is likely to replace individuals whose only job was to write basic code. However, it will empower not only advanced coding experts who will oversee coding agents, but also empower many more people without formal coding training to use coding to turn their ideas into functional programs. If this is hard to visualize, consider how calculators, word processors, and presentation software have democratized the workplace. Not long ago, these functions were carried out by specialists, but are now minimum qualifications for most office jobs.
Education will be critical in how well workers in the US economy adapt to the AI revolution. Economists Claudia Goldin and Lawrence Katz explored this history in The Race Between Education and Technology, recounting how both the supply and demand for education responded throughout the 20th century to the technological needs of the economy. First, the dramatic growth in high school completion helped move the country away from agriculture as the dominant industry. In the second half of the century, the proliferation of higher education responded to the need for specialized expertise in many occupations. In this next wave, a crucial question about how AI may affect inequality will be if high-quality and affordable education and training are widely available not only at the outset but also throughout one's working life.
Yet a focus on AI skills alone is not likely to be the right approach. My instinct is that curiosity, flexibility, and importantly, common sense and human judgment are likely to be critical skills in this new economy. As Ethan Mollick argued in his book Co-Intelligence, success in the AI era depends less on mastering the technology itself than on developing the human capacities to ask insightful questions. This includes being able to distinguish sound reasoning from plausible-sounding but flawed reasoning.
