The real macroeconomic impact of AI is moving from experimental models to real-world data. As the productivity dividend has yet to appear, reshaping wealth distribution rules and introducing a token tax are becoming future realities policymakers must confront.

On September 24, 2026, Harvard Kennedy School hosted a focused dialogue. Jason Furman, former chief economist in the Obama administration and a Harvard professor, and Peter McCrory, chief economist at frontier AI lab Anthropic, held an in-depth discussion on AI's impact on the labor market, macroeconomic guidance, and future policy directions.

As the head of a rare research team that publishes a "Human Economic Index" specifically on large-model applications, McCrory used Anthropic's vast underlying usage data to reveal the real obstacles in AI adoption, the potential economic gains, and the macroeconomic scenarios for 2030.

The J-Curve Effect and the "Missing Unemployment Wave" in the Data

A major mystery in today's macroeconomy is that AI capabilities are exploding exponentially, yet total factor productivity (TFP) has not surged significantly, and the U.S. unemployment rate remains steady at a historic low of 4.1%. This mismatch between perception and macro data has drawn intense market attention.

McCrory offered a core explanation: the time lag in technology diffusion and the "J-curve effect." He noted that while "the range of tasks these models can complete autonomously roughly doubles every four to seven months," firms must endure high restructuring costs before truly gaining productivity dividends.

"To truly unlock the productivity gains this technology could bring, supporting investment is also needed. Capability plus contextual information—both are indispensable," McCrory said. Large firms are constrained by data firewalls and organizational structures: "Unless you invest appropriately to reorganize that data and provide relevant context, there may not be much impact." Companies are currently going through a costly experimental phase, which forms the bottom of the J-curve for AI investment returns.

On the employment side, mass unemployment has not appeared. Anthropic's data confirms this: "This looks more like a skill-biased labor-augmentation model, rather than a technology that fully automates and replaces jobs at scale." McCrory stressed that models rely heavily on high-quality human input when solving complex problems, and human-machine collaboration remains the mainstream form.

Macro Guidance: Is the Fed Underestimating AI's Growth Engine?

Regarding performance guidance and macroeconomic growth for the coming years, Anthropic's estimates offer highly imaginative expectations. Currently, the Congressional Budget Office's (CBO) long-term growth forecast for future fiscal conditions is 1.7%, while the Fed's most optimistic expectation is around 2.4%. But in the view of frontier labs, this baseline may be too conservative.

By tracking the real task time spent by users on the platform, McCrory concluded: "Based on current models and current usage patterns, labor productivity growth could increase by 1.8 percentage points."

That means if these efficiency gains gradually spread across the entire economy over the next decade, combined with assumptions about capital deepening, U.S. macroeconomic growth could return to the high-growth era of the late 1990s and early 2000s. Furman pointed out that economic segments not yet automated—the weak links—may constrain unlimited expansion of overall scale, but a marginal increase of 1.8% in labor productivity is enough to force a reassessment of existing monetary policy and fiscal forecasting frameworks.

Reshaping Wealth Distribution: Labor Income Share May Plunge 15%

On the wealth distribution and inequality issues that markets care deeply about, McCrory acknowledged that AI is likely to worsen inequality between capital and labor.

Within labor income, although AI has the potential to give novices expert-level abilities quickly, actual data presents a harsher reality: "When you look at data on actual adopters, you find the main adopters are high-paid, high-skilled talent. From that perspective, I expect this will be a general phenomenon at the income level, or rather inequality will worsen rather than narrow." Professional expertise, as a multiplier, is amplifying top talent into "100x engineers."

In the contest between capital and labor, macro models provide more cautionary guidance. As AI-driven capital deepening proceeds, "in extreme cases, the labor income share falls by about 15 percentage points relative to the previous 60% share of every $1 of income." That means capital owners (shareholders) will capture returns from AI-driven economic growth far beyond the past, while labor's share of national income will shrink significantly.

Looking Ahead to AI's Future: Three Singularities and the Token and Capital Tax

Looking toward 2030, McCrory summarized the endgame evolution of the AI economy as three singularities:

Software singularity: models get better through recursive self-improvement.

Economic singularity: AI automation breaks through traditional growth accounting models, challenging whether the economy can achieve infinite growth within a finite time.

Coasean singularity: the large-scale application of AI agents will drastically reduce transaction costs, completely changing firm boundaries and the structure of economic exchange.

In tests of the Coasean singularity, Anthropic discovered a striking phenomenon: "More powerful models systematically extracted more rents when negotiating with weaker models; people did not realize this asymmetry in bargaining power." That means firms or individuals controlling frontier models will hold absolute pricing power and rent-extraction ability in future market transactions.

Facing such extreme wealth distribution and business-structure reshaping, McCrory proposed a future policy direction: tax reform and a token tax.

"Currently, the tax structure in the United States and many countries focuses on taxing labor. But if the labor income share falls, other tax systems need to be considered, such as taxing consumption." More importantly, to hedge the negative externalities of excessive automation on society, McCrory said plainly: "In the economic policy framework, I see a token tax as an incentive mechanism to curb excessive adoption... A token tax can be compared to a carbon tax or a tobacco tax." This would ease unemployment shocks and balance fairness and efficiency.

The following is a transcript of the conversation

Good evening, everyone, and welcome to the John F. Kennedy Forum hosted by the Harvard Institute of Politics. My name is Morgan Jay, a sophomore majoring in government and history and a member of the student forum committee.

Before we begin, please note the exits on the park side and on the street side of the Kennedy Forum. In an emergency, please proceed to the nearest exit and gather in Kennedy Park. Tonight's program will be live-streamed and archived online, and will include an audience Q&A session. Audience members who ask questions may voluntarily identify themselves, but are not required to do so. Now please silence your phones and join me in welcoming Harvard College undergraduate Zoe Becker.

Good evening, everyone, and welcome to the Institute of Politics Kennedy Junior Forum. My name is Zoe Becker, a first-year student majoring in English and a member of the forum committee.

Tonight we welcome Peter McCrory, chief economist at Anthropic. McCrory leads Anthropic's economic research team, which focuses on how artificial intelligence will shape the labor market and the broader economy. His team publishes the "Human Economic Index," sharing the latest research from the frontier lab and studying scenario models to understand how AI may affect different regions and industries in the coming years. Just last week, the Human Economic Index released a new scenario model looking at three different possibilities for the economy in 2030.

Co-hosting the forum with Mr. McCrory is Jason Furman, the Aetna Professor of the Practice of Economic Policy jointly appointed at Harvard Kennedy School and the Harvard Department of Economics, director of the Mossavar-Rahmani Center for Business and Government, and a nonresident senior fellow at the Peterson Institute for International Economics. From 2013 to 2017, Furman served as the 28th chairman of the Council of Economic Advisers, serving as President Obama's chief economist. Now please join me in warmly welcoming tonight's guest and host.

Furman: It's a pleasure to meet Mr. Peter McCrory here. I've read many of your articles, but we've never had a conversation before, so let's have one now. I have two dialogue plans, depending on your answer to the first question. I just want to know one number—what is P Doom?

McCrory: Is that a number?

Furman: A number.

McCrory: I don't have a personal number for that.

Furman: My plan was that if it was below 10, I'd ask you about jobs and productivity; if it was above 10, I'd abandon all my previous questions and ask something like "What should we do to survive?"

McCrory: Okay, great. Then we can focus on the ideas I outlined a bit earlier. Could you first briefly describe the responsibilities of Anthropic's chief economist? And what role should economists—not just at Anthropic, but elsewhere—play when thinking about this field that is developing at astonishing speed? We usually study the past to make precise estimates of whether something happened, but you spend a lot of time writing about the future, which is much harder.

The way I see my work and the work my team is doing is this: we are trying to understand the economic impact of artificial intelligence. This is challenging because AI is a general-purpose technology that affects every industry and almost every occupation to some degree, and it is developing very quickly, with capabilities improving very rapidly. There is a famous dashboard chart, which I'm sure you're familiar with—the range of tasks these models can complete autonomously roughly doubles every four to seven months, and the time to complete those tasks has shrunk from months to days. So capabilities are rising quickly.

This is a general-purpose technology that will hit the economy in a concentrated way, and its adoption speed is very fast. There are also other reasons it may automate the innovation process. So uncertainty is reflected not only in the outlook for the future, but also in the present.

The way I think about our team's work is: how can we generate data and research that can only be done inside Anthropic but should be placed in the public domain? So we produce the Economic Index—a privacy-preserving indicator that tracks how people and businesses around the world use Claude across different tasks, occupations, and regions. We make this information public in the hope that independent external researchers can use the data and find it very helpful for understanding this most important technology in real time.

Of course, we can also conduct more research on the most pressing questions, which both helps us find some reasonable answers about what is happening and perhaps shows the world which questions deserve the most attention. This work is like doing research in public, requiring illuminating answers to important questions.

As you said, most academic research revolves around finding a perfect natural experiment or a rigorously designed randomized controlled trial to truly help isolate the causal effects of a new technology or policy. We don't have the luxury of waiting one or two years to see research results; we need some insight into what is happening today. So we do this work publicly, and we also hope to receive feedback and criticism, and I'm sure you can give me some inspiration to bring back to my team.

Furman: Okay, let's see. Then I'll just say what I think. The biggest mystery right now, in my view, is this—let's hear your explanation and its implications for the future. A few years ago, I spoke with top experts at a frontier lab, and they showed me scaling laws, visually extending the curve to tell us where we would be in 2026. I rolled my eyes and thought, even with more processors, that doesn't mean these things will keep getting better. To be fair, they were right, and these capabilities far exceeded their expectations, and perhaps even mine.

They also told me that basically these jobs are gone now, right? I rolled my eyes again. My feeling was that they should now be on the West Coast telling someone else exactly the same story about how they were caught up and how they were wrong.

So why is this? The unemployment rate is 4.1%, basically flat, and one of the lowest levels we've seen. On productivity growth, you might barely say it has picked up slightly, but total factor productivity has not really risen. You could say that, but it simply isn't visible in the economic data. This is clearly completely inconsistent with the scale of this technology. Please explain—I want to first explore the current situation and then infer the implications of your explanation for the future. For now, why isn't it showing up in the economic data?

McCrory: This is a puzzle I often face both inside and outside the lab. As you said, scholars who have studied these scaling laws for more than a decade have done quite well at predicting capabilities, and even predictions made a decade ago about what these models might do in 2026 have come true. And perhaps, as you said, some aspects have exceeded those predictions.

I think perhaps one underappreciated aspect—and economists should know better—is diffusion, which is a long process. If you look back at the development of general-purpose technologies in the 20th century, some academic research suggests it took about 50 years for the technology to diffuse throughout the economy. From that perspective, even if there is a large productivity increase, the productivity level may be smoothed out and harder to parse.

Of course, AI is spreading much faster than previous technologies. I recently wrote a related piece titled "Why AI Has Not Led to Rising Unemployment." My observation is that labor productivity data show relatively stronger productivity growth in economic sectors where AI is widely used—not just a tiny change you have to squint to see.

But as you said, the unemployment rate is now near the lowest level the Fed considers consistent with its mandate, and the prime-age employment-population ratio is at its highest in decades. What explains what we see in many studies? From our data, this looks more like a skill-biased labor-augmentation model rather than a technology that fully automates and replaces jobs at scale.

According to the U.S. Department of Labor's ONET classification, in about half of U.S. jobs, we see that a quarter of tasks are things people use Claude to do. But across the entire classification system, Claude has not automated every task in any one job category. The key hard-to-automate parts of work limit the possibility of full replacement, so human-machine collaboration continues.

In addition, when we look at what really happens before model outputs, we also find complexity. If I ask Claude to build a very complex financial model, or if I see people systematically using Claude to model across tasks and regions, those people are providing complex, detailed inputs. Complex outputs depend to some extent on complex human inputs. We find evidence that models tend to be more successful when humans are more deeply involved—people can solve harder, more complex problems, and there is a tradeoff between complexity and success rates in using AI, so the space for full replacement is relatively limited.

All of this points in one direction: changing and restructuring the reward structure and returning to different forms of expertise, but not completely causing AI to massively displace workers and disrupt lives.

That said, if you look at areas of the economy where you might expect this—such as job turnover for technical writers or data analysts, or disruption of entry-level jobs whose core tasks are exactly what these models are already good at, and where not much of this technology is yet being systematically used to automate such tasks—in the labor market, unemployment has not risen. But these workers are more likely to express concern about potentially losing their jobs in the future, which seems to signal that something bad may happen later, even if it is not yet showing up in the data.

Furman: Yes. I agree with you that some of it may already be showing up in productivity—since 2019 our cumulative productivity growth has been 13%. Of course, you can dispute how many percentage points of that came from AI—half a percentage point, or none at all. I'm not sure, and I think most of it is not from the past year.

You didn't mention the J-curve as part of your explanation. You know, I feel like in the past six months I've written 900,000 lines of code, which will make me more productive for the rest of my life, but at least in terms of output in various respects, the past six months have not been ideal—except for lines of code. How many businesses are going through something similar? That is, something fundamentally different may happen in the next year or two, while the weak points of the current argument may still remain.

McCrory: That's a great point. I was talking about that and then set it aside, and I'm glad you brought it up. About a year ago, we studied how businesses embed Claude through APIs in an automated way. The API model itself has some problems, such as context mismatch, or the model autonomously performing certain actions, perhaps involving internal business operations, processing, or tax filings.

We systematically observed that for the most complex tasks businesses are using, Claude relies on far more contextual information than is needed for simple tasks—very simple tasks like writing an email require much less context. This suggests that to truly unlock the productivity gains this technology could bring, supporting investment is needed, not just capability—capability plus context, both are indispensable.

Think about large multinational organizations that have accumulated and developed over more than 50 years. Their structures are highly decentralized and they use firewalls to protect data. That data could be useful to the model, but the model itself cannot access it. Unless you invest appropriately to reorganize that data and provide relevant context, there may not be much impact. It's similar with organizational processes. If you want Claude to help you develop a sales strategy, you may need tacit information hidden in colleagues' minds. Without processes to extract that information and provide it to the model, the model may struggle to complete the task, even if in some sense it is indeed capable of doing so.

I think this is one way to understand the J-curve. Companies are trying different ways to embed and deploy tools, and this requires a lot of time and costly experimentation to figure out what works. Perhaps this will happen through creative destruction at the edge of the market—new companies can gain this productivity without readjusting existing structures, and they can be AI-native. This is all just beginning, so this may be an entry point for understanding where things are heading in a few years.

Furman: Okay, let's talk about the future. You recently co-authored a great paper—what was it called? "Transformative Scenarios for AI" or something like that. It had more equations than anyone here might want to see, but first, you can omit the equations; second, you can just hand it to Claude and tell it to "rewrite this, but without equations," and I think you can get a very good version. I didn't do that, but anyone here can try.

I want to ask about two scenarios you haven't mentioned, beyond the ones you already have. All your scenarios include combinations of higher productivity, higher economic growth, and higher unemployment—the more productivity growth, the more people unemployed.

The first question is: are there worse scenarios than what you described? For example, a lot of unemployment, chaotic lives, many displaced people, but no productivity growth—Daron Acemoglu thinks we might see that, and that is not one of your scenarios.

The other scenario you omitted, which is natural for many economists, is that productivity growth has nothing to do with replacing jobs. Historically, higher productivity has not necessarily been associated with higher unemployment in any industry, and even during transitions, unemployment may fall. Why do you think higher productivity will lead to higher unemployment this time? Let's settle that part first, and then we'll talk about your various scenarios.

McCrory: The goal of this paper, and the accompanying interactive scenario explorer—which is much easier to use than all the equations in the paper, and I encourage everyone to check it out—is a beautifully designed website where you don't need to look at any math formulas, and it guides you through how the model works.

The idea is that the economy consists of all the different tasks we each do in daily life, including writing emails, conducting statistical analysis, communicating with colleagues, helping patients, and all the things healthcare workers do. Then you can think about how AI affects tasks across the entire economy. There are five key inputs: first, what these models can do; second, how quickly they spread across the economy; third, whether AI is widely applied—by the way, these scenarios envision the situation up to 2030, which is already a very near future, not the super-futuristic kind.

Furman: Yes.

McCrory: We actually set the timeline at 2030 because beyond that you need to consider AI's impact on robotics and the broad uncertainty such impacts bring. Let's focus first on near-term uncertainty.

Autonomy refers to applying AI in an automated way, which can spread and substitute in the labor market but without significantly raising productivity. The fourth input is: how much more efficient are you after using AI? Fifth, how difficult is it for displaced workers to switch jobs—if you lose your job because of AI and need a new one, how hard is it for you to find a new job? Perhaps it depends on training, or it takes time to adjust.

The scenarios highlighted in this paper are only some of many possible scenarios. Depending on how you configure these inputs, you can actually get an outcome that looks a lot like a particular scenario. For example, AI is applied in an automated way but actually has little marginal effect on productivity, leading to displacement while overall gains do not improve much, and workers still need to be reallocated to sectors of the economy that have not felt productivity gains.

The goal of this paper is to promote discussion. Interestingly, after the report came out, some people criticized us for not setting more extreme scenarios, arguing that automation and disruptive innovation should go further; but others criticized us for including substantially extreme scenarios.

On the second point—how AI may differ from previous technologies—I think the key is the breadth of its impact and the speed at which it is being adopted across the entire economy. The U.S. economy is very resilient; we learned a lot about that during the pandemic, and the economy can absorb shocks. Overall, gross job gains and losses in any month are about 10 million, but the net impact—the number everyone pays attention to—may be between 50,000 and 100,000. So the economy can adjust and respond to shocks, but this shock may be larger than the economy can normally accommodate.

And because it will create such enormous reallocation forces, requiring many workers to move from more automated parts of the economy to less automated parts, frictions in the labor market are the fundamental reason unemployment rises in various scenarios, especially in extreme ones.

You know, there is a view that we have centuries of experience with technological progress, farmers have disappeared, and no one can imagine what everyone is doing now, yet everyone still seems to function as they do today. This will be the same.

On the other hand, some argue that you cannot imagine people working again. I think someone even talked about about 50% of white-collar jobs disappearing within one to five years, and he first said that nearly a year ago. Does the data we are seeing now support either of these two views, or is it too early to conclude?

I think, like those ranges of predictions—and this is precisely the motivation for this work—what conditions must be met to achieve extreme scenarios? In extreme cases, GDP growth accelerates to an unprecedented degree, with year-over-year growth of 15%, and overall unemployment rises to 12%. To reach a similar level, you need extremely rapid capability improvement to surpass the vast majority of knowledge work—work that AI systems can do—and it needs to be widely adopted and spread quickly in an automated way.

I think the current data evidence suggests that the pace of capability progress and the pace of automated diffusion are not aligned. But even across three different scenarios, the projected outcomes of the scenarios will not really diverge until next year. That is partly why you released this research—release it now so you can track it against actual data as it emerges and begin to figure out which direction we may be heading.

Furman: To place some of your assumptions within how people think about these issues, your conservative estimate of growth is 2.4% per year. The Federal Open Market Committee, which decides interest rates, publishes forecasts every few months, and the latest number was 2.4%, perhaps the most optimistic over the next four years, even more optimistic than the most optimistic people—their central view is closer to 2%. The Congressional Budget Office, highly respected and trusted for budget analysis and forecasting, bases its analysis of where our fiscal situation is heading on 1.7% growth. You know, 2.4% may not sound very different from 1.7% or 2.1%, but if you work at the Fed or do budget analysis, the difference is enormous. So what you call "moderate" is full of optimism in their eyes.

Now I'm going to the other end—it's not uncommon to hear people in your community talk about 100% annual growth, and I think that may be hard to put into your simulator. So has the Fed made a big mistake on monetary policy? And the CBO too? Does this make us worry more about the fiscal situation than we actually should, or are they just clueless about it?

McCrory: You mentioned looking back at history, and I think that is very valuable. Especially during the long 20th century, from 1870 to 2010, Chad Jones—who is also one of the team members on the scenario explorer—described it as a law of scale for the economy: despite incredible structural adjustment, the U.S. economy still grew at an average rate of 2% over 140 years, accompanied by widespread automation of past work and the emergence of many new types of work.

So I think when participating in this discussion, you need to maintain considerable humility and set the baseline within 2%, which is a very reasonable baseline. That's what we did, and over the past 20 years or so, productivity growth in advanced economies has slowed amid various headwinds, a famous topic widely written about and discussed by many economists in the late 2010s. So I think in that context, even relatively conservative forecasts are an optimistic estimate.

The way I try to understand this is that we did an analysis in November examining how people use Claude and estimating how much time it would take someone to do what Claude is doing if there were no AI. A literature review might take several days, and Claude can do it in five to ten minutes—an astonishing speedup. Using this approach to handle the various things people ask Claude to do, and then using standard macro growth accounting techniques (those interested can look up the relevant literature on Hulten's theorem), we aggregate these task-level efficiency gains to estimate the impact on labor productivity over the next decade—if diffusion takes that long.

Based on current models and current usage patterns, labor productivity growth could increase by 1.8 percentage points. Doesn't that mean... and this seems to take into account some capital deepening assumptions, adding another 1.8 percentage points on top of that?

Furman: Right, otherwise what would happen, probably around 1.5 depending on the situation.

McCrory: Yes, exactly. That would roughly take us back to the prevailing conditions of the late 1990s and early 2000s, and would be more consistent with a moderate, or even slightly above-moderate, scenario. I think I have some reservations about that number, but I think it is a useful framework to help understand how large the potential impact might be based on what we observe among users—users who are trying and learning how to use these tools, and this trying and learning how to use them will spread throughout the economy. If these productivity gains materialize, you might expect progress on that order of magnitude.

In short, you think the Fed and the CBO are somewhat out of touch, because if any of this happens, we will be setting monetary and fiscal policy based on overly conservative forecasts.

Furman: I wouldn't say that. I would say the potential impact on growth and productivity is considerable. Although there are various headwinds, what is the counterfactual? I mean, that 2% change over the past 150 years—perhaps it was precisely the necessary innovation that kept us on that path. So perhaps this is just another series of innovations that keeps us on that trajectory, or perhaps there is some other fundamental constraint preventing us from accelerating further. From this, it can be seen that the constraints of weak links may be very significant.

Chad Jones and Chris Tonetti at Stanford proposed another model in which Moore's Law applies to some extent to all aspects of the economy. Automation is very fast, and even if growth were infinite, with limited time, the economy's growth rate over the next 30 years or so would eventually be about 2.5%, 2.6%, or 6%, because other things that cannot be automated would hinder growth.

McCrory: Exactly. When something is infinite, it becomes very large in both periods, and regardless of its price decline, its importance in growth accounting diminishes and is overshadowed by everything else.

Furman: Yes. I saw Ben Jones at Northwestern describe it this way: accept this model, where these weak links combine to some extent, and then you find that just half the economy can achieve infinite growth, but it will only double the overall size of the economy. The things you fail to automate will greatly constrain the overall impact on the economy. We can see evidence of this in some emerging literature on AI's impact.

Earlier this year, some researchers published a paper examining the introduction of coding agents such as Claude Code—the number of lines of code generated increased by about 20 times, but software output or software releases increased by only 30%. That gap is quite large.

McCrory: Yes. It sounds like you were about to say something harsh about colleagues who think their company can grow 100% a year but actually cannot. Their view of capabilities is no different from yours; they just know less about how the economy works—how to translate those capabilities into growth. I think my baseline looks closer to the typical economist's view, but my margin of error is large.

So I want to understand the root of the disagreement between me and those who have been following this technology. The scenario explorer is one way for me to understand this issue. My feeling after talking with people inside the lab is that the disagreement is mainly about inputs related to capability, adoption, and automation, and less about the underlying economic mechanisms that translate those inputs into output.

Furman: Okay. Next I want to ask you about inequality—there are two types. One is inequality within labor income; for example, if managers earn far more than frontline workers while frontline workers earn the same, that is inequality. The other is inequality between capital and labor—if shareholders get higher returns while others do not, and by the way, richer people own more stocks, this worsens inequality.

Let's first talk about labor inequality. What do you predict AI will do to labor inequality? How uncertain is that effect—both in direction and magnitude?

McCrory: I am uncertain about both direction and magnitude, which is the most honest answer to all such questions.

Furman: Yes, that's a great economist's answer.

McCrory: Here's my answer: there are some excellent papers based on randomized controlled trials showing that AI can help people who are more like novices become more like experts. But when you look at data on actual adopters, you find the main adopters are high-paid, high-skilled talent. From that perspective, I expect this will be a general phenomenon at the income level, or rather inequality will worsen rather than narrow.

But the reason I am uncertain about this is that part of what AI does is not just speed things up or help you complete something faster; it also broadens the range of things you can do. In our survey of about 81,000 people worldwide, asking them various questions about AI, when people talk about productivity gains from AI, they more often mention "expanding my range of abilities and letting me do things I couldn't do before." This may be an equalizing force, compressing the income distribution by forcing you to acquire specific expertise complementary to you.

Furman: Yes, maybe I'll stop there. Alex Imms, who holds a similar role at DeepMind and is a very smart, very creative economist, has been talking about how this can help less-skilled people improve their skills, and he speaks endlessly about the ways he uses this technology—extremely insightful and visionary. As far as I know, it can amplify your talent tenfold, and in ways others cannot imagine.

So do you think this situation—"bad writers become good writers, while great writers get very little help"—is a more common phenomenon than the alternative?

McCrory: I think two forces exist at the same time. One is 10x amplification—software engineers can become 100x software engineers. I think there should be some anecdotal evidence for this, and our research also provides some suggestive evidence based on interactions with domain expertise pointing in this direction. For example, if Claude Code can be used effectively, then the second question is whether 10x engineers are more likely than others to adopt AI—even if the number of engineers doubles and the productivity gain is the same, if adoption rates differ because of pre-existing differences in skills and expertise, then expertise can become a powerful force continuously widening labor market inequality.

Furman: Now let's talk about inequality between labor income and capital income. The data here are very clear—the unemployment rate will not reflect it, productivity growth will not reflect it, but the stock market already reflects it. When I study economic models, including each of yours, and several options I tried, they all show higher returns to capital, but a declining share of income going to labor and a rising share going to capital. Given that people who own capital are already better off, all of this means increasing inequality. Is this the correct expectation for future data?

McCrory: That's a great question. It's worth mentioning that being able to discuss economic models at a forum, with record participation in economic discussions, is an honor, and I hope everyone enjoys it and gains something from these discussions; I myself have certainly learned a lot from these questions.

In the model, what determines the return to capital is the elasticity of capital. In the standard model, you might think capital is infinitely elastic, in which case all gains would go to labor, and you would not see a decline in the labor income share. I warned you earlier that this would be like a conversation between the two of us, and we would pretend you are not here.

This material has infinite elasticity, it is hard to obtain the needed factories, and expansion is equally difficult. To meet this enormous investment demand, sufficient capital must be provided. If this adjustment is harder to make, it creates a stronger downward trend, putting pressure on the labor income share. In extreme cases, the labor income share falls by about 15 percentage points relative to the previous 60% share of every $1 of income.

Next I want to talk first about public policy, and then answer everyone's questions.

Think about a range of policy issues. One is that these models could create weapons that kill everyone. One solution is to tell frontier labs: you have no right to do this; another is to let them go ahead, and the government gives everyone a gas mask.

A trickier problem than weapons is tolerating student cheating, and the solutions to this problem leave an uncomfortable margin. One solution is to make regulations telling labs they cannot help students cheat or answer their questions. Another is to switch to in-class exams. I have reservations about the first—because it is simply impossible to determine who is asking the question and for what purpose. So the only realistic solution is for universities to solve this internally by switching to in-class exams or other methods.

Speaking of employment, where does it fall on this continuum? The disruption you cause is inevitable, and policymakers need to be responsible for cleaning up afterward, which can then be addressed through universal basic income, training programs, wage subsidies, or other means. Conversely, you can actually intervene proactively and guide these models in a certain direction—so that they do not replace people but augment them; not train them to be like people, but train them to help people. Perhaps regulation can push you to do this.

So, regarding the issue we have been discussing—who cleans up the mess?

I think this can be answered from several angles. First, recognize that AI's economic impact is highly uncertain overall, and what we can do and have already done deserves serious consideration as well. At Anthropic, my team and I have a responsibility to produce data and research that can help us understand how this technology affects economic production and labor. This effort aims to help people understand which skills may become increasingly important and which may become less relevant.

More importantly, it is about empowering society to make these choices, rather than deciding unilaterally ourselves. Therefore, the impact of this technology depends both on the direction of capability progress and on the choices we make as a society. My goal on the team is to help society make better decisions.

As for our team's actual work, we mainly and historically focus on data generation and research. I have also written some related content earlier this year on Twitter—I want to use the tools of economics to help us understand the impact of our own decisions, so as to more clearly grasp the tradeoffs between public interest and commercial interest that may be implicit in the decisions we make.

In the language of the economics literature, one approach is: can "directed technical change" be put into practice? I'm actually not sure whether this is feasible, but my team has the opportunity to explore this possibility. Directed technical change refers to guiding technological development to augment labor rather than directly replace it.

Someone asked: what is "augmentation" and what is "replacement"? If you were my research assistant, does this replace you or augment you? I think in this example, the answer is that it augments you... or rather, it replaces workers, just as it replaces your research assistant.

This is actually a good example of the kind of shock AI may cause in the scenarios we discuss. And even at the task level, it is not always clear—whether Claude's role is merely to automate the tasks people give it, or to truly change the nature of work.

The situation you describe is that you are being augmented, which changes the return to certain types of expertise and the complementarity with the skills you have. Historically, new technologies have automated and replaced some work, created new forms of work, and also driven the reintegration of job content. This question remains unclear today. I think this is a genuine puzzle that deserves serious understanding.

Do we have any influence over this? There may be many scenarios. First, governments may set rules prohibiting technology products from doing certain things. Second, companies may take this on voluntarily—if Anthropic takes it on itself, can we ensure other labs follow? In addition, our products are used by many others to develop various products, so the impact will be broader. Third, if we do not fully understand the situation, should we keep trying? Usually, when we do not understand something, we tend to avoid regulation and let markets and dispersed knowledge solve it on their own.

For example, when we do not know whether improving one link will create more jobs elsewhere, we face a dynamic process, and hasty intervention may backfire. So how should ignorance and uncertainty be incorporated into regulation in the economic field? This is different from regulation in the safety field. I think if regulators themselves do not know what they are doing, they are very likely to mess things up.

In a sense, the data and research we try to produce are precisely intended to answer these questions more clearly, perhaps by examining the impact of our own decisions. This type of research ultimately produced some economic indices—for example, the Stanford Digital Economy Lab used related indices to study the employment trajectories of young workers. They documented that young workers associated with automated use on our platform had worse employment prospects. I am not fully convinced by that causal interpretation, but it does show that these data can be used to clarify various contributing factors.

On the policy response, I generally agree: policies should be designed to address this uncertainty and adapt as uncertainty changes. On employment, I think the fundamental tradeoff policymakers need to address is: are we facing a situation of low unemployment, high occupational mobility, and more frequent and severe transitions, or a recession-level unemployment rate accompanied by rapid economic growth?

I think a good insight from the Yale Budget Lab is to measure AI's impact by turnover rates. If you focus on protecting workers' rights, you may be able to facilitate that uncertain transition period, but the overall scale of change may be too large. This may be one reason to incentivize firms to retain employees through retention subsidies, retraining subsidies, and similar measures.

But there is still a lot of uncertainty about what policy mix is reasonable. Given this uncertainty, humility about the effectiveness of various options is necessary. We hope the data can greatly help us construct various scenarios.

Now we move to the Q&A session. Because there are many people here, I will take three questions at a time and then answer them together. Please keep your questions as brief as possible and introduce yourself first.

Question one (Diego Sarmento, second-year Harvard Law School student from Southern California): My question concerns the philosophy of government public policy and AI's impact on the military. Earlier this year, the United States used AI-assisted strike methods, and a recent Pentagon investigation found that AI was used to determine strike targets through a ride-sharing model. My question is: as a member of a company committed to safety, do you endorse the current "human in the loop" requirement? Or do you think a higher standard should be set, such as directly prohibiting AI from being used against any target?

Question two (Liam Damji, Kennedy School MBA student): My question is more from the perspective of competition and market forces. As we said earlier, there are currently large monopolies, and the only thing hindering broad AI adoption is information silos from organizational and data perspectives. Do you think AI will intensify market power concentration among large firms, or disperse the competitive landscape through agile AI-native firms entering the market for the first time?

Question three (Akquilan): I have spent a lot of time thinking about pricing after machine learning model training. You talked a lot about macroeconomic models, but I am curious about how traditional macro and microeconomic models may change when we think about the question of "the pricing of intelligence."

First, one reason I joined Anthropic is that I believe this company places great importance on clearly identifying the risks and benefits of this technology and is committed—even at a cost—to reducing risk as much as possible. I do not know every specific case, but I am grateful to work at a company willing to take these risks seriously. Truly, we must do this well. This is also an expression of the so-called weak-link model—to some extent, delaying the realization of economic benefits may also mean bearing some risks of misuse and malicious use earlier. I will do my best to help us address these risks and benefits at the economic level, ensuring that the benefits are widely shared, risks are minimized, and unfair distribution is avoided. Thank you for the question.

On the question of firm size and AI's impact, I think this is an open question. On the one hand, AI seems to increase the benefits of economies of scale—the more you do, the more information you can manage centrally, allowing you to remain efficient across an increasingly broad range of activities. On the other hand, barriers to entry are much lower—it is much easier to start a one-person company, this model scales very quickly, and one person can access everything needed to run a company, including website development, accounting, human resources, and more.

Rem Koning and her collaborators provide some evidence: they studied the organizational structure of AI-native companies and found that these companies operate with fewer people at a given scale. The situation is still unclear. The data show a large increase in the number of startups, which may be related to AI adoption.

However, one point is worth noting: this macro-level change happened right at the beginning of the pandemic, when business activity fluctuated sharply, new business formation surged, and labor productivity soared—and these changes were clearly not entirely AI-driven. How much came from AI and how much came from remote work and new organizational structures remains an open question.

On the question of pricing intelligence, one thing is clear: the cost of a given level of intelligence is falling sharply. But what the marginal benefit of that intelligence is remains unclear. This is a truly open question, and I do not have a definitive answer right now.

Second round of questions

Question one (Sonia Freire Pearson, joint business school/Kennedy School student): You talked a bit about the paper you recently wrote, and I want to know, in those less catastrophic scenarios, how do you account in economic models for truly strange and hard-to-predict economic consequences—such as large numbers of AI agents operating on the internet in ways where we cannot determine the causes and effects?

Question two (Alex, public policy student): I am very interested in taxation. Anthropic leadership has proposed a new tax system. How do you think such a system would work in the United States and other countries? For example, how should entities that contribute to training data be taxed?

Question three (on the report's forecasts and robotics): Your report forecasts up to 2030, with examples of physical tasks AI currently cannot complete. I am curious about your views on future AI-human cooperation and the autonomization of robotics—which jobs will be replaced and which will not?

First, I think the focus of our research direction is to take seriously the broad and truly strange uncertainties. I like to think about the future in terms of various possible "singularities" and try to understand the economic forces shaping the future.

One of these singularities is the "software singularity"—these AI systems are getting better and better, and recent evidence released by Anthropic shows that these models are getting better at self-improvement, a process sometimes called "recursive self-improvement," whose economic impact deserves in-depth discussion.

The second is the "economic singularity"—what forces determine whether the economy can achieve infinite growth within a finite time? This is unprecedented. If you take into account the degree of innovation automation represented by these models, you get scenarios that are otherwise standard models. So where does the standard model fail? Where is our understanding of innovation off?

The third is the "Coasean singularity"—when more and more agents can take complex actions on your behalf and negotiate with counterparties, transaction costs fall sharply in some areas, and how will firm boundaries and the structure of economic exchange change? We did some exploratory research on this. In the spring we released an experiment called the "Transactions Project," in which Claude interviewed people, asking what items in their homes they would be willing to sell or might be interested in buying, and then built a centralized market for agents to communicate with each other and complete transactions.

This experiment produced some interesting findings: more powerful models systematically extracted more rents when negotiating with weaker models; people did not realize this asymmetry in bargaining power; people's underlying preferences were not fully expressed in interviews but were reflected in the items the models chose to buy for them—for example, someone's AI agent bought a snowboard he already owned, even though he had not explicitly said he liked skiing. This raises important questions about who can access frontier models and how this affects bargaining power, with financial and economic consequences.

On tax policy, this returns to the issue of labor income share versus capital income share. Currently, the tax structure in the United States and many countries focuses on taxing labor. But if the labor income share falls, other tax systems need to be considered. One standard model is taxing consumption, which differs from taxing households' labor supply decisions; taxing consumption may be one response, but more research is needed to consider other potential policy responses.

On the "token tax," my short answer is: in the economic policy framework, I see a token tax as an incentive mechanism to curb excessive adoption. Martin Beraja at Berkeley wrote a great paper discussing the problem that firms do not internalize the externalities of automation—in a world with financial constraints, people find it difficult to mitigate unemployment shocks, so taxing automation has not only fairness reasons but also efficiency considerations. A token tax can be compared to a carbon tax or a tobacco tax.

On robotics and the 2030 forecast, this is precisely why we set the model or scenario in 2030—at least for the next few years, uncertainty is large enough that we want to focus our efforts within this time frame. But in principle, this is a framework that allows us to incorporate robotics considerations in future iterations, and modeling robotics should be regarded as a very high-priority issue. Of course, we must also distinguish the physical accumulation of capital from the universality and availability of non-physical intelligence. This is a very good question.

Final question

Questioner (Vanessa, economics and education major, graduated this May): Regarding access, I would really like to hear your views on the following issue: AI will change the skills valued in the labor market, thereby affecting inequality, but schools are not equally prepared to help students transition smoothly.

This is a very, very important and unresolved question involving skill acquisition, access to human capital, and impacts on human development. We have relatively little research of our own on this, but we can indeed see some clues in our data.

In our research we try to classify the main uses of the product—for example, someone searching for sleep advice at 3 a.m. is mainly personal use; or use for work, or use for learning. We find that learning use cases are actually more common in low-income countries. Therefore, I think access to human capital is actually of global significance: on the one hand, it gives people the ability to acquire knowledge and expertise they may not have had before; on the other hand, there are concerns—for example, students using these tools to cheat, obtaining signals of success without truly internalizing and developing human capital and cognitive endurance. And this cognitive endurance is crucial for getting through such transitions and is very valuable. Thank you.

Finally, I just want to say a few words about this conversation. First, there are no obvious errors in what you said, so your understanding of this topic should be in the top 10%. Second, you approach this with great humility and uncertainty, which I think is enough to put you in the top 10% in this field.

I also want to use this to remind many students here: we still have a lot to learn, and to truly understand all of this, we need to implement some measures now—collecting data, tracking data, and making forecasts. I hope that when we invite you back next time, you can tell us which of your predictions were wrong and why. Admitting mistakes is itself part of humility and part of the learning process. I think you have really set an example for us, and this is something all of us need to learn. For that, I and all of us thank you.