AI and the Future of Economic Growth: A Tale of Two Scenarios
For the past fifteen years, my research has centered on a single, deceptively simple question: what makes
economies grow? Lately, that question has collided head-on with the rise of artificial intelligence. Like many
of you, I find myself thinking about this every day—what kind of world AI is building for us and for our
children. This piece draws on several research papers I have worked on over the last couple of years, and I want
to walk you through a framework that has helped me think more clearly about where we might be headed.
AI is likely to be the most transformative technology of our lifetime. But it is also the latest in a long line
of transformative technologies—electricity, the transistor, semiconductors, information technology, the
internet. A question I keep returning to is this: to what extent is AI different, and to what extent does it
share features with those earlier breakthroughs? The difference, in my view, comes down to one unsettling
possibility. What if machines—AI for cognitive work, and AI running robots for physical work—can eventually
perform every task a human can do? What does it look like to live in that world?
To start, let me lay out two extreme scenarios. Neither is likely to unfold exactly as described; they are
caricatures that help us learn. But somewhere between them lies our actual future.
The Two Extreme Scenarios
Scenario One: AI Dramatically Accelerates Growth
This is the boom scenario you read about almost daily in Silicon Valley. The luminaries of AI—Dario Amodei, Sam
Altman, Demis Hassabis, Geoff Hinton—have been telling us for a decade that these capabilities are coming, and
we have been marching along roughly the schedule they laid out.
The first chapter is AI automating software engineering. Back when Claude Opus 4.5 was released, Anthropic took
the same two-hour take-home exam they give to software engineering candidates and handed it to the model. It
scored higher than any human in history. That was months ago, and the models have only gotten better. In the
next decade—maybe even the next few weeks—it is plausible that AI agents will be able to automate most coding.
Once you have AI agents that can do everything a software engineer can do, you put them to work on AI research
itself. You have them build better algorithms, improve the AI, and create agents that can use a computer the way
a human can. Shortly after that, we could have agents functioning as virtual remote workers—anything you could
ask a colleague on a Zoom call, you could ask an AI instead. Scale that up across millions of GPUs, and you end
up with billions of virtual research assistants, each running perhaps a hundred times faster than we do. Dario
Amodei called this a "country of geniuses in a data center."
You put those geniuses to work discovering new ideas. They design better computer chips, simulate the real world
to engineer better robots, develop new pharmaceuticals. Then, once the robots are designed in virtual reality
and tested in the physical world, you have automated physical tasks too. In the growth models I have studied and
taught for years, once you automate both cognitive and physical tasks, growth explodes. This scenario is
entirely plausible. The question is the horizon—does it happen in three years, five years, or twenty-five? When
growth is accelerating that fast, the distinction almost ceases to matter; the world transforms regardless.
Scenario Two: AI as Business as Usual
Now let me give you the opposite extreme. Here is a graph I have shown my students many times. It plots average
living standards in the United States—real income per person—over 150 years on a logarithmic scale. What you see
is remarkable: you never get too far away from a straight line with a slope of 2% per year. Living standards
have risen at roughly 2% annually for a century and a half, plus or minus a little noise.
Real Income per Person (log scale)
|
| *
| *
| *
| *
| *
| *
| *
| *
| *
+---------------------------------------
1870 1900 1930 1960 1990 2020
Trend: ~2% per year, remarkably steady
Stylized representation of U.S. living standards, 1870--2020. The 2% trend line holds
across waves of transformative technologies.
What is extraordinary about this is that during those 150 years, the U.S. economy absorbed electricity, internal
combustion engines, jet airplanes, antibiotics, vacuum tubes, transistors, semiconductors, information
technology, and the internet. Each of these technologies was wildly transformative. Yet the growth rate stayed
at 2%. How is that possible?
The answer lies in the counterfactual. Within any technology class, ideas get harder to find. The steam engine
runs out of steam. If all you had was the steam engine and you never discovered electricity, growth would have
slowed. Each successive transformative technology did not necessarily lift the growth rate above 2%; rather, it
prevented growth from falling below 2%. It allowed the straight line to continue for another fifty years. The
pessimistic—in quotes—scenario for AI is that it becomes the latest technology that lets 2% growth continue for
another fifty years, rather than igniting an explosion.
Economic history also teaches us that these transformations take decades. When factories switched from steam
power to electric motors, they had to be physically reorganized—instead of one central shaft running through the
building, motors could be distributed everywhere. Information technology required spreadsheets, word processors,
databases, SQL. Complementary innovations and production reorganization take time measured in decades. So do not
be too quick to assume transformative technologies bend the growth curve upward immediately.
The Puzzle and the Power of Weak Links
Both scenarios have merit. So where do we land between them? The concept I find most helpful is weak
links. A chain is only as strong as its weakest link, and business success requires completing
many, many tasks successfully. If you want to release a new iPhone, you have to design it, source all the parts,
manufacture it to exacting tolerances, deliver hundreds of millions of units on schedule, handle retail and
advertising. If any one of those tasks falls down, a lot of value gets lost in the short term. The Space Shuttle
Challenger exploded because a $25 rubber O-ring failed. One small part.
Now apply this to the economy. If you have a chain with twenty links and you make seventeen of them incredibly
strong, that helps—but it does not fundamentally change the overall strength, because three weak links remain.
Here is an example I find stunning: in your pocket right now is a computer with roughly a hundred million times
the transistors of the computers available in the 1970s. Yet I am not a hundred million times more productive at
research. Why not? My computer can invert matrices at blazing speed, but I still have to figure out what data to
put into those matrices, what questions to ask, what theory to test. The weak links—the tasks that have not been
automated—bottleneck everything else.
Weak links are also the source of scarcity, and scarcity is what gives rise to high returns in economics. When we
ask what will happen to the income of our children, the question to ask is: what will be scarce?
What Share of GDP Goes to Computers?
Economists have long been infatuated with the question of who gets GDP. For seventy-five years, the split was
remarkably stable: about two-thirds went to labor, one-third to capital. In the last twenty-five years, labor's
share has fallen by roughly 10%, and economists debate whether automation or market power is the cause. But I
want to go deeper. Within capital income, what share is a return to computing power specifically, and how has
that changed over time?
Computers are everywhere, but their price has fallen dramatically. So which effect dominates—quantity going up or
price going down? Here is the data. During the dot-com boom of the 1990s, the share of GDP paid as a return to
computers rose, peaking in 2000 at just under 4.5%. Since 2000, it has fallen by a third, to about 3%. Computers
are indeed everywhere, and yet they command a smaller share of GDP than before. The price decline dominates the
quantity increase. This is exactly what a weak-link model predicts: computers are the most plentiful thing in
the economy, while humans remain scarce, so the return to computers shrinks.
4.5%
Computer share of GDP (peak, 2000)
3.0%
Computer share of GDP (recent)
-33%
Decline since peak
When we worry that AI might automate everything and leave nothing for humans to do, this graph offers at least a
partial counterpoint. Computers are getting less of GDP, not more, even with a hundred million times more
transistors in our pockets.
Building a Model to Test the Scenarios
To study these two scenarios rigorously, my co-authors and I built a model. It features ideas as the source of
long-run growth, production functions for both goods and ideas that involve weak links, and an automation
process that unfolds endogenously over time. We calibrated the model to fit U.S. data going back to the 1950s,
then ran it forward.
Before showing the forward simulations, here is a thought experiment the model lets us perform. What if we had
infinite amounts of software—anything that uses software, pushed to infinity? How much richer would we be? The
answer turns out to be elegant: because of weak links, infinite amounts of some task raises GDP by that task's
share of GDP. Software is about 2% of GDP. So infinite software would make us roughly 2% richer. Only 2%,
because all the other weak links are bottlenecking us. Automating one thing really well is not enough; you need
to keep automating the weak links.
What the Simulations Reveal
Our model has two key ingredients that mirror the two scenarios. First, automation generates new ideas, which
enable more automation—a flywheel with positive feedback that wants to explode. Second, weak links tell you that
automating some tasks while leaving others untouched means the chain is still held back by its weakest points.
We put both ingredients in, calibrated to history, and ran the model forward under different assumptions.
In the first set of simulations, AI is simply a continuation of the historical patterns of automation we have
seen for 200 years. In the second, more aggressive set, we assume AI represents a break from the past—the entire
economy starts getting automated at the pace of Moore's Law, with machines improving at 10% per year across the
board starting today, and that automation feeds back into more ideas and more automation. I think of the second
scenario as too aggressive and the first as probably not aggressive enough; the truth likely lies somewhere in
between.
| Assumption |
Scenario A: Business as Usual |
Scenario B: Moore's Law Everywhere |
| Automation pace |
Continuation of historical trends (~3% per year for aggregate economy) |
Moore's Law pace everywhere (~10% per year) starting today |
| Growth rate by 2050 |
~2.3% per year |
Above 25% per year |
| Income gain by 2050 vs. baseline |
~4% richer |
Substantially richer; ~50% by 2030 |
| Explosion fully complete |
Centuries |
Around 2060 |
Here is what jumps out. Even in the baseline scenario—where growth looks like the historical 2% line—growth
actually does accelerate. It climbs from 2% to 2.3% by 2050, then 2.6%, then 3%, and eventually settles at an
astonishing 50% per year. But look at the axis: the acceleration is stunningly slow. By 2050, instead of 2%
growth, we are at 2.3%. This is a model that eventually explodes, but it takes a very long time. Why? Weak
links. Just like the computer example: we have a hundred million times the transistors, but we are limited by
everything humans still have to do.
Key insight: Even when growth eventually explodes, the explosion is far slower than the word
"explosion" suggests. In the aggressive Moore's-Law-everywhere scenario, the economy still takes until roughly
2060 before the explosion is fully complete. Weak links act as a powerful brake on the speed of transformation.
Notice something else fascinating about these simulations. The three different futures—full automation, partial
automation with some human-only tasks, and the stable-share baseline—look vastly different 200 years from now.
Yet for the next 75 years, it is remarkably difficult to tell which scenario we are in. The early paths look
similar even when the destinations diverge dramatically.
Jobs, Inequality, and the Radiologist Paradox
In 2016, Geoff Hinton, the Nobel Prize winner and pioneer of deep neural networks, stood up at a conference and
declared, "We should stop training radiologists. In five years, there will be no more radiologists with jobs
because AI will be better than radiologists." He was not wrong about AI becoming better at reading scans—on many
dimensions, that is true today. But here is what actually happened: we have more radiologists now than in 2016,
and they are paid more.
Why? Again, weak links. Jobs are bundles of tasks—perhaps a hundred different things you do in your role. When AI
automates seventy-five of those tasks, the remaining twenty-five become the scarce, high-return activities. The
radiologist can now consult with an AI model that helps detect cancers and other problems, making them more
valuable and more productive. They are still needed to consult on surgeries, talk with other doctors, and
double-check the hardest scans. Automating most tasks can actually raise wages for the humans who handle the
rest.
On the other hand, if you are betting on Uber drivers still being around ten years from now, I think there is a
real chance they will not be. Waymo and other self-driving systems are automating essentially everything an Uber
driver does. But notice how long even this has taken. The DARPA self-driving car competition in 2004 had no
winner; no team completed the course. Stanford won in 2005. That was over twenty years ago. Today, you can hail
a self-driving car in San Francisco, but they remain rare outside the Bay Area, and even there, they are not yet
common. Things take a lot longer than you think.
Meaning in a World of Abundance
Historically, labor is the main asset people trade to get consumption. What happens when machines can do things
better than you? That is a valid worry about the value of your labor. The optimistic take is that the world
where AI changes everything is a world where GDP is incredibly high—a world of abundance. There is plenty to go
around. Rich countries already engage in significant redistribution, and in our simulations, even keeping
current U.S. redistribution programs in place, the consumption of the bottom 10% likely goes up. There is a
genuine chance to make everyone better off.
But I also think about meaning. I use AI models to help with my research constantly now. GPT-5.2 Pro was already
as good as me at math; more recent versions are far better. How long before AI writes better growth papers than
I do? Half of my life's meaning comes from developing growth models. What happens when the AI is better at it?
"What will we do when AI outperforms us at the work that gives our lives meaning?"
The analogy I return to is retirement. When we look at retirees, we do not say they have lost meaning. They seem
happy. They live in a world of abundance, go on cruises, see friends, go dancing. Summer camp is another analogy
I like—making pottery, singing songs, getting together with colleagues while the AI teaches us the latest growth
model. That might be my version of summer camp in an AI-abundant future.
The Downside Risks Are Real
Despite the optimistic notes, I am very nervous about our future. The reason is catastrophic risk, and I think we
need to discuss this honestly, without the pejorative criticism that sometimes accompanies the conversation.
There are two versions people talk about. The first is the bad-actor problem. Imagine a hacker in North Korea—or
anywhere—with access to a jailbroken version of a future frontier model. These models are jailbroken the day
they come out. By GPT-8 or Opus 7, these systems will be able to do anything the smartest humans can do. If it
is possible to design a virus more lethal than Ebola that takes three months to display symptoms, the AI will
figure it out. We got through the nuclear age, so far, because nuclear weapons were rare—only a handful of
people had red buttons that could do serious damage. If eight billion people have access to the red button, can
we make sure no one pushes it?
The second version is more speculative but worth contemplating. Stuart Russell, the computer scientist from
Berkeley, put it starkly: how do we retain power over entities more powerful than us forever? Imagine we learned
tonight that a spaceship is on its way from Pluto toward Earth. We would be excited at first, and then we would
remember that when advanced societies encounter less advanced ones in our history, it has not gone well for the
less advanced.
Here is where the weak-link view takes a darker turn. A chain is only as strong as its weakest link. That means
improvements come slowly—you have to strengthen every link to get the full benefit. But it also means the system
is fragile on the downside. Break one link, and all the value can be lost. Consider Mythos, the model
Anthropic did not release publicly but described as discovering bugs in twenty-five-year-old, battle-tested
software that humans had never found—thousands of them. In six months or a year, there may be an open-source
version anyone can use. How sure are we that a bad actor will not use it to hack the electric grid? The
financial system? The banking system? Zero out everyone's bank balance? That is not an existential problem, but
it is a huge problem—and one we have a real chance of facing in the next three years. The weak-link view says
the benefits come slowly, but the downside risk can arrive very fast.
Conclusion
- Two extremes frame the debate. One scenario has AI igniting explosive economic growth; the
other treats AI as the latest transformative technology that simply allows 2% growth to continue, preventing
decline rather than sparking an explosion.
- Weak links are the central mechanism. A chain is only as strong as its weakest link.
Automating individual tasks yields limited gains until the remaining bottlenecks are also addressed. This
explains why infinite software would make us only 2% richer and why computers command a shrinking share of
GDP despite proliferating everywhere.
- Growth probably does accelerate, but slowly. Even in aggressive simulations where AI
diffuses at Moore's Law pace across the entire economy, the explosion takes decades—reaching full fruition
around 2060 rather than in three or five years. Weak links act as a powerful brake.
- Jobs are bundles of tasks. The radiologist example shows that automating most tasks in a
profession can raise wages for the humans who handle the remaining, scarce, high-value tasks. But
professions where every task can be automated—like driving—face genuine displacement risk.
- Abundance creates possibilities for redistribution. In a world of soaring GDP, there is
enough for everyone to be better off. Whether that happens depends on political economy and social choices
that are far from guaranteed.
- Meaning is a separate challenge. Even in abundance, people derive identity and purpose from
work. The retirement and summer-camp analogies suggest humans can adapt, but the transition deserves serious
thought.
- Catastrophic risk is the most urgent concern. The weak-link logic cuts both ways:
improvements are slow, but breaking one critical link—through a bad actor with a powerful model, for
instance—can cause rapid, severe damage. We should use the intervening years to prepare for these risks,
including labor market disruption, inequality, political economy challenges, and catastrophic scenarios.
- AI will be worth multiple internets. Between 2015 and 2045, AI is likely to change the
world more than the internet did between 1990 and 2020. The transformation will probably take longer than
the most excited voices predict, but that does not make it any less profound.
