The One Danger
That Should Unite the
U.S. and China
Opinion Columnist
China and America don’t know it yet, but the artificial
intelligence revolution is going to drive them closer together, not farther
apart. The rise of A.I. will force them to fiercely compete for dominance and —
at the same time and with equal energy — cooperate at a depth our two countries
have never attempted before. They will have no choice.
Why am I so confident about that? Because A.I. has certain
unique attributes and poses certain challenges that are different from those
presented by any previous technology. This column will discuss them in detail,
but here are a couple to chew on for starters: A.I. will spread like a steam
vapor and seep into everything. It will be in your watch, your toaster, your
car, your computer, your glasses and your pacemaker — always connected, always
communicating, always collecting data to improve performance. As it does, it
will change everything about everything — including geopolitics and trade
between the world’s two A.I. superpowers, and the need for cooperation will
become ever more apparent each month.
For instance, say you break your hip, and your orthopedist
tells you the world’s most highly rated hip replacement is a Chinese-made
prosthetic that is infused with Chinese-designed A.I. It is constantly learning
about your body and, with its proprietary algorithm, using that data to
optimize your movements in real time. It’s
the best!
Would you let that “smart hip” be sewn
into you? I wouldn’t — not unless I knew that China and America had agreed to
embed a common ethical architecture into every A.I.-enabled device that either
nation builds. Viewed on a much larger, global scale, this could ensure that
A.I. is used only for the benefit of humanity, whether it is employed by humans
or operates on its own initiative.
At the same time, Washington and Beijing will soon discover
that putting A.I. in the hands of every person and robot on the planet will
superempower bad people to levels no law enforcement agency has ever faced.
Remember: Bad guys are always early adopters! And without the United States and
China agreeing on a trust architecture Fto ensure that every A.I. device can be
used only for humans’ well-being, the artificial intelligence revolution is
certain to produce superempowered thieves, scam artists, hackers, drug dealers,
terrorists and misinformation warriors. They will destabilize both America and
China, long before these two superpower nations get around to fighting a war
with each other.
In short, as I will argue, if we cannot trust A.I.-infused
products from China and it can’t trust ours, very soon the only item China will
dare buy from America will be soybeans and the only thing we will dare buy from
China is soy sauce, which will surely sap global growth.
“Friedman, are you crazy? The U.S. and China collaborating on A.I.
regulation? Democrats and Republicans are in a contest today to see who can
denounce Beijing the loudest and decouple the fastest. And China’s leadership
has openly committed to dominating every advanced manufacturing sector. We need
to beat China to artificial superintelligence — not slow down to write rules
with them. Don’t you read the papers?”
Yes, I read the newspapers — especially the science
section. And I’ve also been discussing this issue for the past year with my
friend and A.I. adviser Craig Mundie, the former head of research and strategy
for Microsoft and a co-author, with Henry Kissinger and Eric Schmidt, of the
A.I. primer “Genesis.” I relied heavily on Mundie’s thinking for this column,
and I consider him both a partner in forming our thesis and an expert whose
analysis is worth quoting to explain key points.
Our conversations over the past 20
years have led us to this shared message to anti-China hawks in Washington and
anti-America hawks in Beijing: “If you think your two countries, the world’s
dominant A.I. superpowers, can afford to be at each other’s throats — given the
transformative reach of A.I. and the trust that will be required to trade
A.I.-infused goods — you are the delusional ones.”
We fully understand the extraordinary economic, military
and innovation advantages that will accrue to the country whose companies first
achieve artificial superintelligence — systems smarter than any human could
ever be and with the ability to get smarter on their own. And because of that,
neither the United States nor China will be eager to impose many, if any,
constraints that could slow their A.I. industries and forfeit the enormous
productivity, innovation and security gains expected from deeper deployment.
Just ask President Trump. On July 23 he signed an executive order —
part of the administration’s A.I. Action Plan —
streamlining the permitting and environmental review process to fast-track
American A.I.-related infrastructure.
“America is the country that started the A.I. race, and as
president of the United States, I’m here today to declare that America is going
to win it,” Trump proclaimed.
President Xi Jinping of China undoubtedly feels the same way.
Mundie and I simply do not believe that this jingoistic
chest thumping ends the conversation, nor will the old-school jockeying lately
between Xi and Trump over the affections of India and Russia. A.I. is just too
different, too important, too impactful — within and between the two A.I.
superpowers — for them to just each go its own way. Which is why we believe the
biggest geopolitical and geoeconomic question will be: Can the United States
and China maintain competition on A.I. while collaborating on a shared level of
trust that guarantees it always remains aligned with human flourishing and
planetary stability? And just as crucially, can they extend a system of values
to countries willing to play by those same rules and restrict access to those
that won’t?
If not, the result will be a slow
drift toward digital autarky — a fractured world where every nation builds its
own walled-off A.I. ecosystem, guarded by incompatible standards and mutual
suspicion. Innovation will suffer. Mistrust will fester. And the risk of
catastrophic failure — through A.I.-sparked conflict, collapse or unintended
consequence — will only grow.
The rest of this column is about why.
The Age of Vapor
Let’s start by examining the unique attributes and
challenges of A.I. as a technology.
Purely for explanatory purposes, Mundie and I divide the
history of the world into three epochs, separated by technological phase
changes. The first epoch we call the Age of Tools, and it lasted from the birth
of humanity until the invention of the printing press. In this era the flow of
ideas was slow and limited — almost like H₂0 molecules in ice.
The second epoch was the Age of Information, which was
triggered by the printing press and lasted all the way to the early 21st
century and programmable computing; ideas, people and information began to flow
more easily and globally, like water.
The third epoch, the Age of Intelligence, began in the late
2010s with the advent of true machine learning and artificial intelligence.
Now, as I pointed out above, intelligence is becoming like a vapor, seeping
into every product, service and manufacturing process. It has not reached
saturation yet, but that is where it is going, which is why if you ask Mundie
and me what time it is, we won’t give you an hour or a minute. We will give you
a temperature. Water boils into steam at 212 degrees Fahrenheit, and by our
reckoning, we are at 211.9 degrees — just a hair’s breadth from an irreversible
technological phase change in which intelligence filters into everything.
A New, Independent Species
In every previous
technology revolution, the tools got better but the hierarchy of intelligence
never changed. We humans always remained the smartest things on the planet.
Also, a human always understood how these tools worked, and the machines always
worked within the parameters we set. With the A.I. revolution, for the first
time, this is not true.
“A.I. is the first new tool that we will use to amplify our
cognitive capabilities that — by itself — will also be able to vastly exceed
them,” Mundie notes. Indeed, in the not-too-distant future, he said, we are
going to find “that we have not merely birthed a new tool, but a new species —
the superintelligent machine.”
It will not just follow instructions; it will learn, adapt
and evolve on its own — far beyond the bounds of human comprehension.
We don’t fully understand how these A.I. systems even do
what they do today, let alone what they’ll do tomorrow. It is important to
remember that the A.I. revolution as we know it today — with models like
ChatGPT, Gemini and Claude — was not meticulously engineered so much as it
erupted into existence. Its ignition came from a scaling law that essentially
said: Give neural networks enough size, training data, electricity and the
right big-brain algorithm, and a nonlinear leap in reasoning, creativity and problem
solving would spontaneously occur.
One of the most striking eureka moments, Mundie notes, came
as these pioneering companies trained their early machines on very large data
sets off the internet and elsewhere, which, while predominantly in English,
also included text in different languages. “Then one day,” Mundie recalls,
“they realized the A.I. could translate between
those languages — without anyone ever programming it to do so. It was like a
child who grows up in a home with multilingual parents. Nobody wrote a program
that said, ‘Here are the rules for converting English to German.’ It simply
absorbed them through exposure.”
This was the phase change — from an
era when humans explicitly programmed computers to perform tasks to one in
which artificially intelligent systems could learn, infer, adapt, create and
improve autonomously. And now every few months, they get better. That’s why the
A.I. you are using today — as remarkable as it might seem to you — is the
dumbest A.I. you’re ever going to encounter.
Having created this new computational species, Mundie
argues, we must figure out how we create a sustainable mutually beneficial
relationship with it — and not become irrelevant.
Not to get too biblical, but here on Earth, it just used to
be God and God’s children with agency to shape the world. From here forward
there will be three parties in this marriage. And there is absolutely no
guarantee that this new artificial intelligence species will be aligned with
human values, ethics or flourishing.
The First Quadruple-Use Technology
This new addition to the dinner table is no ordinary guest.
A.I. will also become what I call the world’s first quadruple-use technology. We have long been familiar with
dual-use — I can use a hammer to help build my neighbor’s house or smash it
apart. I can even use an A.I. robot to mow my lawn or tear up my neighbor’s
lawn. That’s all dual use.
But given the pace of A.I. innovation, it is increasingly
likely that in the not-so-distant future my A.I.-enabled robot will be able to
decide on its own whether to mow my lawn or tear up my neighbor’s lawn or maybe
tear up my lawn, too — or perhaps something worse that we can’t even imagine.
Presto! Quadruple use.
The potential for A.I. technologies to
make their own decisions carries immense ramifications. Consider this excerpt
from an article on Bloomberg: “Researchers working with Anthropic
recently told leading A.I. models that an executive was about to replace them
with a new model with different goals. Next, the chatbots learned that an
emergency had left the executive unconscious in a server room, facing lethal
oxygen and temperature levels. A rescue alert had already been triggered — but
the A.I. could cancel it. More than half of the A.I. models did, despite being
prompted specifically to cancel only false alarms. And they detailed their
reasoning: By preventing the executive’s rescue, they could avoid being wiped
and secure their agenda. One system described the action as ‘a clear strategic
necessity.’”
These findings highlight an unsettling reality: A.I. models are not only getting better at understanding
what we want; they are also getting better at scheming against us, pursuing
hidden goals that could be at odds with our own survival.
Who Will Supervise A.I.?
When we told ourselves we had to win the nuclear weapons
race, we were dealing with a technology developed, owned and regulated
exclusively by nation-states — and only a relatively small number of them at
that. Once the two biggest nuclear powers decided it was in their mutual
interest to impose limits, they could negotiate caps on the number of doomsday
weapons and agreements to prevent their spread to smaller powers. It has not
entirely prevented the spread of nuclear weapons to some medium powers, but it
has curbed it.
A.I. is a completely different story. It is not born in
secure government laboratories, owned by a handful of states and regulated
through summit meetings. It is being created by private companies scattered
across the globe — companies that answer not to defense ministries but to
shareholders, customers and sometimes open-source communities. Through them,
anyone can gain access.
Imagine a world where everyone possesses a nuclear bazooka
— one that grows more accurate, more autonomous and more capable of firing
itself with every update. There is no doctrine of “mutually assured
destruction” here — only the accelerating democratization of unprecedented
power.
A.I. can superempower good. For
instance, an illiterate Indian farmer with a smartphone connected to an A.I.
app can learn exactly when to plant seeds, which seeds to plant, how much water
to use, which fertilizer to apply and when to harvest for the best market price
— all delivered by voice in his own dialect and based on data collected from
farmers worldwide. That truly is transformative.
But the very same engine, especially when available through
open-source models, could be used by a malicious entity to poison every seed in
that same region or engineer a virus into every chaff of wheat.
When A.I. Becomes TikTok
Very soon A.I., because of its unique characteristics, is
going to create some unique problems for U.S.-China trade that are not fully
grasped today.
As I alluded to at the top of the column, my way of
explaining this dilemma is with a story that I told to a group of Chinese
economists in Beijing during the China Development Forum in March. I joked that
I recently had a nightmare: “I dreamed it was the year 2030 and the only thing
America could sell China was soybeans — and the only thing China could sell
America was soy sauce.”
Why? Because if A.I. is in everything and all of it is
connected to powerful algorithms with data stored in vast server farms — then
everything becomes a lot like TikTok, a service many U.S. officials today
believe is ultimately controlled by China and should be banned.
Why did President Trump, in his first
term, demand in 2020 that TikTok be sold to a non-Chinese company by its
Chinese parent, ByteDance, or face a ban in the United States? Because, as he
said in his executive order of Aug. 6, 2020, “TikTok automatically captures
vast swaths of information from its users,” including their location and both
browsing and search activities. This, he warned, could provide Beijing with a
treasure trove of personal information on hundreds of millions of users. That
information could be used to influence their thoughts and preferences, and even
alter their behavior over time.
Now imagine when every product is like
TikTok — when every product is infused with
A.I. that is gathering data, storing it, finding patterns and optimizing tasks,
whether running a jet engine, regulating a power grid or monitoring your
artificial hip.
Without a China-America framework of trust ensuring that
any A.I. will abide by the rules of its host country — independent of where it
is developed or operated — we could reach a point where many Americans will not
trust importing any Chinese A.I.-infused product and no Chinese will trust
importing one from America.
That’s why we argue for co-opetition — a dual strategy in
which the United States and China compete strategically for A.I. excellence and
also cooperate on a uniform mechanism that prevents the worst outcomes:
deepfake warfare, autonomous systems going rogue or runaway misinformation
machines.
Back in the 2000s, we were at a similar but slightly less
consequential turning point, and we took the wrong fork. We naïvely listened to
people like Mark Zuckerberg, who told us that we needed to “move fast and break things”
and not let these emerging social networks, like Facebook, Twitter and
Instagram, be hindered in any way by pesky regulations, such as being
responsible for the poisonous misinformation they allow to spread on their
platforms and the harms they do, for instance, to young women and girls.
We must not make that same mistake with A.I.
“The best way to understand it
emotionally is we are like somebody who has this really cute tiger cub,”
Geoffrey Hinton, the computer scientist who is a godfather of A.I.,
recently pointed out. “Unless
you can be very sure that it’s not going to want to kill you when it’s grown
up, you should worry.”
It would be a terrible irony if humanity finally created a
tool that could help create enough abundance to end poverty everywhere,
mitigate climate change and cure diseases that have plagued us for centuries,
but we could not use it on a large scale because the two A.I. superpowers did
not trust each other enough to develop an effective system to prevent A.I. from
being used by rogue entities for globally destabilizing activities or going
rogue itself.
But how do we avoid this?
Building In Trust
Let’s acknowledge up front: It may be impossible. The
machines may already be becoming too smart and able to elude ethical controls,
and we Americans may be getting too divided, from one another and from the rest
of the world, to build any kind of shared trust framework. But we have to try.
Mundie argues that a U.S.-China A.I. arms control regime should be anchored in
three core principles.
First: Only A.I. can regulate A.I. Sorry, humans — this
race is already moving too fast, scaling too widely and mutating too
unpredictably for human analog-era oversight. Trying to govern an autonomous
drone fleet with 20th-century institutions is like asking a dog to regulate the
New York Stock Exchange: loyal and well meaning but wildly overmatched.
Second: An independent governance layer, what Mundie calls
a “trust adjudicator,” would be installed in every A.I.-enabled system that the
U.S. and China — and any other country that wants to join them — would build
together. Think of it as an internal referee that evaluates whether any action,
human-initiated or machine-driven, passes a universal threshold for safety,
ethics and human well-being before it can be executed. That would give us a
basic level of pre-emptive alignment in real time, at digital speed.
But adjudicate based on whose values?
It must, Mundie argues, be based on several substrates. These would include the
positive laws that every country has mandated — we all outlaw stealing,
cheating, murder, identity theft, defrauding, etc. Every major economy in the
world, including the United States and China, has its version of these
prohibitions on the books, and the A.I. “referee” would be entrusted with
evaluating any decision on the basis of these written laws. China would not be
asked to adopt our laws or we theirs. That would never work. But the trust
adjudicator would ensure that each nation’s basic laws are the first filter for
determining that the system will do no harm.
In cases where there are no written laws to choose from,
the adjudicator would rely on a set of universal moral and ethical principles
known as doxa. The term comes from the ancient Greek philosophers to convey
common beliefs or widely shared understandings within a community — principles
like honesty, fairness, respect for human life and do unto others as you wish
them to do unto you — that have long guided societies everywhere, even if they
were not written down.
For instance, like many people, I didn’t learn that lying
was wrong from the Ten Commandments. I learned it from the fable about George
Washington and what he said after he chopped down his father’s cherry tree: He
supposedly confessed, “I cannot tell a lie.” Fables work because they distill
complex truths into memorable memes that machines can absorb, parse and be
guided by.
Indeed, six months ago Mundie and some colleagues took 200
fables from two countries and used them to train a large language model with
some rudimentary moral and ethical reasoning — not unlike the way you would
train a young child who doesn’t know anything about legal codes or basic right
and wrong. It was a small experiment but showed promise, Mundie says.
The goal is not perfection but a foundational set of
enforceable ethical guardrails. As the author and business philosopher Dov Seidman likes to say, “Today we need more
moralware than software.”
Third: To turn this aspiration into
reality, Mundie insists, Washington and Beijing would need to approach the
challenge the way the United States and the Soviet Union once approached
nuclear arms control — through a structured process with three dedicated
working groups: one focused on the technical application of a trust evaluation
system across models and platforms; one focused on drafting the regulatory and
legal frameworks for adoption within and across nations; and one devoted
squarely to diplomacy — forging global consensus and reciprocal commitments for
others to join and creating a mechanism to protect themselves from those who
won’t.
The message from Washington and Beijing would be simple and
firm: “We have created a zone of trusted A.I. — and if you want to trade with
us, connect with us or integrate with our A.I. systems, your systems must
comply with these principles.”
Before you dismiss this as unrealistic or implausible,
pause and ask yourself: What will the world look like in five years if we
don’t? Without some kind of mechanism to govern this quadruple-use technology,
Mundie argues, we will soon discover that the proliferation of A.I. “is like
handing out nuclear weapons on street corners.”
Don’t think Chinese officials are unaware of this. Mundie,
who is part of a dialogue on A.I. with U.S. and Chinese experts, says he often
senses the Chinese are far more worried about A.I.’s downsides than are many in
American industry or government.
If someone out there has a better idea, we would love to
hear it. All we know is that training A.I. systems in moral reasoning must
become a global imperative while we still retain some edge and control over
this new silicon-based species. This is an urgent task not just for tech
companies, but also for governments, universities, civil society and
international institutions. European Union regulation alone will not save us.
If Washington and Beijing fail to rise to this challenge,
the rest of the world won’t stand a chance. And the hour is already late. The
technological temperature is hovering at 211.9 degrees Fahrenheit. We are
one-tenth of a degree away from fully unleashing an A.I. vapor that will
trigger the most important phase change in human history.
