The Turing Trap: The Promise & Peril of - Specialized Research AI at MIT

The Turing Trap: The Promise & Peril of
Human-Like Artificial Intelligence

Erik Brynjolfsson

In 1950, Alan Turing proposed a test of whether a machine was intelligent: could a
machine imitate a human so well that its answers to questions were indistinguish-
able from a human’s? Ever since, creating intelligence that matches human intelli-
gence has implicitly or explicitly been the goal of thousands of researchers, engineers,
and entrepreneurs. The benefits of human-like artificial intelligence (HLAI) include
soaring productivity, increased leisure, and perhaps most profoundly a better under-
standing of our own minds. But not all types of AI are human-like–in fact, many
of the most powerful systems are very different from humans–and an excessive fo-
cus on developing and deploying HLAI can lead us into a trap. As machines become
better substitutes for human labor, workers lose economic and political bargaining
power and become increasingly dependent on those who control the technology. In
contrast, when AI is focused on augmenting humans rather than mimicking them,
humans retain the power to insist on a share of the value created. What is more,
augmentation creates new capabilities and new products and services, ultimately
generating far more value than merely human-like AI. While both types of AI can
be enormously beneficial, there are currently excess incentives for automation rath-
er than augmentation among technologists, business executives, and policy-makers.

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

A lan Turing was far from the first to imagine human-like machines.1 Ac-

cording to legend, 3,500 years ago, Dædalus constructed humanoid stat-
ues that were so lifelike that they moved and spoke by themselves.2 Near-
ly every culture has its own stories of human-like machines, from Yanshi’s leather
man described in the ancient Chinese Liezi text to the bronze Talus of the Argo-
nautica and the towering clay Mokkerkalfe of Norse mythology. The word robot
first appeared in Karel Čapek’s influential play Rossum’s Universal Robots and de-
rives from the Czech word robota, meaning servitude or work. In fact, in the first
drafts of his play, Čapek named them labori until his brother Josef suggested sub-
stituting the word robot.3

Of course, it is one thing to tell tales about humanoid machines. It is some-
thing else to create robots that do real work. For all our ancestors’ inspiring sto-
ries, we are the first generation to build and deploy real robots in large numbers.4

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

272

Dozens of companies are working on robots as human-like, if not more so, as
those described in the ancient texts. One might say that technology has advanced
sufficiently to become indistinguishable from mythology.5

The breakthroughs in robotics depend not merely on more dexterous me-
chanical hands and legs, and more perceptive synthetic eyes and ears, but also on
increasingly human-like artificial intelligence (HLAI). Powerful AI systems are
crossing key thresholds: matching humans in a growing number of fundamental
tasks such as image recognition and speech recognition, with applications from
autonomous vehicles and medical diagnosis to inventory management and prod-
uct recommendations.6

These breakthroughs are both fascinating and exhilarating. They also have
profound economic implications. Just as earlier general-purpose technologies
like the steam engine and electricity catalyzed a restructuring of the economy, our
own economy is increasingly transformed by AI. A good case can be made that AI
is the most general of all general-purpose technologies: after all, if we can solve
the puzzle of intelligence, it would help solve many of the other problems in the
world. And we are making remarkable progress. In the coming decade, machine
intelligence will become increasingly powerful and pervasive. We can expect re-
cord wealth creation as a result.

Replicating human capabilities is valuable not only because of its practical po-
tential for reducing the need for human labor, but also because it can help us build
more robust and flexible forms of intelligence. Whereas domain-specific technol-
ogies can often make rapid progress on narrow tasks, they founder when unex-
pected problems or unusual circumstances arise. That is where human-like intel-
ligence excels. In addition, HLAI could help us understand more about ourselves.
We appreciate and comprehend the human mind better when we work to create
an artificial one.

These are all important opportunities, but in this essay, I will focus on the ways

that HLAI could lead to a realignment of economic and political power.

The distributive effects of AI depend on whether it is primarily used to aug-
ment human labor or automate it. When AI augments human capabilities, en-
abling people to do things they never could before, then humans and machines
are complements. Complementarity implies that people remain indispensable for
value creation and retain bargaining power in labor markets and in political deci-
sion-making. In contrast, when AI replicates and automates existing human ca-
pabilities, machines become better substitutes for human labor and workers lose
economic and political bargaining power. Entrepreneurs and executives who have
access to machines with capabilities that replicate those of humans for a given
task can and often will replace humans in those tasks.

Automation increases productivity. Moreover, there are many tasks that are
dangerous, dull, or dirty, and those are often the first to be automated. As more

273

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

151 (2) Spring 2022Erik Brynjolfsson

tasks are automated, a fully automated economy could, in principle, be structured
to redistribute the benefits from production widely, even to those people who are
no longer strictly necessary for value creation. However, the beneficiaries would
be in a weak bargaining position to prevent a change in the distribution that left
them with little or nothing. Their incomes would depend on the decisions of
those in control of the technology. This opens the door to increased concentra-
tion of wealth and power.

This highlights the promise and the peril of achieving HLAI: building machines
designed to pass the Turing Test and other, more sophisticated metrics of hu-
man-like intelligence.7 On the one hand, it is a path to unprecedented wealth, in-
creased leisure, robust intelligence, and even a better understanding of ourselves.
On the other hand, if HLAI leads machines to automate rather than augment hu-
man labor, it creates the risk of concentrating wealth and power. And with that
concentration comes the peril of being trapped in an equilibrium in which those
without power have no way to improve their outcomes, a situation I call the Turing
Trap.

The grand challenge of the coming era will be to reap the unprecedented ben-
efits of AI, including its human-like manifestations, while avoiding the Turing
Trap. Succeeding in this task requires an understanding of how technological
progress affects productivity and inequality, why the Turing Trap is so tempting
to different groups, and a vision of how we can do better.

A rtificial intelligence pioneer Nils Nilsson noted that “achieving real hu-

man-level AI would necessarily imply that most of the tasks that humans
perform for pay could be automated.”8 In the same article, he called for a
focused effort to create such machines, writing that “achieving human-level AI or
‘strong AI’ remains the ultimate goal for some researchers” and he contrasted this
with “weak AI,” which seeks to “build machines that help humans.”9 Not surpris-
ingly, given these monikers, work toward “strong AI” attracted many of the best
and brightest minds to the quest of–implicitly or explicitly–fully automating
human labor, rather than assisting or augmenting it.

For the purposes of this essay, rather than strong versus weak AI, let us use the
terms automation versus augmentation. In addition, I will use HLAI to mean hu-
man-like artificial intelligence, not human-level AI, because the latter mistakenly
implies that intelligence falls on a single dimension, and perhaps even that hu-
mans are at the apex of that metric. In reality, intelligence is multidimensional: a
1970s pocket calculator surpasses the most intelligent human in some ways (such
as for multiplication), as does a chimpanzee (short-term memory). At the same
time, machines and animals are inferior to human intelligence on myriad other
dimensions. The term “artificial general intelligence” (AGI) is often used as a syn-
onym for HLAI. However, taken literally, it is the union of all types of intelligences,

274

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Dædalus, the Journal of the American Academy of Arts & SciencesThe Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence

able to solve types of problems that are solvable by any existing human, animal, or
machine. That suggests that AGI is not human-like.

The good news is that both automation and augmentation can boost labor pro-
ductivity: that is, the ratio of value-added output to labor-hours worked. As pro-
ductivity increases, so do average incomes and living standards, as do our capa-
bilities for addressing challenges from climate change and poverty to health care
and longevity. Mathematically, if the human labor used for a given output declines
toward zero, then labor productivity would grow to infinity.10

The bad news is that no economic law ensures everyone will share this growing
pie. Although pioneering models of economic growth assumed that technologi-
cal change was neutral,11 in practice, technological change can disproportionately
help or hurt some groups, even if it is beneficial on average.12

In particular, the way the benefits of technology are distributed depends to a
great extent on how the technology is deployed and the economic rules and norms
that govern the equilibrium allocation of goods, services, and incomes. When
technologies automate human labor, they tend to reduce the marginal value of
workers’ contributions, and more of the gains go to the owners, entrepreneurs,
inventors, and architects of the new systems. In contrast, when technologies aug-
ment human capabilities, more of the gains go to human workers.13

A common fallacy is to assume that all or most productivity-enhancing innova-
tions belong in the first category: automation. However, the second category, aug-
mentation, has been far more important throughout most of the past two centuries.
One metric of this is the economic value of an hour of human labor. Its market price
as measured by median wages has grown more than tenfold since 1820.14 An entre-
preneur is willing to pay much more for a worker whose capabilities are amplified
by a bulldozer than one who can only work with a shovel, let alone with bare hands.
In many cases, not only wages but also employment grow with the introduc-
tion of new technologies. With the invention of the airplane, a new job category
was born: pilots. With the invention of jet engines, pilot productivity (in passen-
ger-miles per pilot-hour) grew immensely. Rather than reducing the number of
employed pilots, the technology spurred demand for air travel so much that the
number of pilots grew. Although this pattern is comforting, past performance does
not guarantee future results. Modern technologies–and, more important, the ones
under development–are different from those that were important in the past.15

In recent years, we have seen growing evidence that not only is the labor share
of the economy declining, but even among workers, some groups are beginning
to fall even further behind.16 Over the past forty years, the numbers of million-
aires and billionaires grew while the average real wages for Americans with only
a high school education fell.17 Though many phenomena contributed to this, in-
cluding new patterns of global trade, changes in technology deployment are the
single biggest explanation.

275

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

151 (2) Spring 2022Erik Brynjolfsson

If capital in the form of AI can perform more tasks, those with unique assets,
talents, or skills that are not easily replaced with technology stand to benefit dis-
proportionately.18 The result has been greater wealth concentration.19

Ultimately, a focus on more human-like AI can make technology a better sub-
stitute for the many nonsuperstar workers, driving down their market wages,
even as it amplifies the market power of a few.20 This has created a growing fear
that AI and related advances will lead to a burgeoning class of unemployable or
“zero marginal product” people.21

A s noted above, both automation and augmentation can increase produc-

tivity and wealth. However, an unfettered market is likely to create social-
ly excessive incentives for innovations that automate human labor and
provide too weak incentives for technology that augments humans. The first fun-
damental welfare theorem of economics states that under a particular set of con-
ditions, market prices lead to a pareto optimal outcome: that is, one where no one
can be made better off without making someone else worse off. But we should not
take too much comfort in that. The theorem does not hold when there are innova-
tions that change the production possibilities set or externalities that affect peo-
ple who are not part of the market.22

Both innovations and externalities are of central importance to the econom-
ic effects of AI, since AI is not only an innovation itself, but also one that triggers
cascades of complementary innovations, from new products to new production
systems.23 Furthermore, the effects of AI, particularly on work, are rife with ex-
ternalities. When a worker loses opportunities to earn labor income, the costs go
beyond the newly unemployed to affect many others in their community and in
the broader society. With fading opportunities often come the dark horses of al-
coholism, crime, and opioid abuse. Recently, the United States has experienced
the first decline in life expectancies in its recorded history, a result of increasing
deaths from suicide, drug overdose, and alcoholism, what economists Anne Case
and Angus Deaton call “deaths of despair.”24

This spiral of marginalization can grow because concentration of econom-
ic power often begets concentration of political power. In the words attributed
to Louis Brandeis: “We may have democracy, or we may have wealth concen-
trated in the hands of a few, but we can’t have both.” In contrast, when humans
are indispensable to value creation, economic power will tend to be more decen-
tralized. Historically, most economically valuable knowledge–what economist
Simon Kuznets called “useful knowledge”–resided within human brains.25 But
no human brain can contain even a small fraction of the useful knowledge needed
to run even a medium-sized business, let alone a whole industry or economy, so
knowledge had to be distributed and decentralized.26 The decentralization of use-
ful knowledge, in turn, decentralizes economic and political power.

276

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Dædalus, the Journal of the American Academy of Arts & SciencesThe Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence

Unlike nonhuman assets such as property and machinery, much of a person’s
knowledge is inalienable, both in the practical sense that no one person can know
everything that another person knows and in the legal sense that its ownership
cannot be legally transferred.27 In contrast, when knowledge becomes codified
and digitized, it can be owned, transferred, and concentrated very easily. Thus,
when knowledge shifts from humans to machines, it opens the possibility of con-
centration of power. When historians look back on the first two decades of the
twenty-first century, they will note the striking growth in the digitization and cod-
ification of information and knowledge.28 In parallel, machine learning models
are becoming larger, with hundreds of billions of parameters, using more data and
getting more accurate results.29

More formally, incomplete contracts theory shows how ownership of key as-
sets provides bargaining power in relationships between economic agents (such
as employers and employees, or business owners and subcontractors).30 To the
extent that a person controls an indispensable asset (like useful knowledge) need-
ed to create and deliver a company’s products and services, that person can com-
mand not only higher income but also a voice in decision-making. When useful
knowledge is inalienably locked in human brains, so too is the power it confers.
But when it is made alienable, it enables (though does not demand) greater con-
centration of decision-making and power.31

T he risks of the Turing Trap are amplified because three groups of people–

technologists, businesspeople, and policy-makers–each find automa-
tion alluring. Technologists have sought to replicate human intelligence
for decades to address the recurring challenge of what computers could not do.
The invention of computers and the birth of the term “electronic brain” were the
latest fuel for the ongoing battle between technologists and humanist philoso-
phers.32 The philosophers posited a long list of ordinary and lofty human capac-
ities that computers would never be able to do. No machine could play checkers,
master chess, read printed words, recognize speech, translate between human
languages, distinguish images, climb stairs, win at Jeopardy or Go, write poems,
and so forth.

For professors, it is tempting to assign such projects to their graduate students.
Devising challenges that are new, useful, and achievable can be as difficult as solv-
ing them. Rather than specify a task that neither humans nor machines have ever
done before, why not ask the research team to design a machine that replicates an
existing human capability? Unlike more ambitious goals, replication has an exis-
tence proof that such tasks are, in principle, feasible and useful.

While the appeal of human-like systems is clear, the paradoxical reality is that
HLAI can be more difficult and less valuable than systems that achieve super-
human performance.

277

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

151 (2) Spring 2022Erik Brynjolfsson

In 1988, robotics researcher Hans Moravec noted that “it is comparatively easy
to make computers exhibit adult level performance on intelligence tests or play-
ing checkers, and difficult or impossible to give them the skills of a one-year-old
when it comes to perception and mobility.”33 But I would argue that in many do-
mains, Moravec was not nearly ambitious enough. It is often comparatively easier
for a machine to achieve superhuman performance in new domains than to match
ordinary humans in the tasks they do regularly.

Humans have evolved over millions of years to be able to comfort a baby, nav-
igate a cluttered forest, or pluck the ripest blueberry from a bush. These tasks
are difficult if not impossible for current machines. But machines excel when it
comes to seeing X-rays, etching millions of transistors on a fragment of silicon, or
scanning billions of webpages to find the most relevant one. Imagine how feeble
and limited our technology would be if past engineers set their sights on merely
matching human-levels of perception, actuation, and cognition.

Augmenting humans with technology opens an endless frontier of new abili-
ties and opportunities. The set of tasks that humans and machines can do together
is undoubtedly much larger than those humans can do alone (Figure 1). Machines
can perceive things that are imperceptible to humans, they can act on objects in
ways that no human can, and, most intriguingly, they can comprehend things that
are incomprehensible to the human brain. As Demis Hassabis, CEO of DeepMind,
put it, the AI system “doesn’t play like a human, and it doesn’t play like a program.
It plays in a third, almost alien, way . . . it’s like chess from another dimension.”34
Computer scientist Jonathan Schaeffer explains the source of its superiority: “I’m
absolutely convinced it’s because it hasn’t learned from humans.”35 More funda-
mentally, inventing tools that augment the process of invention itself promises to
expand not only our collective abilities, but to accelerate the rate of expansion of
those abilities.

What about businesspeople? They often find that substituting machinery for
human labor is the low-hanging fruit of innovation. The simplest approach is to
implement plug-and-play automation: swap in a piece of machinery for each task
a human is currently doing. That mindset reduces the need for more radical chang-
es to business processes.36 Task-level automation reduces the need to understand
subtle interdependencies and creates easy A-B tests, by focusing on a known task
with easily measurable performance improvement.

Similarly, because labor costs are the biggest line item in almost every company’s
budget, automating jobs is a popular strategy for managers. Cutting costs–which
can be an internally coordinated effort–is often easier than expanding markets.
Moreover, many investors prefer “scalable” business models, which is often a syn-
onym for a business that can grow without hiring and the complexities that entails.
But here again, when businesspeople focus on automation, they often set out
to achieve a task that is both less ambitious and more difficult than it need be.

278

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Dædalus, the Journal of the American Academy of Arts & SciencesThe Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence

Figure 1
Opportunities for Augmenting Humans Are Far Greater than
Opportunities to Automate Existing Tasks

Tasks That
Humans Can Do

Human Tasks
That Machines
Could Automate

New Tasks That
Humans Can Do with
the Help of Machines

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

To understand the limits of substitution-oriented automation, consider a thought
experiment. Imagine that our old friend Dædalus had at his disposal an extreme-
ly talented team of engineers 3,500 years ago and built human-like machines that
fully automated every work-related task that his fellow Greeks were doing.

9 Herding sheep? Automated.
9 Making clay pottery? Automated.
9 Weaving tunics? Automated.
9 Repairing horse-drawn carts? Automated.
9 Incense and chanting for victims of disease? Automated.

The good news is that labor productivity would soar, freeing the ancient
Greeks for a life of leisure. The bad news is that their living standards and health
outcomes would come nowhere near matching ours. After all, there is only so
much value one can get from clay pots and horse-drawn carts, even with unlimit-
ed quantities and zero prices.

In contrast, most of the value that our economy has created since ancient times
comes from new goods and services that not even the kings of ancient empires
had, not from cheaper versions of existing goods.37 In turn, myriad new tasks are

279

151 (2) Spring 2022Erik Brynjolfsson

required: fully 60 percent of people are now employed in occupations that did not
exist in 1940.38 In short, automating labor ultimately unlocks less value than aug-
menting it to create something new.

At the same time, automating a whole job is often brutally difficult. Every job
involves multiple different tasks, including some that are extremely challenging
to automate, even with the cleverest technologies. For example, AI may be able to
read mammograms better than a human radiologist, but it is not very good at the
other twenty-six tasks associated with the job, according to O-NET, such as com-
forting a concerned patient or coordinating on a care plan with other doctors.39
My work with Tom Mitchell and Daniel Rock on the suitability for machine learn-
ing analyzed 950 distinct occupations. We found that machines could perform at
least some tasks in most occupations, but zero in which machine learning could
do 100 percent of the tasks.40

The same principle applies to the more complex production systems that in-
volve multiple people working together.41 To be successful, firms typically need to
adopt a new technology as part of a system of mutually reinforcing organizational
changes.42 Consider another thought experiment: Imagine if Jeff Bezos had “au-
tomated” existing bookstores by simply replacing all the human cashiers with ro-
bot cashiers. That might have cut costs a bit, but the total impact would have been
muted. Instead, Amazon reinvented the concept of a bookstore by combining hu-
mans and machines in a novel way. As a result, they offer vastly greater product
selection, ratings, reviews, and advice, and enable 24/7 retail access from the com-
fort of customers’ homes. The power of the technology was not in automating the
work of humans in the existing retail bookstore concept but in reinventing and
augmenting how customers find, assess, purchase, and receive books and, in turn,
other retail goods.

Third, policy-makers have also often tilted the playing field toward automat-
ing human labor rather than augmenting it. For instance, the U.S. tax code cur-
rently encourages capital investment over investment in labor through effective
tax rates that are much higher on labor than on plants and equipment.43

Consider a third thought experiment: Two potential ventures each use AI to
create $1 billion of profits. If one of them achieves this by augmenting and em-
ploying a thousand workers, the firm will owe corporate and payroll taxes, while
the employees will pay income taxes, payroll taxes, and other taxes. If the second
business has no employees, the government may collect the same corporate taxes,
but no payroll taxes and no taxes paid by workers. As a result, the second business
model pays far less in total taxes.

This disparity is amplified because the tax code treats labor income more
harshly than capital income. In 1986, top tax rates on capital income and labor
income were equalized in the United States, but since then, successive changes
have created a large disparity, with the 2021 top marginal federal tax rates on labor

280

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Dædalus, the Journal of the American Academy of Arts & SciencesThe Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence

income of 37 percent, while long capital gains have a variety of favorable rules, in-
cluding a lower statutory tax rate of 20 percent, the deferral of taxes until capital
gains are realized, and the “step-up basis” rule that resets capital gains to zero,
wiping out the associated taxes, when assets are inherited.

The first rule of tax policy is simple: you tend to get less of whatever you tax.
Thus, a tax code that treats income that uses labor less favorably than income de-
rived from capital will favor automation over augmentation. Treating both busi-
ness models equally would lead to more balanced incentives. In fact, given the
positive externalities of more widely shared prosperity, a case could be made for
treating wage income more favorably than capital income, for instance by expand-
ing the earned income tax credit.44 It is unlikely that any government official can
define in advance exactly which technologies and innovations augment humans
rather than merely substitute for them; indeed, most technologies have elements
of each and the outcome depends a great deal on how they are deployed. Thus,
rather than prescribe or proscribe specific technologies, a broad-based set of in-
centives can gently nudge technologists and managers toward augmentation on
the margin, much as carbon taxes encourage myriad types of cleaner energy or
research and development tax credits encourage greater investments in research.
Government policy in other areas could also do more to steer the economy clear
of the Turing Trap. The growing use of AI, even if only for complementing work-
ers, and the further reinvention of organizations around this new general-purpose
technology imply a great need for worker training or retraining. In fact, for each
dollar spent on machine learning technology, companies may need to spend nine
dollars on intangible human capital.45 However, education and training suffer
from a serious externality issue: companies that incur the costs to train or retrain
workers may reap only a fraction of the benefits of those investments, with the
rest potentially going to other companies, including competitors, as these work-
ers are free to bring their skills to their new employers. At the same time, work-
ers are often cash- and credit-constrained, limiting their ability to invest in their
own skills development.46 This implies that government policy should directly
provide education and training or provide incentives for corporate training that
offset the externalities created by labor mobility.47

In sum, the risks of the Turing Trap are increased not by just one group in our
society, but by the misaligned incentives of technologists, businesspeople, and
policy-makers.

T he future is not preordained. We control the extent to which AI either ex-

pands human opportunity through augmentation or replaces humans
through automation. We can work on challenges that are easy for ma-
chines and hard for humans, rather than hard for machines and easy for humans.
The first option offers the opportunity of growing and sharing the economic pie

281

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

151 (2) Spring 2022Erik Brynjolfsson

by augmenting the workforce with tools and platforms. The second option risks
dividing the economic pie among an ever-smaller number of people by creating
automation that displaces ever-more types of workers.

While both approaches can and do contribute to productivity and progress,
technologists, businesspeople, and policy-makers have each been putting a finger
on the scales in favor of replacement. Moreover, the tendency of a greater concen-
tration of technological and economic power to beget a greater concentration of
political power risks trapping a powerless majority into an unhappy equilibrium:
the Turing Trap.

The backlash against free trade offers a cautionary tale. Economists have long
argued that free trade and globalization tend to grow the economic pie through the
power of comparative advantage and specialization. They have also acknowledged
that market forces alone do not ensure that every person in every country will
come out ahead. So they proposed a grand bargain: maximize free trade to max-
imize wealth creation and then distribute the benefits broadly to compensate any
injured occupations, industries, and regions. It has not worked as they had hoped.
As the economic winners gained power, they reneged on the second part of the bar-
gain, leaving many workers worse off than before.48 The result helped fuel a popu-
list backlash that led to import tariffs and other barriers to free trade. Economists
wept.

Some of the same dynamics are already underway with AI. More and more
Americans, and indeed workers around the world, believe that while the technolo-
gy may be creating a new billionaire class, it is not working for them. The more tech-
nology is used to replace rather than augment labor, the worse the disparity may be-
come, and the greater the resentments that feed destructive political instincts and
actions. More fundamentally, the moral imperative of treating people as ends, and
not merely as means, calls for everyone to share in the gains of automation.

The solution is not to slow down technology, but rather to eliminate or reverse
the excess incentives for automation over augmentation. A good start would be to
replace the Turing Test, and the mindset it embodies, with a new set of practical
benchmarks that steer progress toward AI-powered systems that exceed anything
that could be done by humans alone. In concert, we must build political and eco-
nomic institutions that are robust in the face of the growing power of AI. We can
reverse the growing tech backlash by creating the kind of prosperous society that
inspires discovery, boosts living standards, and offers political inclusion for ev-
eryone. By redirecting our efforts, we can avoid the Turing Trap and create pros-
perity for the many, not just the few.

282

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Dædalus, the Journal of the American Academy of Arts & SciencesThe Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence

author’s note

The core ideas in this essay were inspired by a series of conversations with James
Manyika and Andrew McAfee. I am grateful for valuable comments and sugges-
tions on this work from Matt Beane, Seth Benzell, Avi Goldfarb, Katya Klinova, Ale-
na Kykalova, Gary Marcus, Andrea Meyer, Dana Meyer, and numerous participants
at seminars at the Stanford Digital Economy Lab and the University of Toronto
Creative Destruction Lab, but they should not be held responsible for any errors or
opinions in the essay.

about the author

Erik Brynjolfsson is the Jerry Yang and Akiko Yamazaki Professor and Senior
Fellow at the Institute for Human-Centered AI and Director of the Digital Econ-
omy Lab at Stanford University. He is also the Ralph Landau Senior Fellow at the
Institute for Economic Policy Research and Professor by Courtesy at the Gradu-
ate School of Business and Department of Economics at Stanford University; and a
Research Associate at the National Bureau of Economic Research. He is the author
or coauthor of seven books, including Machine, Platform, Crowd: Harnessing Our Digi-
tal Future (2017), The Second Machine Age: Work, Progress, and Prosperity in a Time of Bril-
liant Technologies (2014), and Race against the Machine: How the Digital Revolution Is Acceler-
ating Innovation, Driving Productivity, and Irreversibly Transforming Employment and the Econ-
omy (2011) with Andrew McAfee, and Wired for Innovation: How Information Technology
Is Reshaping the Economy (2009) with Adam Saunders.

endnotes

1 Alan Turing, “Computing Machinery and Intelligence,” Mind 59 (236): 433–460, https://
doi.org/10.1093/mind/LIX.236.433. An earlier articulation of this test comes from Des-
cartes in The Discourse, in which he wrote,

If there were machines which bore a resemblance to our bodies and imitated our
actions as closely as possible for all practical purposes, we should still have two
very certain means of recognizing that they were not real men. The first is that
they could never use words, or put together signs, as we do in order to declare our
thoughts to others. . . . Secondly, even though some machines might do some things
as well as we do them, or perhaps even better, they would inevitably fail in others,
which would reveal that they are acting not from understanding.

2 Carolyn Price, “Plato, Opinions and the Statues of Daedalus,” OpenLearn, updated
June 19, 2019, https://www.open.edu/openlearn/history-the-arts/philosophy/plato
-opinions-and-the-statues-daedalus; and Andrew Stewart, “The Archaic Period,” Perseus
Digital Library, http://www.perseus.tufts.edu/hopper/text?doc=Perseus:text:1999.04
.0008:part=2:chapter=1&highlight=daedalus.

3 “The Origin of the Word ‘Robot,’” Science Friday, April 22, 2011, https://www.science

friday.com/segments/the-origin-of-the-word-robot/.

4 Millions of people are now working alongside robots. For a recent survey on the diffusion
of robots, AI, and other advanced technologies in the United States, see Nikolas Zolas,

283

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

151 (2) Spring 2022Erik Brynjolfsson

Zachary Kroff, Erik Brynjolfsson, et al., “Advanced Technologies Adoption and Use
by U.S. Firms: Evidence from the Annual Business Survey,” NBER Working Paper No.
28290 (Cambridge, Mass.: National Bureau of Economic Research, 2020).

5 Apologies to Arthur C. Clarke.

6 See, for example, Daniel Zhang, Saurabh Mishra, Erik Brynjolfsson, et al., “The AI Index
2021 Annual Report,” arXiv (2021), esp. chap. 2, https://arxiv.org/abs/2103.06312. In
regard to image recognition, see, for instance, the success of image recognition systems
in Olga Russakovsky, Jia Deng, Hao Su, et al., “Imagenet Large Scale Visual Recogni-
tion Challenge,” International Journal of Computer Vision 115 (3) (2015): 211–252. A broad
array of business application is discussed in Erik Brynjolfsson and Andrew McAfee,
“The Business of Artificial Intelligence,” Harvard Business Review (2017): 3–11.

7 See, for example, Hubert Dreyfus, What Computers Can’t Do (Cambridge, Mass.: MIT Press,
1972); Nils J. Nilsson, “Human-Level Artificial Intelligence? Be Serious!” AI Magazine
26 (4) (2005): 68; and Gary Marcus, Francesca Rossi, and Manuela Veloso, “Beyond
the Turing Test,” AI Magazine 37 (1) (2016): 3–4.
8 Nilsson, “Human-Level Artificial Intelligence?” 68.
9 John Searle was the first to use the terms strong AI and weak AI, writing that with weak AI,
“the principal value of the computer . . . is that it gives us a very powerful tool,” while
strong AI “really is a mind.” Ed Feigenbaum has argued that creating such intelligence
is the “manifest destiny” of computer science. John R. Searle, “Minds, Brains, and Pro-
grams,” Behavioral and Brain Sciences 3 (3) (1980): 417–457.

10 However, this does not necessarily mean living standards would rise without bound.
In fact, if working hours fall faster than productivity rises, it is theoretically possible,
though empirically unlikely, that output and consumption (other than leisure time)
would fall.

11 See, for example, Robert M. Solow, “A Contribution to the Theory of Economic Growth,”

The Quarterly Journal of Economics 70 (1) (1956): 65–94.

12 See, for example, Daron Acemoglu, “Directed Technical Change,” Review of Economic

Studies 69 (4) (2002): 781–809.

13 See, for instance, Erik Brynjolfsson and Andrew McAfee, Race Against the Machine: How
the Digital Revolution Is Accelerating Innovation, Driving Productivity, and Irreversibly Transforming
Employment and the Economy (Lexington, Mass.: Digital Frontier Press, 2011); and Daron
Acemoglu and Pascual Restrepo, “The Race Between Machine and Man: Implications
of Technology for Growth, Factor Shares, and Employment,” American Economic Review
108 (6) (2018): 1488–1542.

14 For instance, the real wage of a building laborer in Great Britain is estimated to have
grown from sixteen times the amount needed for subsistence in 1820 to 167 times that
level by the year 2000, according to Jan Luiten Van Zanden, Joerg Baten, Marco Mira
d’Ercole, et al., eds., How Was Life? Global Well-Being since 1820 (Paris: OECD Publishing,
2014).

15 For instance, a majority of aircraft on U.S. Navy aircraft carriers are likely to be un-
manned. See Oriana Pawlyk, “Future Navy Carriers Could Have More Drones Than
Manned Aircraft, Admiral Says,” Military.com, March 30, 2021. Similarly, companies
like Kittyhawk have developed pilotless aircraft (“flying cars”) for civilian passengers.

284

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Dædalus, the Journal of the American Academy of Arts & SciencesThe Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence

16 Loukas Karabarbounis and Brent Neiman, “The Global Decline of the Labor Share,” The
Quarterly Journal of Economics 129 (1) (2014): 61–103; and David Autor, “Work of the Past,
Work of the Future,” NBER Working Paper No. 25588 (Cambridge, Mass.: National Bu-
reau of Economic Research, 2019). For a broader survey, see Morgan R. Frank, David
Autor, James E. Bessen, et al., “Toward Understanding the Impact of Artificial Intelli-
gence on Labor,” Proceedings of the National Academy of Sciences 116 (14) (2019): 6531–6539.
17 Daron Acemoglu and David Autor, “Skills, Tasks and Technologies: Implications for

Employment and Earnings,” Handbook of Labor Economics 4 (2011): 1043–1171.

18 Seth G. Benzell and Erik Brynjolfsson, “Digital Abundance and Scarce Architects:
Implications for Wages, Interest Rates, and Growth,” NBER Working Paper No. 25585
(Cambridge, Mass.: National Bureau of Economic Research, 2021).

19 Prasanna Tambe, Lorin Hitt, Daniel Rock, and Erik Brynjolfsson, “Digital Capital and
Superstar Firms,” Hutchins Center Working Paper #73 (Washington, D.C.: Hutchins
Center at Brookings, 2021), https://www.brookings.edu/research/digital-capital-and
-superstar-firms.

20 There is some evidence that capital is already becoming an increasingly good substitute
for labor. See, for instance, the discussion in Michael Knoblach and Fabian Stöckl,
“What Determines the Elasticity of Substitution between Capital and Labor? A Litera-
ture Review,” Journal of Economic Surveys 34 (4) (2020): 852.

21 See, for example, Tyler Cowen, Average Is Over: Powering America beyond the Age of the Great
Stagnation (New York: Penguin, 2013). Or more provocatively, Yuval Noah Harari,
“The Rise of the Useless Class,” Ted Talk, February 24, 2017, https://ideas.ted.com/
the-rise-of-the-useless-class/.

22 Anton Korinek and Joseph E. Stiglitz, “Artificial Intelligence and Its Implications for In-
come Distribution and Unemployment,” in The Economics of Artificial Intelligence, ed. Ajay
Agrawal, Joshua Gans, and Avi Goldfarb (Chicago: University of Chicago Press, 2019),
349–390.

23 Erik Brynjolfsson and Andrew McAfee, “Artificial Intelligence, for Real,” Harvard Business

Review, August 7, 2017.

24 Robert D. Putnam, Our Kids: The American Dream in Crisis (New York: Simon and Schuster,
2016) describes the negative effects of joblessness, while Anne Case and Angus Deaton,
Deaths of Despair and the Future of Capitalism (Princeton, N.J.: Princeton University Press,
2021) documents the sharp decline in life expectancy among many of the same people.
25 Simon Smith Kuznets, Economic Growth and Structure: Selected Essays (New York: W. W.

Norton & Co., 1965).

26 Friedrich August Hayek, “The Use of Knowledge in Society,” The American Economic Review

35 (4) (1945): 519–530.

27 Erik Brynjolfsson, “Information Assets, Technology and Organization,” Management

Science 40 (12) (1994): 1645–1662, https://doi.org/10.1287/mnsc.40.12.1645.

28 For instance, in the year 2000, an estimated 85 billion (mostly analog) photos were tak-
en, but by 2020, that had grown nearly twenty-fold to 1.4 trillion (almost all digital)
photos.

285

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

151 (2) Spring 2022Erik Brynjolfsson

29 Andrew Ng, “What Data Scientists Should Know about Deep Learning,” speech pre-
sented at Extract Data Conference, November 24, 2015, https://www.slideshare.net/
ExtractConf/andrew-ng-chief-scientist-at-baidu (accessed September 9, 2021).

30 Sanford J. Grossman and Oliver D. Hart, “The Costs and Benefits of Ownership: A The-
ory of Vertical and Lateral Integration,” Journal of Political Economy 94 (4) (1986): 691–
719; and Oliver D. Hart and John Moore, “Property Rights and the Nature of the Firm,”
Journal of Political Economy 98 (6) (1990): 1119–1158.

31 Erik Brynjolfsson and Andrew Ng, “Big AI Can Centralize Decisionmaking and Power.
And That’s a Problem,” MILA-UNESCO Working Paper (Montreal: MILA-UNESCO,
2021).

32 “Simon Electronic Brain–Complete History of the Simon Computer,” History Com-
puter, January 4, 2021, https://history-computer.com/simon-electronic-brain-complete
-history-of-the-simon-computer/.

33 Hans Moravec, Mind Children: The Future of Robot and Human Intelligence (Cambridge,

Mass.: Harvard University Press, 1988).

34 Will Knight, “Alpha Zero’s ‘Alien’ Chess Shows the Power, and the Peculiarity, of AI,”

Technology Review, December 2017.

35 Richard Waters, “Techmate: How AI Rewrote the Rules of Chess,” Financial Times, Janu-

ary 12, 2018.

36 Matt Beane and Erik Brynjolfsson, “Working with Robots in a Post-Pandemic World,”

MIT Sloan Management Review 62 (1) (2020): 1–5.

37 Timothy Bresnahan and Robert J. Gordon, “Introduction,” The Economics of New Goods

(Chicago: University of Chicago Press, 1996).

38 David Autor, Anna Salomons, and Bryan Seegmiller, “New Frontiers: The Origins and

Content of New Work, 1940–2018,” NBER Preprint, July 26, 2021.

39 David Killock, “AI Outperforms Radiologists in Mammographic Screening,” Nature
Reviews Clinical Oncology 17 (134) (2020), https://doi.org/10.1038/s41571-020-0329-7.
40 Erik Brynjolfsson, Tom Mitchell, and Daniel Rock, “What Can Machines Learn, and
What Does It Mean for Occupations and the Economy?” AEA Papers and Proceedings
(2018): 43–47.

41 Erik Brynjolfsson, Daniel Rock, and Prasanna Tambe, “How Will Machine Learning
Transform the Labor Market?” Governance in an Emerging New World (619) (2019), https:
//www.hoover.org/research/how-will-machine-learning-transform-labor-market.
42 Paul Milgrom and John Roberts, “The Economics of Modern Manufacturing: Technol-
ogy, Strategy, and Organization,” American Economic Review 80 (3) (1990): 511–528.
43 See Daron Acemoglu, Andrea Manera, and Pascual Restrepo, “Does the U.S. Tax Code
Favor Automation?” Brookings Papers on Economic Activity (Spring 2020); and Daron Ace-
moglu, ed., Redesigning AI (Cambridge, Mass.: MIT Press, 2021).

44 This reverses the classic result suggesting that taxes on capital should be lower than taxes
on labor. Christophe Chamley, “Optimal Taxation of Capital Income in General Equi-
librium with Infinite Lives,” Econometrica 54 (3) (1986): 607–622; and Kenneth L. Judd,
“Redistributive Taxation in a Simple Perfect Foresight Model,” Journal of Public Econom-
ics 28 (1) (1985): 59–83.

286

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

Dædalus, the Journal of the American Academy of Arts & SciencesThe Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence

45 Tambe et al., “Digital Capital and Superstar Firms.”
46 Katherine S. Newman, Chutes and Ladders: Navigating the Low-Wage Labor Market (Cam-

bridge, Mass.: Harvard University Press, 2006).

47 While the distinction between complements and substitutes is clear in economic theory,
it can be trickier in practice. Part of the appeal of broad training and/or tax incentives,
rather than specific technology mandates or prohibitions, is that they allow technol-
ogies, entrepreneurs, and, ultimately, the market to reward approaches that augment
labor rather than replace it.

48 See David H. Autor, David Dorn, and Gordon H. Hanson, “The China Shock: Learning
from Labor-Market Adjustment to Large Changes in Trade,” Annual Review of Economics
8 (2016): 205–240.

D
o
w
n
o
a
d
e
d

f
r
o
m
h

t
t

:
/
/

d
i
r
e
c
t
.

i
t
.

e
d
u
d
a
e
d
a
r
t
i
c
e
–
p
d

f
/

1
5
1
2
2
7
2
2
0
6
0
6
0
4
d
a
e
d
_
a
_
0
1
9
1
5
p
d

b
y
g
u
e
s
t

o
n
0
8
S
e
p
e
m
b
e
r
2
0
2
3

287

151 (2) Spring 2022Erik Brynjolfsson
Download pdf