ARTIFICIAL INTELLIGENCE

AND LIFE IN 2030

ONE HUNDRED YEAR STUDY ON ARTIFICIAL INTELLIGENCE | REPORT OF THE 2015 STUDY PANEL | SEPTEMBER 2016

PREFACE
The One Hundred Year Study on
Artificial Intelligence, launched
in the fall of 2014, is a long-
term investigation of the field of
Artificial Intelligence (AI) and
its influences on people, their
communities, and society. It
considers the science, engineering,
and deployment of AI-enabled
computing systems. As its core
activity, the Standing Committee
that oversees the One Hundred
Year Study forms a Study Panel
every five years to assess the
current state of AI. The Study
Panel reviews AIs progress in the years following the immediately prior report, The overarching purpose
envisions the potential advances that lie ahead, and describes the technical and
societal challenges and opportunities these advances raise, including in such arenas as of the One Hundred Year
ethics, economics, and the design of systems compatible with human cognition. The
overarching purpose of the One Hundred Year Studys periodic expert review is to
Studys periodic expert
provide a collected and connected set of reflections about AI and its influences as the review is to provide a
field advances. The studies are expected to develop syntheses and assessments that
provide expert-informed guidance for directions in collected and connected
AI research, development, and systems design, as well as programs and policies to
help ensure that these systems broadly benefit individuals and society.1 set of reflections about
The One Hundred Year Study is modeled on an earlier effort informally known as
AI and its influences as
the AAAI Asilomar Study. During 2008-2009, the then president of the Association
for the Advancement of Artificial Intelligence (AAAI), Eric Horvitz, assembled a the field advances.
group of AI experts from multiple institutions and areas of the field, along with
scholars of cognitive science, philosophy, and law. Working in distributed subgroups,
the participants addressed near-term AI developments, long-term possibilities,
and legal and ethical concerns, and then came together in a three-day meeting at
Asilomar to share and discuss their findings. A short written report on the intensive
meeting discussions, amplified by the participants subsequent discussions with other
colleagues, generated widespread interest and debate in the field and beyond.
The impact of the Asilomar meeting, and important advances in AI that included
AI algorithms and technologies starting to enter daily life around the globe, spurred
the idea of a long-term recurring study of AI and its influence on people and society.
The One Hundred Year Study was subsequently endowed at a university to enable
1 One Hundred Year Study on Artificial Intelligence (AI100), Stanford University, accessed
August 1, 2016, https://ai100.stanford.edu.
 extended deep thought and cross-disciplinary scholarly investigations that could
inspire innovation and provide intelligent advice to government agencies and industry.
This report is the first in the planned series of studies that will continue for at least
a hundred years. The Standing Committee defined a Study Panel charge for the
inaugural Study Panel in the summer of 2015 and recruited Professor Peter Stone,
at the University of Texas at Austin, to chair the panel. The seventeen-member
Study Panel, comprised of experts in AI from academia, corporate laboratories
and industry, and AI-savvy scholars in law, political science, policy, and economics,
was launched in mid-fall 2015. The participants represent diverse specialties and
geographic regions, genders, and career stages.
The Standing Committee extensively discussed ways to frame the Study Panel
charge to consider both recent advances in AI and potential societal impacts on jobs,
the environment, transportation, public safety, healthcare, community engagement,
and government. The committee considered various ways to focus the study,
including surveying subfields and their status, examining a particular technology
such as machine learning or natural language processing, and studying particular
application areas such as healthcare or transportation. The committee ultimately
chose a thematic focus on AI and Life in 2030 to recognize that AIs various uses
and impacts will not occur independently of one another, or of a multitude of other
societal and technological developments. Acknowledging the central role cities have
played throughout most of human experience, the focus was narrowed to the large
urban areas where most people live. The Standing Committee further narrowed the
focus to a typical North American city in recognition of the great variability of urban
settings and cultures around the world, and limits on the first Study Panels efforts.
The Standing Committee expects that the projections, assessments, and proactive
guidance stemming from the study will have broader global relevance and is making
plans for future studies to expand the scope of the project internationally.

TABLE OF CONTENTS
PREFACE
EXECUTIVE SUMMARY 4
OVERVIEW 6
SECTION I: WHAT IS ARTIFICIAL INTELLIGENCE? 12
Defining AI 12
AI Research Trends 14
SECTION II: AI BY DOMAIN 18
Transportation18
Home/Service Robots 24
Healthcare25
Education31
Low-resource Communities 35
Public Safety and Security 36
Employment and Workplace 38
Entertainment40
SECTION III: PROSPECTS AND RECOMMENDATIONS FOR AI PUBLIC POLICY 42
AI Policy, Now and in the Future 42
2 APPENDIX I: A SHORT HISTORY OF AI 50
 As one consequence of the decision to focus on life in North American cities,
military applications were deemed to be outside the scope of this initial report. This
is not to minimize the importance of careful monitoring and deliberation about
the implications of AI advances for defense and warfare, including potentially
destabilizing developments and deployments.
The report is designed to address four intended audiences. For the general public,
it aims to provide an accessible, scientifically and technologically accurate portrayal
of the current state of AI and its potential. For industry, the report describes relevant
technologies and legal and ethical challenges, and may help guide resource allocation.
The report is also directed to local, national, and international governments to help
them better plan for AI in governance. Finally, the report can help AI researchers,
as well as their institutions and funders, to set priorities and consider the ethical and
legal issues raised by AI research and its applications.
Given the unique nature of the One Hundred Year Study on AI, we expect that
future generations of Standing Committees and Study Panels, as well as research
scientists, policy experts, leaders in the private and public sectors, and the general
public, will reflect on this assessment as they make new assessments of AIs future. We
hope that this first effort in the series stretching out before us will be useful for both its
failures and successes in accurately predicting the trajectory and influences of AI.
The Standing Committee is grateful to the members of the Study Panel for
investing their expertise, perspectives, and significant time to the creation of this
inaugural report. We especially thank Professor Peter Stone for agreeing to serve as
chair of the study and for his wise, skillful, and dedicated leadership of the panel,
its discussions, and creation of the report.
Standing Committee of the One Hundred Year Study of Artificial Intelligence
Barbara J. Grosz, Chair Russ Altman Eric Horvitz
Alan Mackworth Tom Mitchell Deirdre Mulligan Yoav Shoham

STUDY PANEL
Peter Stone, University of Texas at Austin, Chair
Rodney Brooks, Rethink Robotics
Erik Brynjolfsson, Massachussets Institute of Technology
Ryan Calo, University of Washington
Oren Etzioni, Allen Institute for AI
Greg Hager, Johns Hopkins University
Julia Hirschberg, Columbia University
Shivaram Kalyanakrishnan, Indian Institute of Technology Bombay
Ece Kamar, Microsoft Research
Sarit Kraus, Bar Ilan University
Kevin Leyton-Brown, University of British Columbia
David Parkes, Harvard University
William Press, University of Texas at Austin
AnnaLee (Anno) Saxenian, University of California, Berkeley
Julie Shah, Massachussets Institute of Technology
Milind Tambe, University of Southern California
Astro Teller, X
Acknowledgments: The members of the Study Panel gratefully acknowledge the
support of and valuable input from the Standing Committee, especially the chair,
Barbara Grosz, who handled with supreme grace the unenviable role of mediating
between two large, very passionate committees. We also thank Kerry Tremain for his
tireless and insightful input on the written product during the extensive editing and
polishing process, which unquestionably strengthened the report considerably. 3
 EXECUTIVE SUMMARY
Artificial Intelligence (AI) is a science and a set of computational technologies that
are inspired bybut typically operate quite differently fromthe ways people use
their nervous systems and bodies to sense, learn, reason, and take action. While the
rate of progress in AI has been patchy and unpredictable, there have been significant
advances since the fields inception sixty years ago. Once a mostly academic area of
study, twenty-first century AI enables a constellation of mainstream technologies that
Substantial increases are having a substantial impact on everyday lives. Computer vision and AI planning,
for example, drive the video games that are now a bigger entertainment industry than
in the future uses of AI Hollywood. Deep learning, a form of machine learning based on layered representations
of variables referred to as neural networks, has made speech-understanding practical
applications, including
on our phones and in our kitchens, and its algorithms can be applied widely to an
more self-driving cars, array of applications that rely on pattern recognition. Natural Language Processing
(NLP) and knowledge representation and reasoning have enabled a machine to beat
healthcare diagnostics the Jeopardy champion and are bringing new power to Web searches.
While impressive, these technologies are highly tailored to particular tasks. Each
and targeted treatment, application typically requires years of specialized research and careful, unique
and physical assistance construction. In similarly targeted applications, substantial increases in the future
uses of AI technologies, including more self-driving cars, healthcare diagnostics
for elder care can be and targeted treatments, and physical assistance for elder care can be expected. AI
and robotics will also be applied across the globe in industries struggling to attract
expected. younger workers, such as agriculture, food processing, fulfillment centers, and
factories. They will facilitate delivery of online purchases through flying drones,
self-driving trucks, or robots that can get up the stairs to the front door.
This report is the first in a series to be issued at regular intervals as a part of the
One Hundred Year Study on Artificial Intelligence (AI100). Starting from a charge
given by the AI100 Standing Committee to consider the likely influences of AI in a
typical North American city by the year 2030, the 2015 Study Panel, comprising experts
in AI and other relevant areas focused their attention on eight domains they considered
most salient: transportation; service robots; healthcare; education; low-resource
communities; public safety and security; employment and workplace; and entertainment.
In each of these domains, the report both reflects on progress in the past fifteen years
and anticipates developments in the coming fifteen years. Though drawing from a
common source of research, each domain reflects different AI influences and challenges,
such as the difficulty of creating safe and reliable hardware (transportation and service
robots), the difficulty of smoothly interacting with human experts (healthcare and
education), the challenge of gaining public trust (low-resource communities and public
safety and security), the challenge of overcoming fears of marginalizing humans
(employment and workplace), and the social and societal risk of diminishing interpersonal
interactions (entertainment). The report begins with a reflection on what constitutes
Artificial Intelligence, and concludes with recommendations concerning AI-related
policy. These recommendations include accruing technical expertise about AI in
government and devoting more resourcesand removing impedimentsto research
on the fairness, security, privacy, and societal impacts of AI systems.
Contrary to the more fantastic predictions for AI in the popular press, the Study
Panel found no cause for concern that AI is an imminent threat to humankind.
No machines with self-sustaining long-term goals and intent have been developed,
nor are they likely to be developed in the near future. Instead, increasingly useful
applications of AI, with potentially profound positive impacts on our society and
economy are likely to emerge between now and 2030, the period this report
4 considers. At the same time, many of these developments will spur disruptions in
how human labor is augmented or replaced by AI, creating new challenges for the
economy and society more broadly. Application design and policy decisions made in
the near term are likely to have long-lasting influences on the nature and directions
of such developments, making it important for AI researchers, developers, social
scientists, and policymakers to balance the imperative to innovate with mechanisms
to ensure that AIs economic and social benefits are broadly shared across society. If
society approaches these technologies primarily with fear and suspicion, missteps that
slow AIs development or drive it underground will result, impeding important work
on ensuring the safety and reliability of AI technologies. On the other hand, if society While drawing on common
approaches AI with a more open mind, the technologies emerging from the field
could profoundly transform society for the better in the coming decades. research and technologies,
AI systems are specialized
Study Panel: Peter Stone, Chair, University of Texas at Austin, Rodney Brooks,
Rethink Robotics, Erik Brynjolfsson, Massachussets Institute of Technology, Ryan to accomplish particular
Calo, University of Washington, Oren Etzioni, Allen Institute for AI, Greg Hager, Johns
Hopkins University, Julia Hirschberg, Columbia University, Shivaram Kalyanakrishnan, tasks. Each application
Indian Institute of Technology Bombay, Ece Kamar, Microsoft Research, Sarit Kraus,
Bar Ilan University. Kevin Leyton-Brown, University of British Columbia, David Parkes,
requires years of focused
Harvard University, William Press, University of Texas at Austin, AnnaLee (Anno) research and a careful,
Saxenian, University of California, Berkeley, Julie Shah, Massachussets Institute of
Technology, Milind Tambe, University of Southern California, Astro Teller, X unique construction.
Standing Committee of the One Hundred Year Study of Artificial Intelligence:
Barbara J. Grosz, Chair, Russ Altman, Eric Horvitz, Alan Mackworth, Tom Mitchell,
Deidre Mulligan, Yoav Shoham

5
 OVERVIEW
The frightening, futurist portrayals of Artificial Intelligence that dominate films and
novels, and shape the popular imagination, are fictional. In reality, AI is already
changing our daily lives, almost entirely in ways that improve human health, safety,
and productivity. Unlike in the movies, there is no race of superhuman robots on the
horizon or probably even possible. And while the potential to abuse AI technologies
must be acknowledged and addressed, their greater potential is, among other things,
Many have already grown to make driving safer, help children learn, and extend and enhance peoples lives. In
fact, beneficial AI applications in schools, homes, and hospitals are already growing
accustomed to touching at an accelerated pace. Major research universities devote departments to AI studies,
and technology companies such as Apple, Facebook, Google, IBM, and Microsoft
and talking to their
spend heavily to explore AI applications they regard as critical to their futures. Even
smart phones. Peoples Hollywood uses AI technologies to bring its dystopian AI fantasies to the screen.
Innovations relying on computer-based vision, speech recognition, and Natural
future relationships with Language Processing have driven these changes, as have concurrent scientific and
technological advances in related fields. AI is also changing how people interact with
machines will become ever technology. Many people have already grown accustomed to touching and talking to
more nuanced, fluid, and their smart phones. Peoples future relationships with machines will become ever more
nuanced, fluid, and personalized as AI systems learn to adapt to individual personalities
personalized. and goals. These AI applications will help monitor peoples well-being, alert them to
risks ahead, and deliver services when needed or wanted. For example, in a mere
fifteen years in a typical North American citythe time frame and scope of this report
AI applications are likely to transform transportation toward self-driving vehicles with
on-time pickup and delivery of people and packages. This alone will reconfigure the
urban landscape, as traffic jams and parking challenges become obsolete.
This studys focus on a typical North American city is deliberate and meant to
highlight specific changes affecting the everyday lives of the millions of people who
inhabit them. The Study Panel further narrowed its inquiry to eight domains where
AI is already having or is projected to have the greatest impact: transportation,
healthcare, education, low-resource communities, public safety and security,
employment and workplace, home/service robots, and entertainment.
Though drawing from a common source of research, AI technologies have
influenced and will continue to influence these domains differently. Each domain
faces varied AI-related challenges, including the difficulty of creating safe and reliable
hardware for sensing and effecting (transportation and service robots), the difficulty
of smoothly interacting with human experts (healthcare and education), the challenge
of gaining public trust (low-resource communities and public safety and security), the
challenge of overcoming fears of marginalizing humans (employment and workplace)
and the risk of diminishing interpersonal interaction (entertainment). Some domains
are primarily business sectors, such as transportation and healthcare, while others are
more oriented to consumers, such as entertainment and home service robots. Some
cut across sectors, such as employment/workplace and low-resource communities.
In each domain, even as AI continues to deliver important benefits, it also raises
important ethical and social issues, including privacy concerns. Robots and other AI
technologies have already begun to displace jobs in some sectors. As a society, we are now
at a crucial juncture in determining how to deploy AI-based technologies in ways that
promote, not hinder, democratic values such as freedom, equality, and transparency. For
individuals, the quality of the lives we lead and how our contributions are valued are
likely to shift gradually, but markedly. Over the next several years, AI research, systems
development, and social and regulatory frameworks will shape how the benefits of AI
6 are weighed against its costs and risks, and how broadly these benefits are spread.
 An accurate and sophisticated picture of AIone that competes with its popular
portrayalis hampered at the start by the difficulty of pinning down a precise
definition of artificial intelligence. In the approaches the Study Panel considered,
none suggest there is currently a general purpose AI. While drawing on common
research and technologies, AI systems are specialized to accomplish particular
tasks, and each application requires years of focused research and a careful, unique
construction. As a result, progress is uneven within and among the eight domains.
A prime example is Transportation, where a few key technologies have catalyzed
the widespread adoption of AI with astonishing speed. Autonomous transportation Society is now at a crucial
will soon be commonplace and, as most peoples first experience with physically
embodied AI systems, will strongly influence the publics perception of AI. As cars juncture in determining
become better drivers than people, city-dwellers will own fewer cars, live further from
how to deploy AI-based
work, and spend time differently, leading to an entirely new urban organization. In
the typical North American city in 2030, physically embodied AI applications will technologies in ways that
not be limited to cars, but are likely to include trucks, flying vehicles, and personal
robots. Improvements in safe and reliable hardware will spur innovation over the next promote rather than hinder
fifteen years, as they will with Home/Service Robots, which have already entered
peoples houses, primarily in the form of vacuum cleaners. Better chips, low-cost 3D
democratic values such
sensors, cloud-based machine learning, and advances in speech understanding will as freedom, equality, and
enhance future robots services and their interactions with people. Special purpose
robots will deliver packages, clean offices, and enhance security. But technical transparency.
constraints and the high costs of reliable mechanical devices will continue to limit
commercial opportunities to narrowly defined applications for the foreseeable future.
In Healthcare, there has been an immense forward leap in collecting useful data
from personal monitoring devices and mobile apps, from electronic health records (EHR)
in clinical settings and, to a lesser extent, from surgical robots designed to assist with
medical procedures and service robots supporting hospital operations. AI-based
applications could improve health outcomes and the quality of life for millions of
people in the coming years. Though clinical applications have been slow to move from
the computer science lab to the real-world, there are hopeful signs that the pace of
innovation will improve. Advances in healthcare can be promoted via the development
of incentives and mechanisms for sharing data and for removing overbearing policy,
regulatory, and commercial obstacles. For many applications, AI systems will have to work
closely with care providers and patients to gain their trust. Advances in how intelligent
machines interact naturally with caregivers, patients, and patients families are crucial.
Enabling more fluid interactions between people and promising AI technologies
also remains a critical challenge in Education, which has seen considerable progress
in the same period. Though quality education will always require active engagement
by human teachers, AI promises to enhance education at all levels, especially by
providing personalization at scale. Interactive machine tutors are now being matched
to students for teaching science, math, language, and other disciplines. Natural
Language Processing, machine learning, and crowdsourcing have boosted online
learning and enabled teachers in higher education to multiply the size of their
classrooms while addressing individual students learning needs and styles. Over the
next fifteen years in a typical North American city, the use of these technologies in
the classroom and in the home is likely to expand significantly, provided they can be
meaningfully integrated with face-to-face learning.
Beyond education, many opportunities exist for AI methods to assist
Low-resource Communities by providing mitigations and solutions to a
variety of social problems. Traditionally, funders have underinvested in AI research
lacking commercial application. With targeted incentives and funding priorities,
7
 AI technologies could help address the needs of low-resource communities, and
budding efforts are promising. Using data mining and machine learning, for example,
AI has been used to create predictive models to help government agencies address
issues such as prevention of lead poisoning in at-risk children and distribution of
food efficiently. These budding efforts suggest more could be done, particularly if
agencies and organizations can engage and build trust with these communities.
Gaining public trust is also a challenge for AI use by Public Safety and Security
professionals. North American cities and federal agencies have already begun to
Longer term, AI may be deploy AI technologies in border administration and law enforcement. By 2030,
they will rely heavily upon them, including improved cameras and drones for
thought of as a radically surveillance, algorithms to detect financial fraud, and predictive policing. The latter
raises the specter of innocent people being unjustifiably monitored, and care must be
different mechanism
taken to avoid systematizing human bias and to protect civil liberties. Well-deployed
for wealth creation in AI prediction tools have the potential to provide new kinds of transparency about
data and inferences, and may be applied to detect, remove, or reduce human bias,
which everyone should rather than reinforcing it.
Social and political decisions are likewise at play in AIs influences on
be entitled to a portion of Employment and Workplace trends, such as the safety nets needed to protect
the worlds AI-produced people from structural changes in the economy. AI is poised to replace people in
certain kinds of jobs, such as in the driving of taxis and trucks. However, in many
treasures. realms, AI will likely replace tasks rather than jobs in the near term, and will also
create new kinds of jobs. But the new jobs that will emerge are harder to imagine in
advance than the existing jobs that will likely be lost. AI will also lower the cost of
many goods and services, effectively making everyone better off. Longer term, AI may
be thought of as a radically different mechanism for wealth creation in which everyone
should be entitled to a portion of the worlds AI-produced treasures. It is not too soon
for social debate on how the economic fruits of AI technologies should be shared.
Entertainment has been transformed by social networks and other platforms for
sharing and browsing blogs, videos, and photos, which rely on techniques actively
developed in NLP, information retrieval, image processing, crowdsourcing, and machine
learning. Some traditional sources of entertainment have also embraced AI to compose
music, create stage performances, and even to generate 3D scenes from natural
language text. The enthusiasm with which people have already responded to AI-driven
entertainment has been surprising. As with many aspects of AI, there is ongoing debate
about the extent to which the technology replaces or enhances sociability. AI will
increasingly enable entertainment that is more interactive, personalized, and engaging.
Research should be directed toward understanding how to leverage these attributes
for individuals and societys benefit.

Whats next for AI research?

The research that fuels the AI revolution has also seen rapid changes. Foremost
among them is the maturation of machine learning, stimulated in part by the rise of
the digital economy, which both provides and leverages large amounts of data. Other
factors include the rise of cloud computing resources and consumer demand for
widespread access to services such as speech recognition and navigation support.
Machine learning has been propelled dramatically forward by impressive
empirical successes of artificial neural networks, which can now be trained with huge
data sets and large-scale computing. This approach has been come to be known as
deep learning. The leap in the performance of information processing algorithms
has been accompanied by significant progress in hardware technology for basic
operations such as sensing, perception, and object recognition. New platforms and
8
markets for data-driven products, and the economic incentives to find new products
and markets, have also stimulated research advances. Now, as it becomes a central
force in society, the field of AI is shifting toward building intelligent systems that
can collaborate effectively with people, and that are more generally human-aware,
including creative ways to develop interactive and scalable ways for people to
teach robots. These trends drive the currently hot areas of AI research into both
fundamental methods and application areas:
Large-scale machine learning concerns the design of learning algorithms, as
well as scaling existing algorithms, to work with extremely large data sets. The field of AI is shifting
Deep learning, a class of learning procedures, has facilitated object recognition
toward building intelligent
in images, video labeling, and activity recognition, and is making significant inroads into
other areas of perception, such as audio, speech, and natural language processing. systems that can
Reinforcement learning is a framework that shifts the focus of machine
learning from pattern recognition to experience-driven sequential decision-making.
collaborate effectively with
It promises to carry AI applications forward toward taking actions in the real world. people, including creative
While largely confined to academia over the past several decades, it is now seeing
some practical, real-world successes. ways to develop interactive
Robotics is currently concerned with how to train a robot to interact with the
and scalable ways for
world around it in generalizable and predictable ways, how to facilitate manipulation
of objects in interactive environments, and how to interact with people. Advances in people to teach robots.
robotics will rely on commensurate advances to improve the reliability and generality
of computer vision and other forms of machine perception.
Computer vision is currently the most prominent form of machine perception.
It has been the sub-area of AI most transformed by the rise of deep learning. For the
first time, computers are able to perform some vision tasks better than people. Much
current research is focused on automatic image and video captioning.
Natural Language Processing, often coupled with automatic speech recognition,
is quickly becoming a commodity for widely spoken languages with large data sets.
Research is now shifting to develop refined and capable systems that are able to
interact with people through dialog, not just react to stylized requests. Great strides
have also been made in machine translation among different languages, with more
real-time person-to-person exchanges on the near horizon.
Collaborative systems research investigates models and algorithms to help develop
autonomous systems that can work collaboratively with other systems and with humans.
Crowdsourcing and human computation research investigates methods to
augment computer systems by making automated calls to human expertise to solve
problems that computers alone cannot solve well.
Algorithmic game theory and computational social choice draw attention
to the economic and social computing dimensions of AI, such as how systems can
handle potentially misaligned incentives, including self-interested human participants
or firms and the automated AI-based agents representing them.
Internet of Things (IoT) research is devoted to the idea that a wide array of
devices, including appliances, vehicles, buildings, and cameras, can be interconnected to
collect and share their abundant sensory information to use for intelligent purposes.
Neuromorphic computing is a set of technologies that seek to mimic
biological neural networks to improve the hardware efficiency and robustness of
computing systems, often replacing an older emphasis on separate modules for input/
output, instruction-processing, and memory.
9
 AI policy, now and in the future
The measure of success for AI applications is the value they create for human lives.
In that light, they should be designed to enable people to understand AI systems
successfully, participate in their use, and build their trust. Public policies should help
ease societys adaptation to AI applications, extend their benefits, and mitigate their
inevitable errors and failures. Debate about how AI is deployed, including concerns
about how privacy is protected and AIs benefits fairly shared, should be encouraged.
Given the speed with which AI technologies are being realized, and concomitant
Misunderstandings concerns about their implications, the Study Panel recommends that all layers of
government acquire technical expertise in AI. Further, research on the fairness,
about what AI is and is security, privacy, and societal implications of AI systems should be encouraged by
not could fuel opposition removing impediments and increasing private and public spending to support it.
Currently in the United States, at least sixteen separate agencies govern sectors
to technologies with of the economy related to AI technologies. Rapid advances in AI research and,
especially, its applications require experts in these sectors to develop new concepts
the potential to benefit and metaphors for law and policy. Who is responsible when a self-driven car crashes
or an intelligent medical device fails? How can AI applications be prevented from
everyone. Poorly informed
promulgating racial discrimination or financial cheating? Who should reap the gains
regulation that stifles of efficiencies enabled by AI technologies and what protections should be afforded
to people whose skills are rendered obsolete? As people integrate AI more broadly
innovation would be a and deeply into industrial processes and consumer products, best practices need to be
spread, and regulatory regimes adapted.
tragic mistake. While the Study Panel does not consider it likely that near-term AI systems will
autonomously choose to inflict harm on people, it will be possible for people to use
AI-based systems for harmful as well as helpful purposes. And though AI algorithms
may be capable of making less biased decisions than a typical person, it remains
a deep technical challenge to ensure that the data that inform AI-based decisions
can be kept free from biases that could lead to discrimination based on race, sexual
orientation, or other factors.
Faced with the profound changes that AI technologies can produce, pressure for
more and tougher regulation is probably inevitable. Misunderstandings about
what AI is and is not could fuel opposition to technologies with the potential to
benefit everyone. Inappropriate regulatory activity would be a tragic mistake. Poorly
informed regulation that stifles innovation, or relocates it to other jurisdictions, would
be counterproductive.2
Fortunately, principles that guide successful regulation of current digital technologies
provide a starting point. In privacy regulation, broad legal mandates coupled with
tough transparency requirements and meaningful enforcementrather than strict
controlsencourage companies to develop processes and professional staff to
enforce privacy controls, engage with outside stakeholders, and adapt their practices
to technological advances. This in turn supports the development of professional
trade associations and standards committees that spread best practices. In AI, too,
regulators can strengthen a virtuous cycle of activity involving internal and external
accountability, transparency, and professionalization, rather than narrow compliance.
A vigorous and informed debate about how to best steer AI in ways that enrich
our lives and our society, while encouraging creativity in the field, is an urgent and
vital need. AI technologies could widen existing inequalities of opportunity if access
to themalong with the high-powered computation and large-scale data that fuel
many of themis unfairly distributed across society. These technologies will improve
2 Kate Crawford, Artificial Intelligences White Guy Problem, The New York Times, June 25,
2016, accessed August 1, 2016, http://www.nytimes.com/2016/06/26/opinion/sunday/artificial-
10 intelligences-white-guy-problem.html.
the abilities and efficiency of people who have access to them. Policies should be
evaluated as to whether they foster democratic values and equitable sharing of AIs
benefits, or concentrate power and benefits in the hands of a fortunate few.
As this report documents, significant AI-related advances have already had
an impact on North American cities over thepast fifteen years, and even more
substantial developments will occur over the next fifteen. Recent advances are largely
due to the growth and analysis of large data sets enabled by the internet, advances
in sensory technologies and, more recently, applications of deep learning. In the
coming years, as the public encounters new AI applications in domains such as
transportation and healthcare, they must be introduced in ways that build trust and
understanding, and respect human and civil rights. While encouraging innovation,
policies and processes should address ethical, privacy, and security implications, and
should work to ensure that the benefits of AI technologies will be spread broadly and
fairly. Doing so will be critical if Artificial Intelligence research and its applications
are to exert a positive influence on North American urban life in 2030 and beyond.

11
 SECTION I: WHAT IS ARTIFICIAL
INTELLIGENCE?
This section describes how researchers and practitioners define Artificial Intelligence, and
the areas of AI research and application that are currently thriving. It proffers definitions of
what AI is and is not, and describes some of the currently hot areas of AI Research. This
section lays the groundwork for Section II, which elaborates on AIs impacts and future in
eight domains and Section III, which describes issues related to AI design and public policy
An accurate and
and makes recommendations for encouraging AI innovation while protecting democratic values.
sophisticated picture of
DEFINING AI
AIone that competes with
Curiously, the lack of a precise, universally accepted definition of AI probably
its popular portrayalis has helped the field to grow, blossom, and advance at an ever-accelerating pace.
Practitioners, researchers, and developers of AI are instead guided by a rough
hampered by the difficulty sense of direction and an imperative to get on with it. Still, a definition remains
important and Nils J. Nilsson has provided a useful one:
of pinning down a precise
Artificial intelligence is that activity devoted to making machines intelligent, and
definition of artificial intelligence is that quality that enables an entity to function appropriately and with
foresight in its environment.3
intelligence. From this perspective, characterizing AI depends on the credit one is willing to
give synthesized software and hardware for functioning appropriately and with
foresight. A simple electronic calculator performs calculations much faster than
the human brain, and almost never makes a mistake.4 Is a calculator intelligent?
Like Nilsson, the Study Panel takes a broad view that intelligence lies on a multi-
dimensional spectrum. According to this view, the difference between an arithmetic
calculator and a human brain is not one of kind, but of scale, speed, degree of
autonomy, and generality. The same factors can be used to evaluate every other
instance of intelligencespeech recognition software, animal brains, cruise-control
systems in cars, Go-playing programs, thermostatsand to place them at some
appropriate location in the spectrum.
Although our broad interpretation places the calculator within the intelligence
spectrum, such simple devices bear little resemblance to todays AI. The frontier of
AI has moved far ahead and functions of the calculator are only one among the
millions that todays smartphones can perform. AI developers now work on improving,
generalizing, and scaling up the intelligence currently found on smartphones.
In fact, the field of AI is a continual endeavor to push forward the frontier of
machine intelligence. Ironically, AI suffers the perennial fate of losing claim to its
acquisitions, which eventually and inevitably get pulled inside the frontier, a repeating
pattern known as the AI effect or the odd paradoxAI brings a new technology
into the common fold, people become accustomed to this technology, it stops being
considered AI, and newer technology emerges.5 The same pattern will continue in the
future. AI does not deliver a life-changing product as a bolt from the blue. Rather,
AI technologies continue to get better in a continual, incremental way.

3 Nils J. Nilsson, The Quest for Artificial Intelligence: A History of Ideas and Achievements (Cambridge,
UK: Cambridge University Press, 2010).
4 Wikimedia Images, accessed August 1, 2016, https://upload.wikimedia.org/wikipedia/
commons/b/b6/SHARP_ELSIMATE_EL-W221.jpg.
5 Pamela McCorduck, Machines Who Think: A Personal Inquiry into the History and Prospects of
Artificial Intelligence, 2nd ed. (Natick, MA: A. K. Peters, Ltd., 2004; San Francisco: W. H. Freeman,
12 1979), Citations are to the Peters edition.
The human measure
Notably, the characterization of intelligence as a spectrum grants no special status
to the human brain. But to date human intelligence has no match in the biological
and artificial worlds for sheer versatility, with the abilities to reason, achieve
goals, understand and generate language, perceive and respond to sensory inputs,
prove mathematical theorems, play challenging games, synthesize and summarize
information, create art and music, and even write histories.6
This makes human intelligence a natural choice for benchmarking the progress
of AI. It may even be proposed, as a rule of thumb, that any activity computers Intelligence lies on a
are able to perform and people once performed should be counted as an instance
of intelligence. But matching any human ability is only a sufficient condition, not a
multi-dimensional
necessary one. There are already many systems that exceed human intelligence, at spectrum. According to
least in speed, such as scheduling the daily arrivals and departures of thousands of
flights in an airport. this view, the difference
AIs long questand eventual successto beat human players at the game of
chess offered a high-profile instance for comparing human to machine intelligence. between an arithmetic
Chess has fascinated people for centuries. When the possibility of building computers
calculator and a human
became imminent, Alan Turing, who many consider the father of computer science,
mentioned the idea of computers showing intelligence with chess as a paradigm.7 brain is not one of kind,
Without access to powerful computers, Turing played a game in which he simulated
the computer, taking about half an hour per move. but of scale, speed,
But it was only after a long line of improvements in the sixties and seventies
contributed by groups at Carnegie Mellon, Stanford, MIT, The Institute for
degree of autonomy, and
Theoretical and Experimental Physics at Moscow, and Northwestern University generality.
that chess-playing programs started gaining proficiency. The final push came
through a long-running project at IBM, which culminated with the Deep Blue
program beating Garry Kasparov, then the world chess champion, by a score of
3.5-2.5 in 1997. Curiously, no sooner had AI caught up with its elusive target than
Deep Blue was portrayed as a collection of brute force methods that wasnt real
intelligence.8 In fact, IBMs subsequent publication about Deep Blue, which gives
extensive details about its search and evaluation procedures, doesnt mention the
word intelligent even once!9 Was Deep Blue intelligent or not? Once again, the
frontier had moved.

An operational definition
AI can also be defined by what AI researchers do. This report views AI
primarily as a branch of computer science that studies the properties of
intelligence by synthesizing intelligence.10 Though the advent of AI has depended
on the rapid progress of hardware computing resources, the focus here on
software reflects a trend in the AI community. More recently, though, progress in
building hardware tailored for neural-network-based computing11 has created a

6 Nilsson, The Quest for Artificial Intelligence.

7 Nilsson, The Quest for Artificial Intelligence, 89.
8 McCorduck, Machines Who Think, 433.
9 Murray Campbell, A. Joseph Hoane Jr., and Feng-hsiung Hsu, Deep Blue, Artificial
Intelligence 134, nos. 1 and 2 (2002): 5783.
10 Herbert A. Simon, Artificial Intelligence: An Empirical Science, Artificial Intelligence 77, no. 2
(1995):95127.
11 Paul Merolla John V. Arthur, Rodrigo Alvarez-Icaza, Andrew S. Cassidy, Jun Sawada, Filipp
Akopyan, Bryan L. Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, Bernard Brezzo, Ivan Vo,
Steven K. Esser, Rathinakumar Appuswamy, Brian Taba, Arnon Amir, Myron D. Flickner, William
P. Risk, Rajit Manohar, and Dharmendra S. Modha, A Million Spiking-Neuron Integrated
Circuit with a Scalable Communication Network and Interface, accessed August 1, 2016, http://
paulmerolla.com/merolla_main_som.pdf. 13
 tighter coupling between hardware and software in advancing AI.
Intelligence remains a complex phenomenon whose varied aspects have attracted
the attention of several different fields of study, including psychology, economics,
neuroscience, biology, engineering, statistics, and linguistics. Naturally, the field of AI has
benefited from the progress made by all of these allied fields. For example, the artificial
neural network, which has been at the heart of several AI-based solutions12 13 was
originally inspired by thoughts about the flow of information in biological neurons.14

Human intelligence has AI RESEARCH TRENDS

Until the turn of the millennium, AIs appeal lay largely in its promise to deliver, but
no match in the biological
in the last fifteen years, much of that promise has been redeemed.15 AI technologies
and artificial worlds for already pervade our lives. As they becomes a central force in society, the field is
shifting from simply building systems that are intelligent to building intelligent systems
sheer versatility, with that are human-aware and trustworthy.
Several factors have fueled the AI revolution. Foremost among them is the
the abilities to reason, maturing of machine learning, supported in part by cloud computing resources
achieve goals, understand and wide-spread, web-based data gathering. Machine learning has been propelled
dramatically forward by deep learning, a form of adaptive artificial neural
and generate language... networks trained using a method called backpropagation.16 This leap in the
performance of information processing algorithms has been accompanied by
create art and music, and significant progress in hardware technology for basic operations such as sensing,
perception, and object recognition. New platforms and markets for data-driven
even write histories.
products, and the economic incentives to find new products and markets, have also
contributed to the advent of AI-driven technology.
All these trends drive the hot areas of research described below. This compilation
is meant simply to reflect the areas that, by one metric or another, currently receive
greater attention than others. They are not necessarily more important or valuable
than other ones. Indeed, some of the currently hot areas were less popular in past
years, and it is likely that other areas will similarly re-emerge in the future.
Large-scale machine learning
Many of the basic problems in machine learning (such as supervised and
unsupervised learning) are well-understood. A major focus of current efforts is to
scale existing algorithms to work with extremely large data sets. For example, whereas
traditional methods could afford to make several passes over the data set, modern
ones are designed to make only a single pass; in some cases, only sublinear methods
(those that only look at a fraction of the data) can be admitted.
Deep learning
The ability to successfully train convolutional neural networks has most benefited the
field of computer vision, with applications such as object recognition, video

12 Gerald Tesauro, Practical Issues in Temporal Difference Learning, Machine Learning, no. 8
(1992): 25777.
13 David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den
Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander
Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap,
Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis, Mastering the game
of Go with deep neural networks and tree search, Nature 529 (2016): 484489.
14 W. McCulloch and W. Pitts, W., A logical calculus of the ideas immanent in nervous activity,
Bulletin of Mathematical Biophysics, 5 (1943): 115133.
15 Appendix I offers a short history of AI, including a description of some of the traditionally
core areas of research, which have shifted over the past six decades.
16 Backpropagation is an abbreviation for backward propagation of errors, a common method
of training artificial neural networks used in conjunction with an optimization method such as
gradient descent. The method calculates the gradient of a loss function with respect to all the
14 weights in the network.
labeling, activity recognition, and several variants thereof. Deep learning is also making
significant inroads into other areas of perception, such as audio, speech, and natural
language processing.
Reinforcement learning
Whereas traditional machine learning has mostly focused on pattern mining,
reinforcement learning shifts the focus to decision making, and is a technology that
will help AI to advance more deeply into the realm of learning about and executing
actions in the real world. It has existed for several decades as a framework for
experience-driven sequential decision-making, but the methods have not found great AI technologies already
success in practice, mainly owing to issues of representation and scaling. However,
the advent of deep learning has provided reinforcement learning with a shot in the pervade our lives. As
arm. The recent success of AlphaGo, a computer program developed by Google
Deepmind that beat the human Go champion in a five-game match, was due in large they become a central
part to reinforcement learning. AlphaGo was trained by initializing an automated
force in society, the
agent with a human expert database, but was subsequently refined by playing a large
number of games against itself and applying reinforcement learning. field is shifting from
Robotics
Robotic navigation, at least in static environments, is largely solved. Current efforts simply building systems
consider how to train a robot to interact with the world around it in generalizable
that are intelligent to
and predictable ways. A natural requirement that arises in interactive environments
is manipulation, another topic of current interest. The deep learning revolution is building intelligent
only beginning to influence robotics, in large part because it is far more difficult to
acquire the large labeled data sets that have driven other learning-based areas of AI. systems that are
Reinforcement learning (see above), which obviates the requirement of labeled data,
may help bridge this gap but requires systems to be able to safely explore a policy human-aware and
space without committing errors that harm the system itself or others. Advances in
trustworthy.
reliable machine perception, including computer vision, force, and tactile perception,
much of which will be driven by machine learning, will continue to be key enablers to
advancing the capabilities of robotics.
Computer vision
Computer vision is currently the most prominentform of machine perception. It has
been the sub-area of AI most transformed by the rise of deep learning. Until just a
few years ago, support vector machines were the method of choice for most visual
classification tasks.But the confluence of large-scale computing, especially on GPUs,
the availability of large datasets, especially via the internet, and refinements of neural
network algorithms has led to dramatic improvements in performance on benchmark
tasks (e.g., classification on ImageNet17). For the first time, computers are able to
perform some (narrowly defined) visual classification tasks better than people. Much
current research is focused on automaticimage and video captioning.
Natural Language Processing
Often coupled with automatic speech recognition, Natural Language Processing is
another very active area of machine perception. It is quickly becoming a commodity
for mainstream languages with large data sets. Google announced that 20% of
current mobile queries are done by voice,18 and recent demonstrations have proven
the possibility of real-time translation. Research is now shifting towards developing
refined and capable systems that are able to interact with people through dialog, not
just react to stylized requests.

17 ImageNet, Stanford Vision Lab, Stanford University, Princeton University, 2016, accessed
August 1, 2016, www.image-net.org/.
18 Greg Sterling, Google says 20% of mobile queries are voice searches, Search Engine Land,
May 18, 2016, accessed August 1, 2016, http://searchengineland.com/google-reveals-20-percent-
queries-voice-queries-249917. 15
 Collaborative systems
Research on collaborative systems investigates models and algorithms to help
develop autonomous systems that can work collaboratively with other systems and
with humans. This research relies on developing formal models of collaboration,
and studies the capabilities needed for systems to become effective partners. There
is growing interest in applications that can utilize the complementary strengths of
humans and machinesfor humans to help AI systems to overcome their limitations,
and for agents to augment human abilities and activities.
Natural Language Crowdsourcing and human computation
Since human abilities are superior to automated methods for accomplishing many
Processing is a very tasks, research on crowdsourcing and human computation investigates methods to
augment computer systems by utilizing human intelligence to solve problems that
active area of machine
computers alone cannot solve well. Introduced only about fifteen years ago, this
perception. Research research now has an established presence in AI. The best-known example of
crowdsourcing is Wikipedia, a knowledge repository that is maintained and updated
is now shifting towards by netizens and that far exceeds traditionally-compiled information sources, such
as encyclopedias and dictionaries, in scale and depth. Crowdsourcing focuses on
developing systems that devising innovative ways to harness human intelligence. Citizen science platforms
are able to interact with energize volunteers to solve scientific problems, while paid crowdsourcing platforms
such as Amazon Mechanical Turk provide automated access to human intelligence on
people through dialog, demand. Work in this area has facilitated advances in other subfields of AI, including
computer vision and NLP, by enabling large amounts of labeled training data and/or
not just react to stylized human interaction data to be collected in a short amount of time. Current research
efforts explore ideal divisions of tasks between humans and machines based on their
requests.
differing capabilities and costs.
Algorithmic game theory and computational social choice
New attention is being drawn to the economic and social computing dimensions of
AI, including incentive structures. Distributed AI and multi-agent systems have been
studied since the early 1980s, gained prominence starting in the late 1990s, and were
accelerated by the internet. A natural requirement is that systems handle potentially
misaligned incentives, including self-interested human participants or firms, as well
as automated AI-based agents representing them. Topics receiving attention include
computational mechanism design (an economic theory of incentive design, seeking
incentive-compatible systems where inputs are truthfully reported), computational
social choice (a theory for how to aggregate rank orders on alternatives), incentive
aligned information elicitation (prediction markets, scoring rules, peer prediction) and
algorithmic game theory (the equilibria of markets, network games, and parlor games
such as Pokera game where significant advances have been made in recent years
through abstraction techniques and no-regret learning).
Internet of Things (IoT)
A growing body of research is devoted to the idea that a wide array of devices can
be interconnected to collect and share their sensory information. Such devices can
include appliances, vehicles, buildings, cameras, and other things. While its a matter
of technology and wireless networking to connect the devices, AI can process and
use the resulting huge amounts of data for intelligent and useful purposes. Currently,
these devices use a bewildering array of incompatible communication protocols. AI
could help tame this Tower of Babel.
Neuromorphic Computing
Traditional computers implement the von Neumann model of computing, which
separates the modules for input/output, instruction-processing, and memory. With
the success of deep neural networks on a wide array of tasks, manufacturers are
16
actively pursuing alternative models of computingespecially those that are inspired
by what is known about biological neural networkswith the aim of improving the
hardware efficiency and robustness of computing systems. At the moment, such
neuromorphic computers have not yet clearly demonstrated big wins, and are just
beginning to become commercially viable. But it is possible that they will become
commonplace (even if only as additions to their von Neumann cousins) in the
near future. Deep neural networks have already created a splash in the application
landscape. A larger wave may hit when these networks can be trained and executed
on dedicated neuromorphic hardware, as opposed to simulated on standard von A growing body of research
Neumann architectures, as they are today.
is devoted to the idea that
Overall trends and the future of AI research
a wide array of devices
The resounding success of the data-driven paradigm has displaced the traditional
paradigms of AI. Procedures such as theorem proving and logic-based knowledge can be interconnected
representation and reasoning are receiving reduced attention, in part because of the
ongoing challenge of connecting with real-world groundings. Planning, which was a to collect and share their
mainstay of AI research in the seventies and eighties, has also received less attention
of late due in part to its strong reliance on modeling assumptions that are hard
sensory information.
to satisfy in realistic applications. Model-based approachessuch as physics-based Such devices can include
approaches to vision and traditional control and mapping in roboticshave by and
large given way to data-driven approaches that close the loop with sensing the results appliances, vehicles,
of actions in the task at hand. Bayesian reasoning and graphical models, which were
very popular even quite recently, also appear to be going out of favor, having been buildings, cameras, and
drowned by the deluge of data and the remarkable success of deep learning.
other things.
Over the next fifteen years, the Study Panel expects an increasing focus on
developing systems that are human-aware, meaning that they specifically model,
and are specifically designed for, the characteristics of the people with whom they
are meant to interact. There is a lot of interest in trying to find new, creative ways
to develop interactive and scalable ways to teach robots. Also, IoT-type systems
devices and the cloudare becoming increasingly popular, as is thinking about
social and economic dimensions of AI. In the coming years, new perception/object
recognition capabilities and robotic platforms that are human-safe will grow, as will
data-driven products and their markets.
The Study Panel also expects a reemergence of some of the traditional forms of
AI as practitioners come to realize the inevitable limitations of purely end-to-end deep
learning approaches. We encourage young researchers not to reinvent the wheel,
but rather to maintain an awareness of the significant progress in many areas of
AI during the first fifty years of the field, and in related fields such as control theory,
cognitive science, and psychology.

17
 SECTION II: AI BY DOMAIN
Though different instances of AI research and practice share common technologies, such
as machine learning, they also vary considerably in different sectors of the economy and
society. We call these sectors domains, and in this section describe the different states
of AI research and implementation, as well as impacts and distinct challenges, in eight of
them: transportation; home/service robotics; healthcare; education; low-resource communities;
public safety and security; employment and workplace; and entertainment. Based on these
Autonomous analyses, we also predict trends in a typical North American city over the next fifteen years.
Contrary to AIs typical depiction in popular culture, we seek to offer a balanced overview
transportation will soon of the ways in which AI is already beginning to transform everyday life, and how those
transformations are likely to grow by the year 2030.
be commonplace and,
as most peoples first TRANSPORTATION
experience with physically Transportation is likely to be one of the first domains in which the general public
will be asked to trust the reliability and safety of an AI system for a critical task.
embodied AI systems, Autonomous transportation will soon be commonplace and, as most peoples first
experience with physically embodied AI systems, will strongly influence the publics
will strongly influence the perception of AI. Once the physical hardware is made sufficiently safe and robust, its
introduction to daily life may happen so suddenly as to surprise the public, which will
publics perception of AI.
require time to adjust. As cars will become better drivers than people, city-dwellers
will own fewer cars, live further from work, and spend time differently, leading to an
entirely new urban organization. Further, in the typical North American city in 2030,
changes wont be limited to cars and trucks, but are likely to include flying vehicles
and personal robots, and will raise social, ethical and policy issues.
A few key technologies have already catalyzed the widespread adoption of AI
in transportation. Compared to 2000, the scale and diversity of data about personal
and population-level transportation available todayenabled by the adoption of
smartphones and decreased costs and improved accuracies for variety of sensorsis
astounding. Without the availability of this data and connectivity, applications such as
real-time sensing and prediction of traffic, route calculations, peer-to-peer ridesharing
and self-driving cars would not be possible.

Smarter cars
GPS was introduced to personal vehicles in 2001 with in-car navigation devices and
has since become a fundamental part of the transportation infrastructure.19 GPS
assists drivers while providing large-scale information to technology companies and
cities about transportation patterns. Widespread adoption of smartphones with GPS
technology further increased connectivity and the amount of location data shared by
individuals.
Current vehicles are also equipped with a wide range of sensing capabilities.
An average automobile in the US is predicted to have seventy sensors including
gyroscopes, accelerometers, ambient light sensors, and moisture sensors.20 Sensors
are not new to vehicles. Automobiles built before 2000 had sensors for the
internal state of the vehicle such as its speed, acceleration, and wheel position.21

19 Mark Sullivan, A brief history of GPS, PCWorld, August 9, 2012, accessed August 1, 2016,
http://www.pcworld.com/article/2000276/a-brief-history-of-gps.html.
20 William J. Fleming, New Automotive Sensors - A Review, IEEE Sensors Journal 8, no 11,
(2008): 1900-1921.
21 Jean Jacques Meneu, ed., Automotive Sensors: Now and in the Future, Arrow, September 24,
2015, accessed August 1, 2016, https://www.arrow.com/en/research-and-events/articles/automotive-
18 sensors-now-and-in-the-future.
They already had a number of functionalities that combined real-time sensing with
perception and decision-making such as Anti-lock Braking Systems (ABS), airbag
control, Traction Control Systems (TCS), and Electronic Stability Control (ESC).22
Automated capabilities have been introduced into commercial cars gradually since
2003 as summarized in the following table.

Context Automated Functionality Release Date

Parking Intelligent Parking Assist System Since 200323 As cars will become better

Parking Summon Since 201624 drivers than people, city-

dwellers will own fewer
Arterial & Highway Lane departure system Since 2004 in North America25
cars, live further from
Arterial & Highway Adaptive cruise control Since 2005 in North America26
work, and spend time
Highway Blind spot monitoring 200727
differently, leading to
Highway Lane changing 201528 an entirely new urban
These functionalities assist drivers or completely take over well-defined activities organization.
for increased safety and comfort. Current cars can park themselves, perform adaptive
cruise control on highways, steer themselves during stop-and-go traffic, and alert
drivers about objects in blind spots during lane changes. Vision and radar technology
were leveraged to develop pre-collision systems that let cars autonomously brake
when risk of a collision is detected. Deep learning also has been applied to improve
automobiles capacity to detect objects in the environment and recognize sound.29

Self-driving vehicles
Since the 1930s, science fiction writers dreamed of a future with self-driving cars,
and building them has been a challenge for the AI community since the 1960s. By
the 2000s, the dream of autonomous vehicles became a reality in the sea and sky, and
even on Mars, but self-driving cars existed only as research prototypes in labs. Driving
in a city was considered to be a problem too complex for automation due to factors
like pedestrians, heavy traffic, and the many unexpected events that can happen
outside of the cars control. Although the technological components required to

22 Carl Liersch, Vehicle Technology Timeline: From Automated to Driverless, Robert Bosch
(Australia) Pty. Ltd., 2014, accessed August 1, 2016, http://dpti.sa.gov.au/__data/assets/pdf_
file/0009/246807/Carl_Liersch_Presentation.pdf.
23 Intelligent Parking Assist System, Wikipedia, last modified July 26, 2016, accessed August 1,
2016, https://en.wikipedia.org/wiki/Intelligent_Parking_Assist_System.
24 The Tesla Motors Team, Summon Your Tesla from Your Phone, Tesla, January 10, 2016,
accessed August 1, 2016, https://www.teslamotors.com/blog/summon-your-tesla-your-phone.
25 Lane departure warning system, Wikipedia, last modified July 24, 2016, accessed August 1,
2016, https://en.wikipedia.org/wiki/Lane_departure_warning_system.
26 Autonomous cruise control system, Wikipedia, last modified July 30, 2016, accessed August 1,
2016, https://en.wikipedia.org/wiki/Autonomous_cruise_control_system.
27 Blind spot monitor, Wikipedia, last modified April 20, 2016, accessed August 1, 2016,
https://en.wikipedia.org/wiki/Blind_spot_monitor.
28 Dana Hull, Tesla Starts Rolling Out Autopilot Features, Boomberg Technology, October 14,
2015, accessed August 1, 2016, http://www.bloomberg.com/news/articles/2015-10-14/tesla-
software-upgrade-adds-automated-lane-changing-to-model-s.
29 Aaron Tilley, New Qualcomm Chip Brings Deep Learning To Cars, Forbes, January 5, 2016,
accessed August 1, 2016, http://www.forbes.com/sites/aarontilley/2016/01/05/along-with-nvidia-
new-qualcomm-chip-brings-deep-learning-to-cars/#4cb4e9235357. 19
 make such autonomous driving possible were available in 2000and indeed some
autonomous car prototypes existed30 31 32few predicted that mainstream companies
would be developing and deploying autonomous cars by 2015. During the first
Defense Advanced Research Projects Agency (DARPA) grand challenge on
autonomous driving in 2004, research teams failed to complete the challenge in a
limited desert setting.
But in eight short years, from 2004-2012, speedy and surprising progress occurred
in both academia and industry. Advances in sensing technology and machine learning
We will see self-driving and for perception tasks has sped progress and, as a result, Googles autonomous vehicles
and Teslas semi-autonomous cars are driving on city streets today. Googles self-
remotely controlled delivery driving cars, which have logged more than 1,500,000 miles (300,000 miles without an
accident),33 are completely autonomousno human input needed. Tesla has widely
vehicles, flying vehicles,
released self-driving capability to existing cars with a software update.34 Their cars are
and trucks. Peer-to-peer semi-autonomous, with human drivers expected to stay engaged and take over if they
detect a potential problem. It is not yet clear whether this semi-autonomous approach
transportation services is sustainable, since as people become more confident in the cars capabilities, they
are likely to pay less attention to the road, and become less reliable when they are
such as ridesharing are most needed. The first traffic fatality involving an autonomous car, which occurred in
also likely to utilize self- June of 2016, brought this question into sharper focus.35
In the near future, sensing algorithms will achieve super-human performance for
driving vehicles. capabilities required for driving. Automated perception, including vision, is already
near or at human-level performance for well-defined tasks such as recognition and
tracking. Advances in perception will be followed by algorithmic improvements
in higher level reasoning capabilities such as planning. A recent report predicts
self-driving cars to be widely adopted by 2020.36 And the adoption of self-driving
capabilities wont be limited to personal transportation. We will see self-driving and
remotely controlled delivery vehicles, flying vehicles, and trucks. Peer-to-peer
transportation services such as ridesharing are also likely to utilize self-driving vehicles.
Beyond self-driving cars, advances in robotics will facilitate the creation and adoption
of other types of autonomous vehicles, including robots and drones.
It is not yet clear how much better self-driving cars need to become to encourage
broad acceptance. The collaboration required in semi-self-driving cars and its
implications for the cognitive load of human drivers is not well understood. But
if future self-driving cars are adopted with the predicted speed, and they exceed
human-level performance in driving, other significant societal changes will follow.
Self-driving cars will eliminate one of the biggest causes of accidental death and
injury in United States, and lengthen peoples life expectancy. On average, a

30 Navlab, Wikipedia, last updated June 4, 2016, accessed August 1, 2016, https://en.wikipedia.
org/wiki/Navlab.
31 Navlab: The Carnegie Mellon University Navigation Laboratory, Carnegie Mellon
University, accessed August 1, 2016, http://www.cs.cmu.edu/afs/cs/project/alv/www/.
32 Eureka Prometheus Project, Wikipedia, last modified February 12, 2016, accessed August 1,
2016, https://en.wikipedia.org/wiki/Eureka_Prometheus_Project.
33 Google Self-Driving Car Project, Google, accessed August 1, 2016, https://www.google.
com/selfdrivingcar/. 33 Molly McHugh, Teslas Cars Now Drive Themselves, Kinda, Wired,
October 14, 2015, accessed August 1, 2016, http://www.wired.com/2015/10/tesla-self-driving-over-
air-update-live/.
34 Molly McHugh, Teslas Cars Now Drive Themselves, Kinda, Wired, October 14, 2015,
accessed August 1, 2016, http://www.wired.com/2015/10/tesla-self-driving-over-air-update-live/.
35 Anjali Singhvi and Karl Russell, Inside the Self-Driving Tesla Fatal Accident, The New
York Times, Last updated July 12, 2016, accessed August 1, 2016, http://www.nytimes.com/
interactive/2016/07/01/business/inside-tesla-accident.html.
36 John Greenough, 10 million self-driving cars will be on the road by 2020, Business Insider,
June 15, 2016, accessed August 1, 2016, http://www.businessinsider.com/report-10-million-self-
20 driving-cars-will-be-on-the-road-by-2020-2015-5-6.
commuter in US spends twenty-five minutes driving each way.37 With self-driving car
technology, people will have more time to work or entertain themselves during their
commutes. And the increased comfort and decreased cognitive load with self-driving
cars and shared transportation may affect where people choose to live. The reduced
need for parking may affect the way cities and public spaces are designed. Self-driving
cars may also serve to increase the freedom and mobility of different subgroups of
the population, including youth, elderly and disabled.
Self-driving cars and peer-to-peer transportation services may eliminate the need
to own a vehicle. The effect on total car use is hard to predict. Trips of empty vehicles Shared transportation
and peoples increased willingness to travel may lead to more total miles driven.
Alternatively, shared autonomous vehiclespeople using cars as a service rather than may displace the need for
owning their ownmay reduce total miles, especially if combined with well-constructed
public transportation
incentives, such as tolls or discounts, to spread out travel demand, share trips, and
reduce congestion. The availability of shared transportation may displace the need or public transportation
for public transportationor public transportation may change form towards
personal rapid transit, already available in four cities,38 which uses small capacity may change form towards
vehicles to transport people on demand and point-to-point between many stations.39
As autonomous vehicles become more widespread, questions will arise over their
personal rapid transit
security, including how to ensure that technologies are safe and properly tested that uses small capacity
under different road conditions prior to their release. Autonomous vehicles and the
connected transportation infrastructure will create a new venue for hackers to exploit vehicles to transport
vulnerabilities to attack. Ethical questions are also involved in programming cars to
act in situations in which human injury or death is inevitable, especially when there people on demand.
are split-second choices to be made about whom to put at risk. The legal systems in
most states in the US do not have rules covering self-driving cars. As of 2016, four
states in the US (Nevada, Florida, California, and Michigan), Ontario in Canada,
the United Kingdom, France, and Switzerland have passed rules for the testing
of self-driving cars on public roads. Even these laws do not address issues about
responsibility and assignment of blame for an accident for self-driving and semi-self-
driving cars.40

Transportation planning
By 2005, cities had started investing in the transportation infrastructure to develop
sensing capabilities for vehicle and pedestrian traffic.41 The sensors currently used
include inductive loops, video cameras, remote traffic microwave sensors, radars, and
GPS.42 For example, in 2013 New York started using a combination of microwave
sensors, a network of cameras, and pass readers to detect vehicle traffic in the city.43
37 Brian McKenzie and Melanie Rapino, Commuting in the United States: 2009, American
Community Survey Reports, United States Census Bureau, September 2011, accessed August 1, 2016,
https://www.census.gov/prod/2011pubs/acs-15.pdf.
38 Morgantown, West Virginia; Masdar City, UAE; London, England; and Suncheon, South
Korea.
39 Personal rapid transit, Wikipedia, Last modified July 18, 2016, accessed August 1, 2016,
https://en.wikipedia.org/wiki/Personal_rapid_transit.
40 Patrick Lin, The Ethics of Autonomous Cars, The Atlantic, October 8, 2013, accessed
August 1, 2016, http://www.theatlantic.com/technology/archive/2013/10/the-ethics-of-
autonomous-cars/280360/.
41 Steve Lohr, Bringing Efficiency to the Infrastructure, The New York Times, April 29,
2009, accessed August 1, 2016, http://www.nytimes.com/2009/04/30/business/energy-
environment/30smart.html.
42 Intelligent transportation system, Wikipedia, last modified July 28, 2016, accessed August 1, 2016,
https://en.wikipedia.org/wiki/Intelligent_transportation_system.
43 Access Science Editors, Active traffic management: adaptive traffic signal control, Access Science,
2014, accessed August 1, 2016, http://www.accessscience.com/content/active-traffic-management-
adaptive-traffic-signal-control/BR0106141. 21
 Cities use AI methods to optimize services in several ways, such as bus and subway
schedules, and tracking traffic conditions to dynamically adjust speed limits or apply
smart pricing on highways, bridges, and HOV lanes.44 45 46 Using sensors and cameras
in the road network, they can also optimize traffic light timing for improving traffic
flow and to help with automated enforcement.47 48 These dynamic strategies are aimed
at better utilizing the limited resources in the transportation network, and are made
possible by the availability of data and the widespread connectivity of individuals.
Before the 2000s, transportation planners were forced to rely on static pricing
Ethical questions arise strategies tied to particular days or times of day, to manage demand. As dynamic
pricing strategies are adopted, this raises new issues concerning the fair distribution
when programming cars to of public goods, since market conditions in high-demand situations may make
services unavailable to segments of the public.
act in situations in which
The availability of large-scale data has also made transportation an ideal domain
human injury or death is for machine learning applications. Since 2006, applications such as Mapquest, Google
Maps, and Bing Maps have been widely used by the public for routing trips, using public
inevitable, especially when transportation, receiving real-time information and predictions about traffic conditions,
49 50
and finding services around a location.51 52 Optimal search algorithms have been
there are split-second applied to the routing of vehicles and pedestrians to a given destination (i.e.,53 54).
choices to be made about Despite these advances, the widespread application of sensing and optimization
techniques to city infrastructure has been slower than the application of these
whom to put at risk. techniques to individual vehicles or people. Although individual cities have
implemented sensing and optimization applications, as yet there is no standardization
of the sensing infrastructure and AI techniques used. Infrastructure costs, differing
priorities among cities, and the high coordination costs among the parties involved
have slowed adoption, as have public concerns over privacy related to sensing. Still,

44 Kitae Jang, Koohong Chung, and Hwasoo Yeo, A Dynamic Pricing Strategy for High
Occupancy Toll Lanes, Transportation Research Part A: Policy and Practice 67 (2014): 6980.
45 Seattle Variable Tolling Study, City of Seattle Department of Transportation, May 2009,
accessed August 1, 2016, http://www.seattle.gov/transportation/docs/FINAL%20Tolling%20
Study%20report%20revised%206.25.10.pdf.
46 James F. Peltz, Dynamic Pricing Is Catching On in the Public and Private Sectors,
Government Technology, March 21, 2016, accessed August 1, 2016, http://www.govtech.com/budget-
finance/Dynamic-Pricing-Is-Catching-On-in-the-Public-and-Private-Sectors.html.
47 Arthur G Sims and Kenneth W. Dobinson. The Sydney Coordinated Adaptive Traffic
(SCAT) System Philosophy and Benefits. IEEE Transactions on Vehicular Technology 29, no. 2 (1980):
130137.
48 New York City Launches Nations Most Sophisticated Active Traffic Management System
Powered by TransCores TransSuite Traffic Management Software and RFID Technology,
Business Wire, September 27, 2009, accessed August 1, 2016, http://www.businesswire.com/news/
home/20110927005530/en/York-City-Launches-Nation%E2%80%99s-Sophisticated-Active-Traffic.
49 Eric Horvitz, Johnson Apacible, Raman Sarin, and Lin Liao, Prediction, Expectation, and
Surprise: Methods, Designs, and Study of a Deployed Traffic Forecasting Service, Proceedings of the
Twenty-First Conference on Uncertainty and Artificial Intelligence (2005) (Arlington, Virginia: AUAI Press,
July 2005), 275283.
50 Timothy Hunter, Ryan Herring, Pieter Abbeel, and Alexandre Bayen, Path and Travel Time
Inference from GPS Probe Vehicle Data, NIPS Analyzing Networks and Learning with Graphs 12, no. 1 (2009).
51 John Krumm and Eric Horvitz, Predestination: Inferring Destinations from Partial
Trajectories, UbiComp 2006: Ubiquitous Computing, Proceedings of the 8th International Conference,
September 2006, (Springer Berlin, Heidelberg, 2006), 243260.
52 Jill Duffy, Get Organized: Using Location-Based Reminders, PC Magazine, June 30, 2014,
accessed August 1, 2016, http://www.pcmag.com/article2/0,2817,2460207,00.asp.
53 Robert J. Szczerba, Peggy Galkowski, I. S. Glicktein, and Noah Ternullo. Robust Algorithm
for Real-time Route Planning, IEEE Transactions on Aerospace and Electronic Systems 36, no. 3 (2000):
869878.
54 Matt Duckham and Lars Kulik, Simplest Paths: Automated Route Selection for
Navigation, Spatial Information Theory. Foundations of Geographic Information Science,
Proceedings of the International Conference, COSIT 2003, September 2003 (Springer-Verlag
22 Berlin Heidelberg, 2003), 169-185.
AI is likely to have an increasing impact on city infrastructure. Accurate predictive
models of individuals movements, their preferences, and their goals are likely to
emerge with the greater availability of data. The ethical issues regarding such an
emergence are discussed in Section III of this report.
The United States Department of Transportation released a call for proposals in
2016 asking medium-size cities to imagine smart city infrastructure for transportation.55
This initiative plans to award forty million dollars to a city to demonstrate how
technology and data can be used to reimagine the movement of people as well as goods.
One vision is a network of connected vehicles that can reach a high level of safety Our Study Panel doesnt
in driving with car-to-car communication.56 If this vision becomes reality, we expect
advances in multi-agent coordination, collaboration, and planning will have a expect drones that can
significant impact on future cars and play a role in making the transportation system
fly, swim, and drive, or
more reliable and efficient. Robots are also likely to take part in transportation by
carrying individuals and packages (c.f., Segway robot). For transportation of goods, flying quadcoptors to
interest in drones has increased, and Amazon is currently testing a delivery system using
them,57 although questions remain about the appropriate safety rules and regulations. become a common means
The increased sensing capabilities, adoption of drones, and the connected
transportation infrastructure will also raise concerns about the privacy of individuals
of transportation by 2030
and the safety of private data. In coming years, these and related transportation issues (although prototypes exist
will need to be addressed either by preemptive action on the part of industry or within
the legal framework. As noted in the Section III policy discussion, how well this is done today).
will affect the pace and scope of AI-related advances in the transportation sector.

On-demand transportation
On-demand transportation services such as Uber and Lyft have emerged as another
pivotal application of sensing, connectivity, and AI,58 with algorithms for matching
drivers to passengers by location and suitability (reputation modeling).59 60
Through dynamic pricing, these services ration access by willingness-to-pay, with
dynamic pricing also encouraging an increase in the supply of drivers, and have
become a popular method for transportation in cities. With their rapid advance have
come multiple policy and legal issues, such as competition with existing taxi services
and concerns about lack of regulation and safety. On-demand transportation services
seem likely to be a major force towards self-driving cars.
Carpooling and ridesharing have long been seen as a promising approach to
decrease traffic congestion and better utilize personal transportation resources.
Services such as Zimride and Nuride bring together people sharing similar routes for a
joint trip. But this approach to carpooling has failed to gain traction on a large scale.

55 U.S. Department of Transportation Launches Smart City Challenge to Create a City of

the Future, Transportation.gov, U.S. Department of Transportation, December 7, 2015, accessed
August 1, 2016, https://www.transportation.gov/briefing-room/us-department-transportation-
launches-smart-city-challenge-create-city-future.
56 Will Knight, Car-to-Car Communication: A simple wireless technology promises to
make driving much safer., MIT Technology Review, accessed August 1, 2016, https://www.
technologyreview.com/s/534981/car-to-car-communication/.
57 Amazon Prime Air, Amazon, accessed August 1, 2016, http://www.amazon.com/
b?node=8037720011.
58 Jared Meyer, Uber and Lyft are changing the way Americans move about their country,
National Review, June 7, 2016, accessed August 1, 2016, http://www.nationalreview.com/
article/436263/uber-lyft-ride-sharing-services-sharing-economy-are-future.
59 Alexander Howard, How Digital Platforms Like LinkedIn, Uber And TaskRabbit Are
Changing The On-Demand Economy, The Huffington Post, July 14, 2015, accessed August 1, 2016,
http://www.huffingtonpost.com/entry/online-talent-platforms_us_55a03545e4b0b8145f72ccf6.
60 Announcing UberPool, Uber Newsroom, August 5, 2014, accessed August 1, 2016, https://
newsroom.uber.com/announcing-uberpool/. 23
 Interacting with people
For decades, people have imagined wildly different, futuristic-looking transportation
vehicles. Although future cars will be smarter and drones will be available widely, it
is unlikely that by 2030 we will have widely adopted transportation vehicles that look
and function differently than the ones we have today. Our Study Panel doesnt expect
drones that can fly, swim, and drive, or flying quadcoptors to become a common
means of transportation in this time horizon (although prototypes exist today).
We do expect humans to become partners to self-driving cars and drones in their
Over the next fifteen years, training, execution, and evaluation. This partnering will happen both when humans
are co-located with machines and also virtually. We predict advances in algorithms to
coincident advances facilitate machine learning from human input. We also expect models and algorithms
in mechanical and AI for modeling of human attention, and to support communication and coordination
between humans and machine. This is an integral part of the development of
technologies promise to future vehicles.

increase the safe and

HOME/SERVICE ROBOTS
reliable use and utility of Robots have entered peoples homes in the past fifteen years. Disappointingly slow
growth in the diversity of applications has occurred simultaneously with increasingly
home robots in a typical sophisticated AI deployed on existing applications. AI advances are often inspired by
North American city. mechanical innovations, which in turn prompt new AI techniques to be introduced.
Over the next fifteen years, coincident advances in mechanical and AI technologies
promise to increase the safe and reliable use and utility of home robots in a typical
North American city. Special purpose robots will deliver packages, clean offices,
and enhance security, but technical constraints and the high costs of reliable
mechanical devices will continue to limit commercial opportunities to narrowly
defined applications for the foreseeable future. As with self-driving cars and other
new transportation machines, the difficulty of creating reliable, market-ready
hardware is not to be underestimated.

Vacuum cleaners
In 2001, after many years of development, the Electrolux Trilobite, a vacuum
cleaning robot, became the first commercial home robot. It had a simple control
system to do obstacle avoidance, and some navigation. A year later, iRobot introduced
Roomba, which was a tenth the price of the Trilobite and, with only 512 bytes of
RAM, ran a behavior based controller. The most intelligent thing it did was to avoid
falling down stairs. Since then, sixteen million Roombas have been deployed all over
the world and several other competing brands now exist.
As the processing power and RAM capacity of low cost embedded processors
improved from its dismal state in the year 2000, the AI capabilities of these robots
also improved dramatically. Simple navigation, self-charging, and actions for dealing
with full dust bins were added, followed by ability to deal with electrical cords and
rug tassels, enabled by a combination of mechanical improvements and sensor
based perception. More recently, the addition of full VSLAM (Visual Simultaneous
Location and Mapping) an AI technology that had been around for twenty years
has enabled the robots to build a complete 3D world model of a house as they clean,
and become more efficient in their cleaning coverage.
Early expectations that many new applications would be found for home robots have
not materialized. Robot vacuum cleaners are restricted to localized flat areas, while real
homes have lots of single steps, and often staircases; there has been very little research
on robot mobility inside real homes. Hardware platforms remain challenging to build,
24 and there are few applications that people want enough to buy. Perceptual algorithms
for functions such as image labeling, and 3D object recognition, while common at AI
conferences, are still only a few years into development as products.

Home robots 2030

Despite the slow growth to date of robots in the home, there are signs that this will
change in the next fifteen years. Corporations such as Amazon Robotics and Uber
are developing large economies of scale using various aggregation technologies. Also:
System in Module (SiM), with a lot of System on Chip (SoC) subsystems, are
now being pushed out the door by phone-chip makers (Qualcomms SnapDragon, Special purpose robots
Samsungs Artik, etc.). These are better than supercomputers of less than ten years
ago with eight or more sixty-four-bit cores, and specialized silicon for cryptography,
will deliver packages,
camera drivers, additional DSPs, and hard silicon for certain perceptual algorithms. clean offices, and enhance
This means that low cost devices will be able to support much more onboard AI than
we have been able to consider over the last fifteen years. security, but technical
Cloud (someone elses computer) is going to enable more rapid release of new
software on home robots, and more sharing of data sets gathered in many different constraints and high
homes, which will in turn feed cloud-based machine learning, and then power
costs will continue to limit
improvements to already deployed robots.
The great advances in speech understanding and image labeling enabled by deep commercial opportunities
learning will enhance robots interactions with people in their homes.
Low cost 3D sensors, driven by gaming platforms, have fueled work on 3D for the foreseeable future.
perception algorithms by thousands of researchers worldwide, which will speed the
development and adoption of home and service robots.
In the past three years, low cost and safe robot arms have been introduced to
hundreds of research labs around the world, sparking a new class of research on
manipulation that will eventually be applicable in the home, perhaps around 2025.
More than half a dozen startups around the world are developing AI-based robots for
the home, for now concentrating mainly on social interaction. New ethics and privacy
issues may surface as a result.

HEALTHCARE
For AI technologies, healthcare has long been viewed as a promising domain.
AI-based applications could improve health outcomes and quality of life for
millions of people in the coming yearsbut only if they gain the trust of doctors,
nurses, and patients, and if policy, regulatory, and commercial obstacles are removed.
Prime applications include clinical decision support, patient monitoring and
coaching, automated devices to assist in surgery or patient care, and management
of healthcare systems. Recent successes, such as mining social media to infer
possible health risks, machine learning to predict patients at risk, and robotics
to support surgery, have expanded a sense of possibility for AI in healthcare.
Improvements in methods for interacting with medical professionals and patients
will be a critical challenge.
As in other domains, data is a key enabler. There has been an immense forward
leap in collecting useful data from personal monitoring devices and mobile apps, from
electronic health records (EHR) in clinical settings and, to a lesser extent, from robots
designed to assist with medical procedures and hospital operations. But using this
data to enable more finely-grained diagnostics and treatments for both individual
patients and patient populations has proved difficult. Research and deployment have
been slowed by outdated regulations and incentive structures. Poor human-computer
interaction methods and the inherent difficulties and risks of implementing
technologies in such a large and complex system have slowed realization of AIs 25
 promise in healthcare.61 The reduction or removal of these obstacles, combined with
innovations still on the horizon, have the potential to significantly improve health
outcomes and quality of life for millions of people in the coming years.

The clinical setting

For decades, the vision of an AI-powered clinicians assistant has been a near clich.
Although there have been successful pilots of AI-related technology in healthcare,62
the current healthcare delivery system unfortunately remains structurally ill-suited to
AI-based applications absorb and deploy rapid advances. Incentives provided by the Affordable Care Act
have accelerated the penetration of electronic health records (EHRs) into clinical
could improve health practice, but implementation has been poor, eroding clinicians confidence in their
outcomes and quality of usefulness. A small group of companies control the EHR market, and user interfaces
are widely considered substandard, including annoying pop-ups that physicians
life for millions of people routinely dismiss. The promise of new analytics using data from EHRs, including AI,
remains largely unrealized due to these and other regulatory and structural barriers.
in the coming yearsbut Looking ahead to the next fifteen years, AI advances, if coupled with sufficient data
and well-targeted systems, promise to change the cognitive tasks assigned to human
only if they gain the trust
clinicians. Physicians now routinely solicit verbal descriptions of symptoms from
of doctors, nurses, and presenting patients and, in their heads, correlate patterns against the clinical
presentation of known diseases. With automated assistance, the physician could
patients. instead supervise this process, applying her or his experience and intuition to guide the
input process and to evaluate the output of the machine intelligence. The literal
hands-on experience of the physician will remain critical. A significant challenge is to
optimally integrate the human dimensions of care with automated reasoning processes.
To achieve future advances, clinicians must be involved and engaged at the outset to
ensure that systems are well-engineered and trusted. Already, a new generation of more
tech savvy physicians routinely utilize specialized apps on mobile devices. At the same
time, workloads on primary care clinicians have increased to the point that they are
grateful for help from any quarter. Thus, the opportunity to exploit new learning methods,
to create structured patterns of inference by mining the scientific literature automatically,
and to create true cognitive assistants by supporting free-form dialogue, has never
been greater. Provided these advances are not stymied by regulatory, legal, and social
barriers, immense improvements to the value of healthcare are within our grasp.

Healthcare analytics
At the population level, AIs ability to mine outcomes from millions of patient clinical
records promises to enable finer-grained, more personalized diagnosis and treatment.
Automated discovery of genotype-phenotype connections will also become possible
as full, once-in-a-lifetime genome sequencing becomes routine for each patient.
A related (and perhaps earlier) capability will be to find patients like mine as a way
to inform treatment decisions based on analysis of a similar cohort. Traditional and
non-traditional healthcare data, augmented by social platforms, may lead to the emergence
of self-defined subpopulations, each managed by a surrounding ecosystem of healthcare
providers augmented with automated recommendation and monitoring systems.
These developments have the potential to radically transform healthcare

61 LeighAnne Olsen, Dara Aisner, and J. Michael McGinnis, eds., Institute of Medicine
(US) Roundtable on Evidence-Based Medicine, The Learning Healthcare System: Workshop Summary.
(Washington (DC): National Academies Press (US); 2007), accessed August 1, 2016, http://www.ncbi.
nlm.nih.gov/books/NBK53500/.
62 Katherine E. Henry, David N. Hager, Peter J. Pronovost, and Suchi Saria, A Targeted Real-
time Early Warning Score (TREWScore) for Septic Shock, Science Translational Medicine 7, (299),
26 299ra122.
delivery as medical procedures and lifetime clinical records for hundreds of millions
of individuals become available. Similarly, the automated capture of personal
environmental data from wearable devices will expand personalized medicine. These
activities are becoming more commercially viable as vendors discover ways to engage
large populations (e.g. ShareCare)63 and then to create population-scale data that can
be mined to produce individualized analytics and recommendations.
Unfortunately, the FDA has been slow to approve innovative diagnostic software,
and there are many remaining barriers to rapid innovation. HIPAA (Health Insurance
Portability and Accountability Act) requirements for protecting patient privacy create A small group of companies
legal barriers to the flow of patient data to applications that could utilize AI technologies.
Unanticipated negative effects of approved drugs could show up routinely, sooner, and
control the EHR market,
more rigorously than they do today, but mobile apps that analyze drug interactions and user interfaces
may be blocked from pulling the necessary information from patient records. More
generally, AI research and innovation in healthcare are hampered by the lack of are widely considered
widely accepted methods and standards for privacy protection. The FDA has been
slow to approve innovative software, in part due to an unclear understanding of the substandard, including
cost/benefit tradeoffs of these systems. If regulators (principally the FDA) recognize
that effective post-marketing reporting is a dependable hedge against some safety risks,
annoying pop-ups that
faster initial approval of new treatments and interventions may become possible. physicians routinely
Automated image interpretation has also been a promising subject of study for
decades. Progress on interpreting large archives of weakly-labeled images, such as dismiss.
large photo archives scraped from the web, has been explosive. At first blush, it is
surprising that there has not been a similar revolution in interpretation of medical
images. Most medical imaging modalities (CT, MR, ultrasound) are inherently digital,
the images are all archived, and there are large, established companies with internal
R&D (e.g. Siemens, Philips, GE) devoted to imaging.
But several barriers have limited progress to date. Most hospital image archives
have only gone digital over the past decade. More importantly, the problem in
medicine is not to recognize what is in the image (is this a liver or a kidney?), but
rather to make a fine-grained judgement about it (does the slightly darker smudge
in the liver suggest a potentially cancerous tumor?). Strict regulations govern these
high-stakes judgements. Even with state-of-the-art technologies, a radiologist will still
likely have to look at the images, so the value proposition is not yet compelling. Also,
healthcare regulations preclude easy federation of data across institutions. Thus, only
very large organizations of integrated care, such as Kaiser Permanente, are able to
attack these problems.
Still, automated/augmented image interpretation has started to gain momentum.
The next fifteen years will probably not bring fully automated radiology, but initial forays
into image triage or second level checking will likely improve the speed and cost-
effectiveness of medical imaging. When coupled with electronic patient record systems,
large-scale machine learning techniques could be applied to medical image data. For
example, multiple major healthcare systems have archives of millions of patient scans,
each of which has an associated radiological report, and most have an associated
patient record. Already, papers are appearing in the literature showing that deep neural
networks can be trained to produce basic radiological findings, with high reliability,
by training from this data.64

63 Sharecare, accessed August 1, 2016, https://www.sharecare.com.

64 Hoo-Chang Shin, Holger R. Roth, Mingchen Gao, Le Lu, Ziyue Xu, Isabella Nogues,
Jianhua Yao, Daniel Mollura, and Ronald M. Summers, Deep Convolutional Neural Networks for
Computer-aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning,
IEEE Transactions on Medical Imaging 35, no. 5 (2016): 12851298. 27
 Healthcare robotics
Fifteen years ago, healthcare robotics was largely science fiction. One company
called Robodoc,65 a spin-out from IBM, developed robotic systems for orthopedic
surgeries, such as hip and knee replacements. The technology worked, but the
company struggled commercially, and was ultimately shut down and acquired for
its technology.66 More recently, though, the research and practical use of surgical
robotics has exploded.
In 2000 Intuitive Surgical67 introduced the da Vinci system, a novel technology
The problem in medicine initially marketed to support minimally invasive heart bypass surgery, and then gained
substantial market traction for treatment of prostate cancer and merged with its
is not to recognize what is only major competition, Computer Motion, in 2003. The da Vinci, now in its fourth
in the imageis this a liver generation, provides 3D visualization (as opposed to 2D monocular laparoscopy) and
wristed instruments in an ergonomic platform. It is considered the standard of care
or a kidney?but rather in multiple laparoscopic procedures, and used in nearly three quarters of a million
procedures a year,68 providing not only a physical platform, but also a new data
to make a fine-grained platform for studying the process of surgery.
The da Vinci anticipates a day when much greater insight into how medical
judgement about it. Strict
professionals carry out the process of providing interventional medical care will
regulations govern these be possible. The presence of the da Vinci in day-to-day operation has also opened
the doors to new types of innovationfrom new instrumentation to image fusion
high-stakes judgements. to novel biomarkerscreating its own innovation ecosystem. The success of the
platform has inspired potential competitors in robotic surgery, most notably the
Alphabet spin-off Verb, in collaboration with J&J/Ethicon.69 There are likely to be
many more, each exploring a unique niche or space and building out an ecosystem of
sensing, data analytics, augmentation, and automation.
Intelligent automation in hospital operations has been less successful. The story
is not unlike surgical robotics. Twenty years ago, one company, HelpMate, created
a robot for hospital deliveries,70 such as meals and medical records, but ultimately
went bankrupt. More recently, Aethon71 introduced TUG Robots for basic deliveries,
but few hospitals have invested in this technology to date. However, robotics in
other service industries such as hotels and warehouses, including Amazon Robotics
(formerly Kiva), are demonstrating that these technologies are practical and cost
effective in at least some large-scale settings, and may ultimately spur additional
innovation in health care.
Looking ahead, many tasks that appear in healthcare will be amenable to
augmentation, but will not be fully automated. For example, robots may be able to
deliver goods to the right room in a hospital, but then require a person to pick them
up and place them in their final location. Walking a patient down the corridor may

65 ROBODOC, accessed August 1, 2016, http://www.robodoc.com/professionals.html.

66 THINK Surgical, accessed August 1, 2016, http://thinksurgical.com/history.
67 Intuitive Surgical, accessed August 1, 2016, http://www.intuitivesurgical.com.
68 Trefis Team, Intuitive Surgical Maintains Its Growth Momentum With Strong Growth In
Procedure Volumes, Forbes, January 22, 2016, accessed August 1, 2016, http://www.forbes.com/
sites/greatspeculations/2016/01/22/intuitive-surgical-maintains-its-growth-momentum-with-
strong-growth-in-procedure-volumes/#22ae6b0939a1.
69 Evan Ackerman, Google and Johnson & Johnson Conjugate to Create Verb Surgical,
Promise Fancy Medical Robots, IEEE Spectrum, December 17, 2015, accessed August 1, 2016,
http://spectrum.ieee.org/automaton/robotics/medical-robots/google-verily-johnson-johnson-
verb-surgical-medical-robots.
70 John M. Evans and Bala Krishnamurthy, HelpMate, the trackless robotic courier: A
perspective on the development of a commercial autonomous mobile robot, Lecture Notes in Control
and Information Sciences 236, June 18, 2005 (Springer-Verlag London Limited, 1998), 182210,
accessed August 1, 2016, http://link.springer.com/chapter/10.1007%2FBFb0030806.
28 71 Aethon, accessed August 1, 2016, http://www.aethon.com.
be relatively simple once a patient is standing in a walker (though will certainly not
be trivial for patients recovering from surgery and/or elderly patients, especially in
corridors crowded with equipment and other people). Driving a needle to place a
suture is relatively straightforward once the needle is correctly placed.72 This implies
that many future systems will involve intimate interaction between people and
machines and require technologies that facilitate collaboration between them.
The growth of automation will enable new insights into healthcare processes.
Historically, robotics has not been a strongly data-driven or data-oriented science.
This is changing as (semi)automation infiltrates healthcare. As the new surgical, Specialized motion
delivery, and patient care platforms come online, the beginnings of quantification
and predictive analytics are being built on top of data coming from these platforms.73 tracking devices... and
This data will be used to assess quality of performance, identify deficiencies, errors,
the emerging (inter)
or potential optimizations, and will be used as feedback to improve performance. In
short, these platforms will facilitate making the connection between what is done, and connectedness between
the outcome achieved, making true closed-loop medicine a real possibility.
the home environment and
Mobile health
To date, evidence-driven analytics on healthcare have relied on traditional healthcare
health-monitoring devices
datamainly the electronic medical records discussed above. In the clinical setting, have created a vibrant new
there are hopeful trends towards bringing new data to bear. For example, Tele-
Language enables a human clinician to conduct language therapy sessions with sector of innovation.
multiple patients simultaneously with the aid of an AI agent trained by the clinician.
And Lifegraph, which extracts behavioral patterns and creates alerts from data
passively collected from a patients smartphone, has been adopted by psychiatrists in
Israel to detect early signs of distressful behavior in patients.
Looking ahead, driven by the mobile computing revolution, the astonishing growth
of biometrics in the wildand the explosion of platforms and applications that use
themis a hopeful and unanticipated trend. Thousands of mobile apps now offer
information, introduce behavior modification, or identify groups of people like me.
This, combined with the emerging trend of more specialized motion tracking devices,
such as Fitbit, and the emerging (inter)connectedness between the home environment
and health-monitoring devices, has created a vibrant new sector of innovation.
By combining social and healthcare data, some healthcare apps can perform data
mining, learning, and prediction from captured data, though their predictions are
relatively rudimentary. The convergence of data and functionality across applications
will likely spur new and even obvious products, such as an exercise app that not only
proposes a schedule for exercise but also suggests the best time to do it, and provides
coaching to stick to that schedule.

72 Azad Shademan, Ryan S. Decker, Justin D. Opfermann, Simon Leonard, Axel Krieger, and
Peter CW Kim, Supervised Autonomous Robotic Soft Tissue Surgery, Science Translational Medicine
8, no. 337 (2016): 337ra64337ra64.
73 Carolyn Chen, Lee White, Timothy Kowalewski, Rajesh Aggarwal, Chris Lintott, Bryan
Comstock, Katie Kuksenok, Cecilia Aragon, Daniel Holst, and Thomas Lendvay, Crowd-Sourced
Assessment of Technical Skills: a novel method to evaluate surgical performance. Journal of Surgical
Research 187, no. 1 (2014): 6571.
29
 Elder care
Over the next fifteen years the number of elderly in the United States will grow
by over 50%.74 The National Bureau of Labor Statistics projects that home health
aides will grow 38% over the next ten years. Despite the broad opportunities in this
domainbasic social support, interaction and communication devices, home health
monitoring, a variety of simple in-home physical aids such as walkers, and light meal
preparationlittle has happened over the past fifteen years. But the coming generational
shift will accompany a change in technology acceptance among the elderly. Currently,
Better hearing aids and someone who is seventy was born in 1946 and may have first experienced some form
of personalized IT in middle age or later, while a fifty-year-old today is far more
visual assistive devices technology-friendly and savvy. As a result, there will be a growing interest and market
will mitigate the effects for already available and maturing technologies to support physical, emotional, social,
and mental health. Here are a few likely examples by category:
of hearing and vision Life quality and independence
Automated transportation will support continued independence and expanded
loss, improving safety social horizons.
Sharing of information will help families remain engaged with one another at a
and social connection.
distance, and predictive analytics may be used to nudge family groups toward
Personalized rehabilitation positive behaviors, such as reminders to call home.
Smart devices in the home will help with daily living activities when needed,
and in-home therapy such as cooking and, if robot manipulation capabilities improve sufficiently,
dressing and toileting.
will reduce the need for
Health and wellness
hospital stays. Mobile applications that monitor movement and activities, coupled with social
platforms, will be able to make recommendations to maintain mental and
physical health.
In-home health monitoring and health information access will be able to detect
changes in mood or behavior and alert caregivers.
Personalized health management will help mitigate the complexities associated
with multiple co-morbid conditions and/or treatment interactions.
Treatments and devices
Better hearing aids and visual assistive devices will mitigate the effects of
hearing and vision loss, improving safety and social connection.
Personalized rehabilitation and in-home therapy will reduce the need for
hospital or care facility stays.
Physical assistive devices (intelligent walkers, wheel chairs, and exoskeletons) will
extend the range of activities of an infirm individual.
The Study Panel expects an explosion of low-cost sensing technologies that can
provide substantial capabilities to the elderly in their homes. In principle, social
agents with a physical presence and simple physical capabilities (e.g. a mobile robot
with basic communication capabilities) could provide a platform for new innovations.
However, doing so will require integration across multiple areas of AINatural
Language Processing, reasoning, learning, perception, and roboticsto create
a system that is useful and usable by the elderly.
These innovations will also introduce questions regarding privacy within various
circles, including friends, family, and care-givers, and create new challenges to
accommodate an evermore active and engaged population far past retirement.

74 Jennifer M. Ortman, Victoria A. Velkoff, and Howard Hogan, An Aging Nation: The
Older Population in the United States: Population Estimates and Projections, Current Population
Reports, U.S Census Bureau (May 2014), accessed August 1, 2016, https://www.census.gov/
prod/2014pubs/p25-1140.pdf.
30
EDUCATION
The past fifteen years have seen considerable AI advances in education. Applications
are in wide use by educators and learners today, with some variation between K-12
and university settings. Though quality education will always require active engagement
by human teachers, AI promises to enhance education at all levels, especially by
providing personalization at scale. Similar to healthcare, resolving how to best
integrate human interaction and face-to-face learning with promising AI technologies
remains a key challenge.
Though quality education
Robots have long been popular educational devices, starting with the early
Lego Mindstorms kits developed with the MIT Media Lab in the 1980s. Intelligent will always require active
Tutoring Systems (ITS) for science, math, language, and other disciplines match
students with interactive machine tutors. Natural Language Processing, especially when engagement by human
combined with machine learning and crowdsourcing, has boosted online learning
and enabled teachers to multiply the size of their classrooms while simultaneously
teachers, AI promises to
addressing individual students learning needs and styles. The data sets from large enhance education at
online learning systems have fueled rapid growth in learning analytics.
Still, schools and universities have been slow in adopting AI technologies primarily all levels, especially by
due to lack of funds and lack of solid evidence that they help students achieve
learning objectives. Over the next fifteen years in a typical North American city, the providing personalization
use of intelligent tutors and other AI technologies to assist teachers in the classroom
at scale.
and in the home is likely to expand significantly, as will learning based on virtual
reality applications. But computer-based learning systems are not likely to fully
replace human teaching in schools.

Teaching robots
Today, more sophisticated and versatile kits for use in K-12 schools are available
from a number of companies that create robots with new sensing technologies
programmable in a variety of languages. Ozobot is a robot that teaches children
to code and reason deductively while configuring it to dance or play based on
color-coded patterns.75 Cubelets help teach children logical thinking through
assembling robot blocks to think, act, or sense, depending upon the function of the
different blocks.76 Wonder Workshops Dash and Dot span a range of programming
capabilities. Children eight years old and older can create simple actions using a
visual programming language, Blockly, or build iOS and Android applications using
C or Java.77 PLEO rb is a robot pet that helps children learn biology by teaching
the robot to react to different aspects of the environment.78 However, while fun and
engaging for some, in order for such kits to become widespread, there will need to be
compelling evidence that they improve students academic performance.

Intelligent Tutoring Systems (ITS) and online learning

ITS have been developed from research laboratory projects such as Why-2 Atlas,
which supported human-machine dialogue to solve physics problems early in the
era.79 The rapid migration of ITS from laboratory experimental stages to real use is

75 Ozobot, accessed August 1, 2016, http://ozobot.com/.

76 Cubelets, Modular Robotics, accessed August 1, 2016, http://www.modrobotics.com/
cubelets.
77 Meet Dash, Wonder Workshop, accessed August 1, 2016, https://www.makewonder.com/
dash.
78 Pleo rb, Innvo Labs, accessed August 1, 2016, http://www.pleoworld.com/pleo_rb/eng/
lifeform.php.
79 Kurt VanLehn, Pamela W. Jordan, Carolyn P. Ros, Dumisizwe Bhembe, Michael Bttner,
Andy Gaydos, Maxim Makatchev, Umarani Pappuswamy, Michael Ringenberg, Antonio Roque,
Stephanie Siler, and Ramesh Srivastava, The Architecture of Why2-Atlas: A Coach for Qualitative 31
 surprising and welcome. Downloadable software and online systems such as Carnegie
Speech or Duolingo provide foreign language training using Automatic Speech
Recognition (ASR) and NLP techniques to recognize language errors and help users
correct them.80 Tutoring systems such as the Carnegie Cognitive Tutor81 have been
used in US high schools to help students learn mathematics. Other ITS have been
developed for training in geography, circuits, medical diagnosis, computer literacy
and programming, genetics, and chemistry. Cognitive tutors use software to mimic
the role of a good human tutor by, for example, providing hints when a student gets
It can be argued that AI stuck on a math problem. Based on the hint requested and the answer provided, the
tutor offers context specific feedback.
is the secret sauce that Applications are growing in higher education. An ITS called SHERLOCK82 is
has enabled instructors, beginning to be used to teach Air Force technicians to diagnose electrical systems
problems in aircraft. And the University of Southern Californias Information
particularly in higher Sciences Institute has developed more advanced avatar-based training modules to
train military personnel being sent to international posts in appropriate behavior
education, to multiply the when dealing with people from different cultural backgrounds. New algorithms for
personalized tutoring, such as Bayesian Knowledge Tracing, enable individualized
size of their classrooms
mastery learning and problem sequencing.83
by a few orders of Most surprising has been the explosion of the Massive Open Online Courses
(MOOCs) and other models of online education at all levelsincluding the use
magnitudeclass sizes of of tools like Wikipedia and Khan Academy as well as sophisticated learning
management systems that build in synchronous as well as asynchronous education
a few tens of thousands and adaptive learning tools. Since the late 1990s, companies such as the Educational
are not uncommon. Testing Service and Pearson have been developing automatic NLP assessment tools to
co-grade essays in standardized testing.84 Many of the MOOCs which have become
so popular, including those created by EdX, Coursera, and Udacity, are making use
of NLP, machine learning, and crowdsourcing techniques for grading short-answer
and essay questions as well as programming assignments.85 Online education systems
that support graduate-level professional education and lifelong learning are also
expanding rapidly. These systems have great promise because the need for face-to-face
interaction is less important for working professionals and career changers. While not
the leaders in AI-supported systems and applications, they will become early adopters
as the technologies are tested and validated.
It can be argued that AI is the secret sauce that has enabled instructors, particularly
in higher education, to multiply the size of their classrooms by a few orders of
magnitudeclass sizes of a few tens of thousands are not uncommon. In order to
continually test large classes of students, automated generation of the questions is

Physics Essay Writing, Intelligent Tutoring Systems: Proceedings of the 6th International Conference, (Springer
Berlin Heidelberg, 2002), 158167.
80 VanLehn et al, The Architecture of Why2-Atlas.
81 Resources and Support, Carnegie Learning, accessed August 1, 2016, https://www.
carnegielearning.com/resources-support/.
82 Alan Lesgold, Suzanne Lajoie, Marilyn Bunzo, and Gary Eggan, SHERLOCK: A Coached
Practice Environment for an Electronics Troubleshooting Job, in J. H. Larkin and R. W. Chabay,
eds., Computer-Assisted Instruction and Intelligent Tutoring Systems: Shared Goals and Complementary Approaches
(Hillsdale, New Jersey: Lawrence Erlbaum Associates, 1988).
83 Michael V. Yudelson, Kenneth R. Koedinger, and Geoffrey J. Gordon, (2013). Individualized
Bayesian Knowledge Tracing Models, Artificial Intelligence in Education, (Springer Berlin Heidelberg,
2013), 171180.
84 Jill Burstein, Karen Kukich, Susanne Wolff, Chi Lu, Martin Chodorow, Lisa Braden-Harder,
and Mary Dee Harris, Automated Scoring Using a Hybrid Feature Identification Technique in
Proceedings of the Annual Meeting of the Association of Computational Linguistics, Montreal, Canada, August
1998, accessed August 1, 2016, https://www.ets.org/Media/Research/pdf/erater_acl98.pdf.
85 EdX, https://www.edx.org/, Coursera, https://www.coursera.org/, Udacity, https://www.
32 udacity.com/, all accessed August 1, 2016.
also possible, such as those designed to assess vocabulary,86 wh (who/what/when/
where/why) questions,87 and multiple choice questions,88 using electronic resources
such as WordNet, Wikipedia, and online ontologies. With the explosion of online
courses, these techniques are sure to be eagerly adopted for use in online education.
Although the long term impact of these systems will have on the educational system
remains unclear, the AI community has learned a great deal in a very short time.

Learning analytics
Data sets being collected from massive scale online learning systems, ranging from The current absence of
MOOCs to Khan Academy, as well as smaller scale online programs, have fueled
the rapid growth of the field of learning analytics. Online courses are not only good sophisticated use of AI
for widespread delivery, but are natural vehicles for data collection and experimental
technologies in schools,
instrumentation that will contribute to scientific findings and improving the quality
of learning at scale. Organizations such as the Society for Learning Analytics colleges, and universities
Research (SOLAR), and the rise of conferences including the Learning Analytics and
Knowledge Conference89 and the Learning at Scale Conference (L@S)90 reflect this may be explained by
trend. This community applies deep learning, natural language processing, and other
AI techniques to analysis of student engagement, behavior, and outcomes.
the lack of financial
Current projects seek to model common student misconceptions, predict resources as well as the
which students are at risk of failure, and provide real-time student feedback that is
tightly integrated with learning outcomes. Recent work has also been devoted to lack of data establishing
understanding the cognitive processes involved in comprehension, writing, knowledge
acquisition, and memory, and to applying that understanding to educational practice the technologies
by developing and testing educational technologies.
effectiveness.
Challenges and opportunities
One might have expected more and more sophisticated use of AI technologies in
schools, colleges, and universities by now. Much of its absence can be explained
by the lack of financial resources of these institutions as well as the lack of data
establishing the technologies effectiveness. These problems are being addressed,
albeit slowly, by private foundations and by numerous programs to train primarily
secondary school teachers in summer programs. As in other areas of AI, excessive
hype and promises about the capabilities of MOOCs have meant that expectations
frequently exceed the reality. The experiences of certain institutions, such as San Jose
State Universitys experiment with Udacity,91 have led to more sober assessment of
the potential of the new educational technologies.

86 Jonathan C. Brown, Gwen A. Frishkoff , and Maxine Eskenazi, Automatic Question

Generation for Vocabulary Assessment, Proceedings of Human Language Technology Conference and
Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), Vancouver, October
2005, (Association for Computational Linguistics, 2005), 819826.
87 Michael Heilman, Automatic Factual Question Generation from Text, PhD thesis CMU-
LTI-11-004, (Carnegie Mellon University, 2011), accessed August 1, 2016, http://www.cs.cmu.
edu/~ark/mheilman/questions/papers/heilman-question-generation-dissertation.pdf.
88 Tahani Alsubait, Bijan Parsia, and Uli Sattler, Generating Multiple Choice Questions from
Ontologies: How Far Can We Go?, in eds. P. Lambrix, E. Hyvnen. E. Blomqvist, V. Presutti, G.
Qi, U. Sattler, Y. Ding, and C. Ghidini, Knowledge Engineering and Knowledge Management: EKAW 2014
Satellite Events, VISUAL, EKM1, and ARCOE-Logic Linkping, Sweden, November 2428, 2014 Revised
Selected Papers, (Switzerland: Springer International Publishing, 2015), 6679.
89 The 6th International Learning Analytics & Knowledge Conference, accessed August 1, 2016,
http://lak16.solaresearch.org/.
90 Third Annual ACM Conference on Learning at Scale, http://learningatscale.acm.org/
las2016/.
91 Ry Rivard, Udacity Project on Pause, Inside Higher Ed, July 18, 2013, accessed August 1,
2016, https://www.insidehighered.com/news/2013/07/18/citing-disappointing-student-outcomes-
san-jose-state-pauses-work-udacity. 33
 In the next fifteen years, it is likely that human teachers will be assisted by AI
technologies with better human interaction, both in the classroom and in the home.
The Study Panel expects that more general and more sophisticated virtual reality
scenarios in which students can immerse themselves in subjects from all disciplines
will be developed. Some steps in this direction are being taken now by increasing
collaborations between AI researchers and researchers in the humanities and social
sciences, exemplified by Stanfords Galileo Correspondence Project92 and Columbias
Making and Knowing Project.93 These interdisciplinary efforts create interactive
While formal education experiences with historical documents and the use of Virtual Reality (VR) to explore
interactive archeological sites.94 VR techniques are already being used in the natural
will not disappear, the sciences such as biology, anatomy, geology and astronomy to allow students to interact
Study Panel believes that with environments and objects that are difficult to engage with in the real world. The
recreation of past worlds and fictional worlds will become just as popular for studies
MOOCs and other forms of arts and other sciences.
AI techniques will increasingly blur the line between formal, classroom education
of online education will and self-paced, individual learning. Adaptive learning systems, for example, are going
to become a core part of the teaching process in higher education because of the
become part of learning
pressures to contain cost while serving a larger number of students and moving students
at all levels, from K-12 through school more quickly. While formal education will not disappear, the Study
Panel believes that MOOCs and other forms of online education will become part of
through university, in learning at all levels, from K-12 through university, in a blended classroom experience.
This development will facilitate more customizable approaches to learning, in which
a blended classroom students can learn at their own pace using educational techniques that work best for
experience. them. Online education systems will learn as the students learn, supporting rapid
advances in our understanding of the learning process. Learning analytics, in turn, will
accelerate the development of tools for personalized education.
The current transition from hard copy books to digital and audio media and texts
is likely to become prevalent in education as well. Digital reading devices will also
become much smarter, providing students with easy access to additional information
about subject matter as they study. Machine Translation (MT) technology will
also make it easier to translate educational material into different languages with a
fair degree of accuracy, just as it currently translates technical manuals. Textbook
translation services that currently depend only upon human translators will
increasingly incorporate automatic methods to improve the speed and affordability of
their services for school systems.
Online learning systems will also expand the opportunity for adults and working
professionals to enhance their knowledge and skills (or to retool and learn a new field)
in a world where these fields are evolving rapidly. This will include the expansion
of fully online professional degrees as well as professional certifications based on
online coursework.

Broader societal consequences

In countries where education is difficult for the broad population to obtain, online
resources may have a positive effect if the population has the tools to access them.
The development of online educational resources should make it easier for
foundations that support international educational programs to provide quality
92 Stanford University: Galileo Correspondence Project, accessed August 1, 2016, http://galileo.
stanford.edu.
93 The Making and Knowing Project: Reconstructing the 16th Century Workshop of BNF MS.
FR. 640 at Columbia University, accessed August 1, 2016, http://www.makingandknowing.org.
94 Paul James, 3D Mapped HTC Vive Demo Brings Archaeology to Life, Road to VR, August
31, 2015, accessed August 1, 2016, http://www.roadtovr.com/3d-mapped-htc-vive-demo-brings-
34 archaeology-to-life/.
education by providing tools and relatively simple amounts of training in their use.
For example, large numbers of educational apps, many of them free, are being
developed for the iPad. On the negative side, there is already a major trend among
students to restrict their social contacts to electronic ones and to spend large amounts
of time without social contact, interacting with online programs. If education also
occurs more and more online, what effect will the lack of regular, face-to-face contact
with peers have on students social development? Certain technologies have even
been shown to create neurological side effects.95 On the other hand, autistic children
have benefited from interactions with AI systems already.96 With targeted incentives
and funding priorities, AI
LOW-RESOURCE COMMUNITIES
Many opportunities exist for AI to improve conditions for people in low-resource technologies could help
communities in a typical North American cityand, indeed, in some cases it
address the needs of low-
already has. Understanding these direct AI contributions may also inform potential
contributions in the poorest parts of the developing world. There has not been a resource communities.
significant focus on these populations in AI gatherings, and, traditionally, AI funders
have underinvested in research lacking commercial application. With targeted incentives Budding efforts are
and funding priorities, AI technologies could help address the needs of low-resource
communities. Budding efforts are promising. Counteracting fears that AI may contribute
promising.
to joblessness and other societal problems, AI may provide mitigations and solutions,
particularly if implemented in ways that build trust in them by the affected communities.
Machine learning, data mining approaches
Under the banner of data science for social good, AI has been used to create
predictive models to help government agencies more effectively use their limited
budgets to address problems such as lead poisoning,97 a major public health concern
that has been in the news due to ongoing events in Flint, Michigan. Children may
be tested for elevated lead levels, but that unfortunately means the problem is only
detected after they have already been poisoned. Many efforts are underway to use
predictive models to assist government agencies in prioritizing children at
risk, including those who may not yet have been exposed.98 Similarly, the Illinois
Department of Human Services (IDHS) uses predictive models to identify pregnant
women at risk for adverse birth outcomes in order to maximize the impact of
prenatal care. The City of Cincinnati uses them to proactively identify and deploy
inspectors to properties at risk of code violations.
Scheduling, planning
Task assignment scheduling and planning techniques have been applied by many
different groups to distribute food before it spoils from those who may have excess,
such as restaurants, to food banks, community centers and individuals.99

95 Scientist have studied, for example, the way reliance on GPS may lead to changes in the
hypocampus. Kim Tingley, The Secrets of the Wave Pilots, The New York Times, March 17, 2016,
accessed August 1, 2016, http://www.nytimes.com/2016/03/20/magazine/the-secrets-of-the-wave-
pilots.html.
96 Judith Newman, To Siri, With Love: How One Boy With Autism Became BFF With Apples
Siri, The New York Times, October 17, 2014, accessed August 1, 2016, http://www.nytimes.
com/2014/10/19/fashion/how-apples-siri-became-one-autistic-boys-bff.html.
97 Eric Potash, Joe Brew, Alexander Loewi, Subhabrata Majumdar, Andrew Reece, Joe Walsh,
Eric Rozier, Emile Jorgensen, Raed Mansour, and Rayid Ghani, Predictive Modeling for Public
Health: Preventing Childhood Lead Poisoning, Proceedings of the 21th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining (New York: Association for Computing Machinery,
2015), 20392047.
98 Data Science for Social Good, University of Chicago, accessed August 1, 2016,
http://dssg.uchicago.edu/.
99 Senay Solak, Christina Scherrer, and Ahmed Ghoniem, The Stop-and-Drop Problem in
Nonprofit Food Distribution Networks, Annals of Operations Research 221, no. 1 (October 2014): 35
 Reasoning with social networks and influence maximization
Social networks can be harnessed to create earlier, less-costly interventions involving
large populations. For example, AI might be able to assist in spreading health-
related information. In Los Angeles, there are more than 5,000 homeless youth (ages
thirteen-twenty-four). Individual interventions are difficult and expensive, and the
youths mistrust of authority dictates that key messages are best spread through peer
leaders. AI programs might be able to leverage homeless youth social networks to
strategically select peer leaders to spread health-related information, such as how to
One of the more successful
avoid spread of HIV. The dynamic, uncertain nature of these networks does pose
uses of AI analytics is in challenges for AI research.100 Care must also be taken to prevent AI systems from
reproducing discriminatory behavior, such as machine learning that identifies people
detecting white collar through illegal racial indicators, or through highly-correlated surrogate factors, such
crime, such as credit as zip codes. But if deployed with great care, greater reliance on AI may well result
in a reduction in discrimination overall, since AI programs are inherently more easily
card fraud. Cybersecurity audited than humans.

(including spam) is a PUBLIC SAFETY AND SECURITY

widely shared concern, Cities already have begun to deploy AI technologies for public safety and security.
By 2030, the typical North American city will rely heavily upon them. These include
and machine learning is cameras for surveillance that can detect anomalies pointing to a possible crime,
making an impact. drones, and predictive policing applications. As with most issues, there are benefits
and risks. Gaining public trust is crucial. While there are legitimate concerns that
policing that incorporates AI may become overbearing or pervasive in some contexts,
the opposite is also possible. AI may enable policing to become more targeted and
used only when needed. And assuming careful deployment, AI may also help remove
some of the bias inherent in human decision-making.
One of the more successful uses of AI analytics is in detecting white collar
crime, such as credit card fraud.101 Cybersecurity (including spam) is a widely shared
concern, and machine learning is making an impact. AI tools may also prove useful
in helping police manage crime scenes or search and rescue events by helping
commanders prioritize tasks and allocate resources, though these tools are not yet
ready for automating such activities. Improvements in machine learning in general,
and transfer learning in particularfor speeding up learning in new scenarios based
on similarities with past scenariosmay facilitate such systems.
The cameras deployed almost everywhere in the world today tend to be more
useful for helping solve crimes than preventing them.102 103 This is due to the low
quality of event identification from videos and the lack of manpower to look at
massive video streams. As AI for this domain improves, it will better assist crime
prevention and prosecution through greater accuracy of event classification and
efficient automatic processing of video to detect anomaliesincluding, potentially,

407426.
100 Jordan Pearson, Artificial Intelligence Could Help Reduce HIV Among Homeless Youths,
Teamcore, University of Southern California, February 4. 2015, accessed August 1, 2016, http://
teamcore.usc.edu/news/motherboard_news_ai_could_help_reduce_HIV.pdf.
101 RSA Adaptive Authentication, RSA, accessed August 1, 2016, https://www.rsa.com/en-us/
products-services/fraud-prevention/adaptive-authentication.
102 Takeshi Arikuma and Yasunori Mochizuki, Intelligent multimedia surveillance system for
safer cities APSIPA Transactions on Signal and Information Processing 5 (2016): 18.
103 Big Op-Ed: Shifting Opinions On Surveillance Cameras,, Talk of the Nation, NPR, April 22,
2013, accessed August 1, 2016, http://www.npr.org/2013/04/22/178436355/big-op-ed-shifting-
36 opinions-on-surveillance-cameras.
evidence of police malpractice. These improvements could lead to even more
widespread surveillance. Some cities have already added drones for surveillance
purposes, and police use of drones to maintain security of ports, airports, coastal
areas, waterways, industrial facilities is likely to increase, raising concerns about
privacy, safety, and other issues.
The New York Police Departments CompStat was the first tool pointing toward
predictive policing,104 and many police departments now use it.105 Machine learning
significantly enhances the ability to predict where and when crimes are more likely
to happen and who may commit them. As dramatized in the movie Minority Report, As dramatized in the
predictive policing tools raise the specter of innocent people being unjustifiably
targeted. But well-deployed AI prediction tools have the potential to actually remove movie Minority Report,
or reduce human bias, rather than reinforcing it, and research and resources should
predictive policing tools
be directed toward ensuring this effect.
AI techniques can be used to develop intelligent simulations for training law- raise the specter of
enforcement personnel to collaborate. While international criminal organizations and
terrorists from different countries are colluding, police forces from different countries innocent people being
still face difficulty in joining forces to fight them. Training international groups of law
enforcement personnel to work as teams is very challenging. The European Union,
unjustifiably targeted. But
through the Horizon 2020 program, currently supports such attempts in projects such well-deployed AI prediction
as LawTrain.106 The next step will be to move from simulation to actual investigations
by providing tools that support such collaborations. tools have the potential to
Tools do exist for scanning Twitter and other feeds to look for certain types
of events and how they may impact security. For example, AI can help in social actually remove or reduce
network analysis to prevent those at risk from being radicalized by ISIS or other
human bias.
violent groups. Law enforcement agencies are increasingly interested in trying to
detect plans for disruptive events from social media, and also to monitor activity at
large gatherings of people to analyze security. There is significant work on crowd
simulations to determine how crowds can be controlled. At the same time, legitimate
concerns have been raised about the potential for law enforcement agencies to
overreach and use such tools to violate peoples privacy.
The US Transportation Security Administration (TSA), Coast Guard, and the
many other security agencies that currently rely on AI will likely increase their
reliance to enable significant efficiency and efficacy improvements.107 AI techniques
vision, speech analysis, and gait analysis can aid interviewers, interrogators, and
security guards in detecting possible deception and criminal behavior. For example,
the TSA currently has an ambitious project to redo airport security nationwide.108
Called DARMS, the system is designed to improve efficiency and efficacy of airport
security by relying on personal information to tailor security based on a persons risk
categorization and the flights being taken. The future vision for this project is a tunnel
that checks peoples security while they walk through it. Once again, developers of
this technology should be careful to avoid building in bias (e.g. about a persons risk
level category) through use of datasets that reflect prior bias.109

104 Walter L. Perry, Brian McInnis, Carter C. Price, Susan Smith, and John S. Hollywood, The
Role of Crime Forecasting in Law Enforcement Operations, Rand Corporation Report 233 (2013).
105 CompStat, Wikipedia, last modified July 28, 2016, accessed August 1, 2016, https://
en.wikipedia.org/wiki/CompStat.
106 LAW-TRAIN, accessed August 1, 2016, http://www.law-train.eu/.
107 Milind Tambe, Security and Game Theory: Algorithms, Deployed Systems, Lessons Learned (New York:
Cambridge University Press, 2011).
108 Peter Neffenger, TSAs 2017 BudgetA Commitment to Security (Part I), Department
of Homeland Security, March 1, 2016, accessed August 1, 2016, https://www.tsa.gov/news/
testimony/2016/03/01/hearing-fy17-budget-request-transportation-security-administration.
109 Crawford, AIs White Guy Problem. 37
 EMPLOYMENT AND WORKPLACE
While AI technologies are likely to have a profound future impact on employment
and workplace trends in a typical North American city, it is difficult to accurately
assess current impacts, positive or negative. In the past fifteen years, employment
has shifted due to a major recession and increasing globalization, particularly with
Chinas introduction to the world economy, as well as enormous changes in non-AI
digital technology. Since the 1990s, the US has experienced continued growth in
productivity and GDP, but median income has stagnated and the employment to
AI will likely replace tasks population ratio has fallen.
rather than jobs in the There are clear examples of industries in which digital technologies have had
profound impacts, good and bad, and other sectors in which automation will likely
near term, and will also make major changes in the near future. Many of these changes have been driven
strongly by routine digital technologies, including enterprise resource planning,
create new kinds of jobs. networking, information processing, and search. Understanding these changes should
But the new jobs that provide insights into how AI will affect future labor demand, including the shift in
skill demands. To date, digital technologies have been affecting workers more in the
will emerge are harder to skilled middle, such as travel agents, rather than the very lowest-skilled or highest
skilled work.110 On the other hand, the spectrum of tasks that digital systems can
imagine in advance than do is evolving as AI systems improve, which is likely to gradually increase the scope
of what is considered routine. AI is also creeping into high end of the spectrum,
the existing jobs that will including professional services not historically performed by machines.
likely be lost. To be successful, AI innovations will need to overcome understandable human
fears of being marginalized. AI will likely replace tasks rather than jobs in the near
term, and will also create new kinds of jobs. But the new jobs that will emerge are
harder to imagine in advance than the existing jobs that will likely be lost. Changes
in employment usually happen gradually, often without a sharp transition, a trend
likely to continue as AI slowly moves into the workplace. A spectrum of effects will
emerge, ranging from small amounts of replacement or augmentation to complete
replacement. For example, although most of a lawyers job is not yet automated,111
AI applied to legal information extraction and topic modeling has automated parts of
first-year lawyers jobs.112 In the not too distant future, a diverse array of job-holders,
from radiologists to truck drivers to gardeners, may be affected.
AI may also influence the size and location of the workforce. Many organizations
and institutions are large because they perform functions that can be scaled only by
adding human labor, either horizontally across geographical areas or vertically
in management hierarchies. As AI takes over many functions, scalability no longer
implies large organizations. Many have noted the small number of employees of
some high profile internet companies, but not of others. There may be a natural scale
of human enterprise, perhaps where the CEO can know everyone in the company.
Through the creation of efficiently outsourced labor markets enabled by AI,
enterprises may tend towards that natural size.
AI will also create jobs, especially in some sectors, by making certain tasks more
important, and create new categories of employment by making new modes of
interaction possible. Sophisticated information systems can be used to create new
110 Jeremy Ashkenas and Alicia Parlapiano, How the Recession Reshaped the Economy, in 255
Charts, The New York Times, June 6, 2014, accessed August 1, 2016, http://www.nytimes.com/
interactive/2014/06/05/upshot/how-the-recession-reshaped-the-economy-in-255-charts.html.
111 R Dana Remus and Frank S. Levy, Can Robots Be Lawyers? Computers, Lawyers, and the
Practice of Law, Social Science Research Network, last modified February 12, 2016, accessed August 1,
2016, http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2701092.
112 John Markoff, Armies of Expensive Lawyers, Replaced by Cheaper Software, The New
York Times, March 4, 2011, accessed August 1, 2016, http://www.nytimes.com/2011/03/05/
38 science/05legal.html.
markets, which often have the effect of lowering barriers to entry and increasing
participationfrom app stores to AirBnB to taskrabbit. A vibrant research
community within AI studies further ways of creating new markets and making
existing ones operate more efficiently.
While work has intrinsic value, most people work to be able to purchase goods and
services they value. Because AI systems perform work that previously required human
labor, they have the effect of lowering the cost of many goods and services, effectively
making everyone richer. But as exemplified in current political debates, job loss is
more salient to peopleespecially those directly affectedthan diffuse economic As labor becomes a
gains, and AI unfortunately is often framed as a threat to jobs rather than a boon to
living standards. less important factor in
There is even fear in some quarters that advances in AI will be so rapid as
production as compared
to replace all human jobsincluding those that are largely cognitive or involve
judgmentwithin a single generation. This sudden scenario is highly unlikely, but to owning intellectual
AI will gradually invade almost all employment sectors, requiring a shift away from
human labor that computers are able to take over. capital, a majority of
The economic effects of AI on cognitive human jobs will be analogous to
the effects of automation and robotics on humans in manufacturing jobs. Many
citizens may find the value
middle-aged workers have lost well-paying factory jobs and the socio-economic status of their labor insufficient
in family and society that traditionally went with such jobs. An even larger fraction
of the total workforce may, in the long run, lose well-paying cognitive jobs. As labor to pay for a socially
becomes a less important factor in production as compared to owning intellectual
capital, a majority of citizens may find the value of their labor insufficient to pay for acceptable standard of
a socially acceptable standard of living. These changes will require a political, rather
living.
than a purely economic, response concerning what kind of social safety nets should
be in place to protect people from large, structural shifts in the economy. Absent
mitigating policies, the beneficiaries of these shifts may be a small group at the
upper stratum of the society.113
In the short run, education, re-training, and inventing new goods and services may
mitigate these effects. Longer term, the current social safety net may need to evolve into
better social services for everyone, such as healthcare and education, or a guaranteed
basic income. Indeed, countries such as Switzerland and Finland have actively considered
such measures. AI may be thought of as a radically different mechanism of wealth
creation in which everyone should be entitled to a portion of the worlds AI-produced
treasure.114 It is not too soon for social debate on how the economic fruits of
AI-technologies should be shared. As children in traditional societies support their
aging parents, perhaps our artificially intelligent children should support us, the
parents of their intelligence.

113 For example, Brynjolfsson and McAfee, Second Machine Age, have two chapters of devoted to
this (Erik Brynjolfsson and Andrew McAfee, The Second Machine Age: Work, Progress, and Prosperity in a
Time of Brilliant Technologies, (New York: W. W. Norton & Company, Inc., 2014)) and Brynjolfsson,
McAfee, and Spence describe policy responses for the combination of globalization and digital
technology (Erik Brynjolfsson, Andrew McAfee, and Michael Spence, Foreign Affairs, July/August
2014, accessed August 1, 2016, https://www.foreignaffairs.com/articles/united-states/
2014-06-04/new-world-order).
114 GDP does not do a good job of measuring the value of many digital goods. When society
cant manage what isnt measured, badpolicy decisions result. One alternative is to look at
consumer surplus, not just dollar flows. As AI is embodied in more goods, this issue becomes more
salient. It may look like GDP goes down but people have better well-being through access to these
digital goods. See Erik Brynjolfsson and Adam Saunders, What the GDP Gets Wrong (Why
Managers Should Care), Sloan Management Review, vol. 51, no. 1 (October 1, 2009): 9596. 39
 ENTERTAINMENT
With the explosive growth of the internet over the past fifteen years, few can imagine
their daily lives without it. Powered by AI, the internet has established user-generated
content as a viable source of information and entertainment. Social networks such
as Facebook are now pervasive, and they function as personalized channels of
social interaction and entertainmentsometimes to the detriment of interpersonal
interaction. Apps such as WhatsApp and Snapchat enable smart-phone users to
remain constantly in touch with peers and share sources of entertainment and
AI will increasingly enable information. In on-line communities such as Second Life and role-playing games such
entertainment that is more as World of Warcraft, people imagine an alternative existence in a virtual world.115
Specialized devices, such as Amazons Kindle have also redefined the essentials of
interactive, personalized, long-cherished pastimes. Books can now be browsed and procured with a few swipes
of the finger, stored by the thousands in a pocket-sized device, and read in much the
and engaging. Research same way as a handheld paperback.
should be directed toward Trusted platforms now exist for sharing and browsing blogs, videos, photos, and
topical discussions, in addition to a variety of other user-generated information. To
understanding how to operate at the scale of the internet, these platforms must rely on techniques that are
being actively developed in natural language processing, information retrieval, image
leverage these attributes processing, crowdsourcing, and machine learning. Algorithms such as collaborative
filtering have been developed, for example, to recommend relevant movies, songs, or
for individuals and articles based on the users demographic details and browsing history.116
societys benefit. Traditional sources of entertainment have also embraced AI to keep pace with the
times. As exemplified in the book and movie Moneyball, professional sport is now subjected
to intensive quantitative analysis.117 Beyond aggregate performance statistics, on-field
signals can be monitored using sophisticated sensors and cameras. Software has been
created for composing music118 and recognizing soundtracks.119 Techniques from computer
vision and NLP have been used in creating stage performances.120 Even the lay user can
exercise his or her creativity on platforms such as WordsEye, which automatically
generates 3D scenes from natural language text.121 AI has also come to the aid of
historical research in the arts, and is used extensively in stylometry and, more recently,
in the analysis of paintings.122
The enthusiasm with which humans have responded to AI-driven entertainment
has been surprising and led to concerns that it reduces interpersonal interaction
among human beings. Few predicted that people would spend hours on end
interacting with a display. Children often appear to be genuinely happier playing
at home on their devices rather than outside with their friends. AI will increasingly
enable entertainment that is more interactive, personalized, and engaging. Research
should be directed toward understanding how to leverage these attributes for
individuals and societys benefit.

115 Second Life, accessed August 1, 2016, http://secondlife.com; World of Warcraft, Blizzard
Entertainment, Inc, accessed August 1, 2016, http://us.battle.net/wow/en/.
116 John S. Breese, David Heckerman, and Carl Kadie, Empirical Analysis of Predictive
Algorithms for Collaborative Filtering, Proceedings of the 14th Conference on Uncertainty in Artificial
Intelligence (July 1998), accessed August 1, 2016, http://arxiv.org/pdf/1301.7363.pdf, 4352.
117 Michael Lewis, Moneyball: The Art of Winning an Unfair Game (New York: W. W. Norton &
Company, Inc., 2003): http://www.imdb.com/title/tt1210166/).
118 MuseScore, accessed August 1, 2016, https://musescore.org/.
119 Shazam, accessed August 1, 2016, http://www.shazam.com/.
120 Annie Dorsen, accessed August 1, 2016, http://www.anniedorsen.com/.
121 WordsEye, accessed August 1, 2016, https://www.wordseye.com/.
122 Stylometry, Wikipedia, last modified August 4, 2016, accessed August 1, 2016, https://
40 en.wikipedia.org/wiki/Stylometry; http://arxiv.org/pdf/1408.3218v1.pdf.
Imagining the Future
The success of any form of entertainment is ultimately determined by the individuals
and social groups that are its subjects. The modes of entertainment that people find
appealing are diverse and change over time. It is therefore hard to predict the forms
entertainment will take in the next fifteen years precisely. Nevertheless, current trends
suggest at least a few features that the future entertainment landscape is likely to contain.
To date, the information revolution has mostly unfolded in software. However,
with the growing availability of cheaper sensors and devices, greater innovation in the
hardware used in entertainment systems is expected. Virtual reality and haptics could More sophisticated tools
enter our living roomspersonalized companion robots are already being developed.123
With the accompanying improvements in Automatic Speech Recognition, the Study
and apps will become
Panel expects that interaction with robots and other entertainment systems will become available to make it
dialogue-based, perhaps constrained at the start, but progressively more human-like.
Equally, the interacting systems are predicted to develop new characteristics such as even easier to produce
emotion, empathy, and adaptation to environmental rhythms such as time of day.124
Today, an amateur with a video camera and readily-available software tools can high-quality content, for
make a relatively good movie. In the future, more sophisticated tools and apps will
example, to compose
become available to make it even easier to produce high-quality content, for example,
to compose music or to choreograph dance using an avatar. The creation and music or to choreograph
dissemination of entertainment will benefit from the progress of technologies such as
ASR, dubbing, and Machine Translation, which will enable content to be customized dance using an avatar.
to different audiences inexpensively. This democratization and proliferation of AI-
created media makes it difficult to predict how humans taste for entertainment, which
are already fluid, will evolve.
With content increasingly delivered digitally, and large amounts of data being
logged about consumers preferences and usage characteristics, media powerhouses
will be able to micro-analyze and micro-serve content to increasingly specialized
segments of the populationdown to the individual.125 Conceivably the stage is set
for the emergence of media conglomerates acting as Big Brothers who are able to
control the ideas and online experiences to which specific individuals are exposed.
It remains to be seen whether broader society will develop measures to prevent their
emergence. This topic, along with others pertaining to AI-related policy, is treated in
more detail in the next section.

123 Emoters, accessed August 1, 2016, http://emoterbots.com/.

124 Siri, Apple, Inc., accessed August 1, 2016, http://www.apple.com/in/ios/siri/.
125 Ryan Calo, Digital Market Manipulation, George Washington Law Review 82, no. 4 (2014):
9951051. 41
 SECTION III: PROSPECTS AND
RECOMMENDATIONS FOR AI PUBLIC POLICY
The goal of AI applications must be to create value for society. Our policy recommendations
flow from this goal, and, while this report is focused on a typical North American city in
2030, the recommendations are broadly applicable to other places over time. Strategies that
enhance our ability to interpret AI systems and participate in their use may help build trust
The measure of success and prevent drastic failures. Care must be taken to augment and enhance human capabilities
and interaction, and to avoid discrimination against segments of society. Research to encourage
for AI applications is this direction and inform public policy debates should be emphasized. Given the current
sector-specific regulation of US industries, new or retooled laws and policies will be needed
the value they create to address the widespread impacts AI is likely to bring. Rather than more or stricter
regulation, policies should be designed to encourage helpful innovation, generate and transfer
for human lives. Going
expertise, and foster broad corporate and civic responsibility for addressing critical societal
forward, the ease with issues raised by these technologies. In the long term, AI will enable new wealth creation that
will require social debate on how the economic fruits of AI technologies should be shared.
which people use and
adapt to AI applications AI POLICY, NOW AND IN THE FUTURE
Throughout history, humans have both shaped and adapted to new technologies. This
will likewise largely report anticipates that advances in AI technologies will be developed and fielded
graduallynot in sudden, unexpected jumps in the techniques themselvesand will
determine their success.
build on what exists today, making this adaptation easier. On the other hand, small
improvements to techniques, computing power, or availability of data can occasionally
lead to novel, game-changing applications. The measure of success for AI applications
is the value they create for human lives. Going forward, the ease with which people
use and adapt to AI applications will likewise largely determine their success.
Conversely, since AI applications are susceptible to errors and failures, a mark of
their success will be how users perceive and tolerate their shortcomings. As AI becomes
increasingly embedded in daily lives and used for more critical tasks, system mistakes
may lead to backlash from users and negatively affect their trust. Though accidents
in a self-driving car may be less probable than those driven by humans, for example,
they will attract more attention. Design strategies that enhance the ability of humans
to understand AI systems and decisions (such as explicitly explaining those decisions),
and to participate in their use, may help build trust and prevent drastic failures. Likewise,
developers should help manage peoples expectations, which will affect their happiness
and satisfaction with AI applications. Frustration in carrying out functions promised
by a system diminishes peoples trust and reduces their willingness to use the system
in the future.
Another important consideration is how AI systems that take over certain tasks
will affect peoples affordances and capabilities. As machines deliver super-human
performances on some tasks, peoples ability to perform them may wither. Already,
introducing calculators to classrooms has reduced childrens ability to do basic
arithmetic operations. Still, humans and AI systems have complementary abilities.
People are likely to focus on tasks that machines cannot do as well, including complex
reasoning and creative expression.
Already, children are increasingly exposed to AI applications, such as interacting
with personal assistants on cell phones or with virtual agents in theme parks. Having
early exposure will improve childrens interactions with AI applications, which
will become a natural part of their daily lives. As a result, gaps will appear in how
younger and older generations perceive AIs influences on society.
42
 Likewise, AI could widen existing inequalities of opportunity if access to AI
technologiesalong with the high-powered computation and large-scale data that
fuel many of themis unfairly distributed across society. These technologies will
improve the abilities and efficiency of people who have access to them. A person with
access to accurate Machine Translation technology will be better able to use learning
resources available in different languages. Similarly, if speech translation technology is
only available in English, people who do not speak English will be at a disadvantage.
Further, AI applications and the data they rely upon may reflect the biases of their
designers and users, who specify the data sources. This threatens to deepen existing AI could widen
social biases, and concentrate AIs benefits unequally among different subgroups
of society. For example, some speech recognition technologies do not work well for existing inequalities of
women and people with accents. As AI is increasingly used in critical applications,
opportunity if access to
these biases may surface issues of fairness to diverse groups in society. On the other
hand, compared to the well-documented biases in human decision-making, AI-based AI technologiesalong
decision-making tools have the potential to significantly reduce the bias in critical
decisions such as who is lent money or sent to jail. with the high-powered
Privacy concerns about AI-enabled surveillance are also widespread, particularly
in cities with pervasive instrumentation. Sousveillance, the recording of an activity
computation and large-
by a participant, usually with portable personal devices, has increased as well. Since scale data that fuel
views about bias and privacy are based on personal and societal ethical and value
judgments, the debates over how to address these concerns will likely grow and resist many of themis unfairly
quick resolution. Similarly, since AI is generating significant wealth, debates will grow
regarding how the economic fruits of AI technologies should be sharedespecially as distributed across society.
AI expertise and the underlying data sets that fuel applications are concentrated in a
small number of large corporations.
To help address these concerns about the individual and societal implications
of rapidly evolving AI technologies, the Study Panel offers three general policy
recommendations:
1. Define a path toward accruing technical expertise in AI at all
levels of government. Effective governance requires more experts
who understand and can analyze the interactions between AI
technologies, programmatic objectives, and overall societal values.
Absent sufficient technical expertise to assess safety or other metrics, national or
local officials may refuse to permit a potentially promising application. Or insufficiently
trained officials may simply take the word of industry technologists and green light a
sensitive application that has not been adequately vetted. Without an understanding
of how AI systems interact with human behavior and societal values, officials will be
poorly positioned to evaluate the impact of AI on programmatic objectives.
2. Remove the perceived and actual impediments to research on the
fairness, security, privacy, and social impacts of AI systems.
Some interpretations of federal laws such as the Computer Fraud and Abuse Act
and the anti-circumvention provision of the Digital Millennium Copyright Act are
ambiguous regarding whether and how proprietary AI systems may be reverse engineered
and evaluated by academics, journalists, and other researchers. Such research is critical
if AI systems with physical and other material consequences are to be properly vetted
and held accountable.
3. Increase public and private funding for interdisciplinary studies of
the societal impacts of AI.
As a society, we are underinvesting resources in research on the societal implications
of AI technologies. Private and public dollars should be directed toward interdisciplinary
43
 teams capable of analyzing AI from multiple angles. Research questions range from
basic research into intelligence to methods to assess and affect the safety, privacy,
fairness, and other impacts of AI.
Questions include: Who is responsible when a self-driven car crashes or an
intelligent medical device fails? How can AI applications be prevented from unlawful
discrimination? Who should reap the gains of efficiencies enabled by AI technologies
and what protections should be afforded to people whose skills are rendered obsolete?
As AI becomes integrated more broadly and deeply into industrial and consumer
As a society, we are products, it enters areas in which established regulatory regimes will need to be
adapted to AI innovations or in some cases fundamentally reconfigured according to
underinvesting resources broadly accepted goals and principles.
The approach in the United States to date has been sector-specific, with oversight
in research on the
by a variety of agencies. The use of AI in devices that deliver medical diagnostics and
societal implications of treatments is subject to aggressive regulation by the Food and Drug Administration
(FDA), both in defining what the product is and specifying the methods by which it is
AI technologies. Private produced, including standards of software engineering. The use of drones in regulated
airspace falls under the authority of the Federal Aviation Administration (FAA).126 For
and public dollars should consumer-facing AI systems, regulation by the Federal Trade Commission (FTC) comes
be directed toward into play. Financial markets using AI technologies, such as in high-frequency trading,
come under regulation by the Security Exchange Commission (SEC).
interdisciplinary teams In addition to sector-specific approaches, the somewhat ambiguous and broad
regulatory category of critical infrastructure may apply to AI applications.127 The
capable of analyzing AI Obama Administrations Presidential Policy Directive (PPD) 21 broadly defines critical
infrastructure as composed of the assets, systems, and networks, whether physical or
from multiple angles.
virtual, so vital to the United States that their incapacitation or destruction would have
a debilitating effect on security, national economic security, national public health or
safety, or any combination thereof. Today, an enterprise does not come under federal
regulation solely by falling under that broad definition. Instead, the general trend of
federal policy is to seek regulation in sixteen sectors of the economy.128
As regards AI, critical infrastructure is notably defined by the end-user application,
and not the technology or sector that actually produces AI software.129 Software

126 FAA controls the ways drones fly, requires drones to be semiautonomous as opposed to
autonomous, requires visual connection to the drone, and enforces no-fly zones close to airports.
127 Presidential Policy Directive (PPD-21)Critical Infrastructure Security and Resilience, The
White House, February 12, 2013, accessed August 1, 2016, https://www.whitehouse.gov/the-press-
office/2013/02/12/presidential-policy-directive-critical-infrastructure-security-and-resil.
128 PPD 21 identifies agencies responsible in each case. Chemical: Department of Homeland
Security; Commercial Facilities: Department of Homeland Security; Communications: Department
of Homeland Security; Critical Manufacturing: Department of Homeland Security; Dams:
Department of Homeland Security; Defense Industrial Base: Department of Defense; Emergency
Services: Department of Homeland Security; Energy: Department of Energy; Financial Services:
Department of the Treasury; Food and Agriculture: U.S. Department of Agriculture and
Department of Health and Human Services; Government Facilities: Department of Homeland
Security and General Services Administration; Healthcare and Public Health: Department of
Health and Human Services; Information Technology: Department of Homeland Security;
Nuclear Reactors, Materials, and Waste: Department of Homeland Security; Transportation
Systems: Department of Homeland Security and Department of Transportation; Water and
Wastewater Systems: Environmental Protection Agency.
129 In ICYMI- Business Groups Urge White House to Rethink Cyber Security Order, Internet
Association, March 5, 2013, accessed August 1, 2016, https://internetassociation.org/030513gov-3/:
Obamas Feb. 12 order says the government cant designate commercial information technology
products or consumer information technology services as critical U.S. infrastructure targeted for
voluntary computer security standards, ... Obamas order isnt meant to get down to the level
of products and services and dictate how those products and services behave, said David LeDuc,
senior director of public policy for the Software & Information Industry Association, a Washington
44 trade group that lobbied for the exclusions.
companies such as Google, Facebook, and Amazon have actively lobbied to avoid
being designated as critical to the economy, arguing that this would open the door to
regulation that would inevitably compromise their rapid product development cycles
and ability to innovate.130 Nonetheless, as the companies creating, operating, and
maintaining critical infrastructure use AI, interest will grow in regulating that software.
Some existing regulatory regimes for software safety (for example, the FDAs
regulation of high consequence medical software) require specific software
engineering practices at the developer level. However, modern software systems
are often assembled from library components which may be supplied by multiple Absent sufficient technical
vendors, and are relatively application-independent. It doesnt seem feasible or
desirable to subject all such developers to the standards required for the most critical, expertise to assess
rare applications. Nor does it seem advisable to allow unregulated use of such
safety or other metrics,
components in safety critical applications. Tradeoffs between promoting innovation
and regulating for safety are difficult ones, both conceptually and in practice. At a national or local officials
minimum, regulatory entities will require greater expertise going forward in order to
understand the implications of standards and measures put in place by researchers, may refuse to permit a
government, and industry.131
potentially promising
Policy and legal considerations applicationor green light
While a comprehensive examination of the ways artificial intelligence (AI) interacts
with the law is beyond the scope of this inaugural report, this much seems clear: a sensitive application that
as a transformative technology, AI has the potential to challenge any number of
has not been adequately
legal assumptions in the short, medium, and long term. Precisely how law and policy
will adapt to advances in AIand how AI will adapt to values reflected in law and vetted.
policydepends on a variety of social, cultural, economic, and other factors, and
is likely to vary by jurisdiction.
American law represents a mixture of common law, federal, state, and local
statutes and ordinances, andperhaps of greatest relevance to AIregulations.
Depending on its instantiation, AI could implicate each of these sources of law. For
example, Nevada passed a law broadly permitting autonomous vehicles and instructed
the Nevada Department of Motor Vehicles to craft requirements. Meanwhile, the
National Highway Transportation Safety Administration has determined that a self-
driving car system, rather than the vehicle occupants, can be considered the driver
of a vehicle. Some car designs sidestep this issue by staying in autonomous mode
only when hands are on the wheel (at least every so often), so that the human driver
has ultimate control and responsibility. Still, Teslas adoption of this strategy did not
prevent the first traffic fatality involving an autonomous car, which occurred in June
of 2016. Such incidents are sure to influence public attitudes towards autonomous
driving. And as most peoples first experience with embodied agents, autonomous
transportation will strongly influence the publics perception of AI.
Driverless cars are, of course, but one example of the many instantiations of
AI in services, products, and other contexts. The legal effect of introducing AI into
the provision of tax advice, automated trading on the stock market, or generating
medical diagnoses will also vary in accordance to the regulators that govern
these contexts and the rules that apply within them. Many other examples of AI
applications fall within current non-technology-specific policy, including predictive

130 Eric Engleman, Google Exception in Obamas Cyber Order Questioned as Unwise Gap,
Bloomberg Technology, March 4, 2013, accessed August 1, 2016, http://www.bloomberg.com/news/
articles/2013-03-05/google-exception-in-obama-s-cyber-order-questioned-as-unwise-gap.
131 Ryan Calo, The Case for a Federal Robotics Commission, Brookings Report, September 15,
2014, accessed August 1, 2016, http://www.brookings.edu/research/reports2/2014/09/case-for-
federal-robotics-commission. 45
 policing, non-discriminatory loans, healthcare applications such as eldercare and
drug delivery, systems designed to interact with children (for example, autonomous
tutoring systems are required to respect laws in regard to balanced handling of
evolution vs. intelligent design), and interactive entertainment.
Given the present structure of American administrative law, it seems unlikely
that AI will be treated comprehensively in the near term. Nevertheless, it is possible
to enumerate broad categories of legal and policy issues that AI tends to raise in
various contexts.
As AI applications engage Privacy
Private information about an individual can be revealed through decisions and
in behavior that, were it predictions made by AI. While some of the ways that AI implicates privacy mirror
those of technologies such as computers and the internet, other issues may be unique
done by a human, would
to AI. For example, the potential of AI to predict future behavior based on previous
constitute a crime, courts patterns raises challenging questions. Companies already use machine learning to
predict credit risk. And states run prisoner details through complex algorithms to
and other legal actors will predict the likelihood of recidivism when considering parole. In these cases, it is a
technical challenge to ensure that factors such as race and sexual orientation are not
have to puzzle through being used to inform AI-based decisions. Even when such features are not directly
whom to hold accountable provided to the algorithms, they may still correlate strongly with seemingly innocuous
features such as zip code. Nonetheless, with careful design, testing, and deployment,
and on what theory. AI algorithms may be able to make less biased decisions than a typical person.
Anthropomorphic interfaces increasingly associated with AI raise novel privacy
concerns. Social science research suggests people are hardwired to respond to
anthropomorphic technology as though it were human. Subjects in one study
were more likely to answer when they were born if the computer first stated
when it was built.132 In another, they skipped sensitive questions when posed by an
anthropomorphic interface.133 At a basic level lies the question: Will humans continue
to enjoy the prospect of solitude in a world permeated by apparently social agents
living in our houses, cars, offices, hospital rooms, and phones?134
Innovation policy
Early law and policy decisions concerning liability and speech helped ensure the
commercial viability of the Internet. By contrast, the software industry arguably
suffers today from the decision of firms to pivot from open and free software to the
more aggressive pursuit of intellectual property protections, resulting in what some
have termed patent thickets. Striking the proper balance between incentivizing
innovation in AI while promoting cooperation and protection against third party
harm will prove a central challenge.
Liability (civil)
As AI is organized to directly affect the world, even physically, liability for harms
caused by AI will increase in salience. The prospect that AI will behave in ways
designers do not expect challenges the prevailing assumption within tort law that
courts only compensate for foreseeable injuries. Courts might arbitrarily assign
liability to a human actor even when liability is better located elsewhere for reasons
of fairness or efficiency. Alternatively, courts could refuse to find liability because
the defendant before the court did not, and could not, foresee the harm that the AI
caused. Liability would then fall by default on the blameless victim. The role

132 Youngme Moon, Intimate Exchanges: Using Computers to Elicit Self-Disclosure from
Consumers, Journal of Consumer Research 26, no. 4 (March 2000): 323339.
133 Lee Sproull, Mani Subramani, Sara Kiesler , Janet H. Walker, and Keith Waters, When the
Interface is a Face, Human-Computer Interaction 11, no. 2 (1996): 97124.
134 M. Ryan Calo, People Can Be So Fake: A New Dimension to Privacy andTechnology
46 Scholarship, Penn State Law Review 114, no. 3 (2010): 809855.
of product liabilityand the responsibility that falls to companies manufacturing
these products will likely grow when human actors become less responsible for the
actions of a machine.
Liability (criminal)
If tort law expects harms to be foreseeable, criminal law goes further to expect that
harms be intended. US law in particular attaches great importance to the concept of
mens reathe intending mind. As AI applications engage in behavior that, were it
done by a human, would constitute a crime, courts and other legal actors will have to
puzzle through whom to hold accountable and on what theory. AI applications could
Agency
The above issues raise the question of whether and under what circumstances an increasingly shift
AI system could operate as the agent of a person or corporation. Already regulatory
investment from payroll
bodies in the United States, Canada, and elsewhere are setting the conditions under
which software can enter into a binding contract.135 The more AI conducts legally and income to capital
salient activities, the greater the challenge to principles of agency under the law.
Certification expenditure. Depending on
The very notion of artificial intelligence suggests a substitution for human skill
and ingenuity. And in many contexts, ranging from driving to performing surgery or
a state budgets reliance
practicing law, a human must attain some certification or license before performing on payroll and income
a given task. Accordingly, law and policy will have toand already doesgrapple
with how to determine competency in an AI system. For example, imagine a tax, such a shift could be
robotics company creates a surgical platform capable of autonomously removing an
appendix. Or imagine a law firm writes an application capable of rendering legal destabilizing.
advice. Today, it is unclear from a legal perspective who in this picture would have to
pass the medical boards or legal bar, let alone where they would be required to do so.136
Labor
As AI substitutes for human roles, some jobs will be eliminated and new jobs will
be created. The net effect on jobs is ambiguous, but labor markets are unlikely to
benefit everyone evenly. The demand for some types of skills or abilities will likely
drop significantly, negatively affecting the employment levels and wages of people
with those skills.137 While the ultimate effects on income levels and distribution are not
inevitable, they depend substantially on government policies, on the way companies
choose to organize work, and on decisions by individuals to invest in learning new
skills and seeking new types of work and income opportunities. People who find their
employment altered or terminated as a consequence of advances of AI may seek
recourse in the legislature and courts. This may be why Littler Mendelson LLP
perhaps the largest employment law firm in the worldhas an entire practice group
to address robotics and artificial intelligence.
Taxation
Federal, state, and local revenue sources may be affected. Accomplishing a task using AI
instead of a person can be faster and more accurateand avoid employment taxes.
As a result, AI applications could increasingly shift investment from payroll and
income to capital expenditure. Depending on a state budgets reliance on payroll
and income tax, such a shift could be destabilizing. AI may also display different

135 Ian R. Kerr, Ensuring the Success of Contract Formation in Agent-Mediated Electronic
Commerce, Electronic Commerce Research1 (2001): 183202.
136 Ryan Calo, Digital Agendas public discussion on The effects of robotics on economics,
labour and society, Ausschuss Digitale Agenda, (Deutsche Bundestag: Ausschussdrucksache A-Drs.
18(24)102), June 22, 2016, accessed August 1, 2016, https://www.bundestag.de/blob/428266/195a
1cde8d5347849accbbe60ed91865/a-drs-18-24-102-data.pdf.
137 Brynjolfsson and McAfee, Race Against the Machine: How the Digital Revolution is
Accelerating Innovation, Driving Productivity, and Irreversibly Transforming Employment and the
Economy (2011); Brynjolfsson and McAfee, Second Machine Age. 47
 habits than people, resulting in still fewer revenue sources. The many municipalities
relying on income from speeding or parking tickets will have to find alternatives
if autonomous cars can drop people off and find distance parking, or if they are
programmed not to violate the law. As a result, government bodies trying to balance
their budgets in light of advances in AI may pass legislation to slow down or alter the
course of the technology.
Politics
AI technologies are already being used by political actors in gerrymandering and
Like other technologies, targeted robocalls designed to suppress votes, and on social media platforms in the
form of bots.138 They can enable coordinated protest as well as the ability to predict
AI has the potential to be protests, and promote greater transparency in politics by more accurately pinpointing
who said what, when. Thus, administrative and regulatory laws regarding AI can be
used for good or nefarious
designed to promote greater democratic participation or, if ill-conceived, to reduce it.
purposes. A vigorous and This list is not exhaustive and focuses largely on domestic policy in the United
States, leaving out many areas of law that AI is likely to touch. One lesson that
informed debate about might be drawn concerns the growing disconnect between the context-specific way
in which AI is governed today and a wider consideration of themes shared by AI
how to best steer AI in technologies across industries or sectors of society. It could be tempting to create new
ways that enrich our lives institutional configurations capable of amassing expertise and setting AI standards
across multiple contexts. The Study Panels consensus is that attempts to regulate AI
and our society is an in general would be misguided, since there is no clear definition of AI (it isnt any
one thing), and the risks and considerations are very different in different domains.
urgent and vital need. Instead, policymakers should recognize that to varying degrees and over time, various
industries will need distinct, appropriate, regulations that touch on software built
using AI or incorporating AI in some way. The government will need the expertise to
scrutinize standards and technology developed by the private and public sector, and
to craft regulations where necessary.

Guidelines for the future

Faced with the profound changes that AI technologies can produce, pressure for
more and tougher regulation is inevitable. Misunderstanding about what AI is
and is not, especially against a background of scare-mongering, could fuel opposition
to technologies that could benefit everyone. This would be a tragic mistake.
Regulation that stifles innovation, or relocates it to other jurisdictions, would be
similarly counterproductive.
Fortunately, principles that guide successful regulation of current digital
technologies can be instructive. A recent multi-year study comparing privacy
regulation in four European countries and the United States, for example, yielded
counter-intuitive results.139 Those countries, such as Spain and France, with strict
and detailed regulations bred a compliance mentality within corporations, which
had the effect of discouraging both innovation and robust privacy protections.
Rather than taking responsibility for privacy protection internally and developing a
professional staff to foster it in business and manufacturing processes, or engaging with
privacy advocates or academics outside their walls, these companies viewed privacy as
a compliance activity. Their focus was on avoiding fines or punishments, rather than
proactively designing technology and adapting practices to protect privacy.
By contrast, the regulatory environment in the United States and Germany,
which combined more ambiguous goals with tough transparency requirements and
meaningful enforcement, were more successful in catalyzing companies to view
138 Political Bots, accessed August 1, 2016, http://politicalbots.org/.
139 Kenneth A. Bamberger and Deirdre K. Mulligan, Privacy on the Ground: Driving Corporate Behavior
48 in the United States and Europe (Cambridge, Massachusetts: MIT Press, 2015).
privacy as their responsibility. Broad legal mandates encouraged companies to
develop a professional staff and processes to enforce privacy controls, engage with
outside stakeholders, and to adapt their practices to technology advances. Requiring
greater transparency enabled civil society groups and media to become credible
enforcers both in court and in the court of public opinion, making privacy more
salient to corporate boards and leading them to further invest in privacy protection.
In AI, too, regulators can strengthen a virtuous cycle of activity involving internal
and external accountability, transparency, and professionalization, rather than narrow
compliance. As AI is integrated into cities, it will continue to challenge existing
protections for values such as privacy and accountability. Like other technologies, AI
has the potential to be used for good or nefarious purposes. This report has tried to
highlight the potential for both. A vigorous and informed debate about how to best
steer AI in ways that enrich our lives and our society, while encouraging creativity
in the field, is an urgent and vital need. Policies should be evaluated as to whether
they democratically foster the development and equitable sharing of AIs benefits,
or concentrate power and benefits in the hands of a fortunate few. And since future
AI technologies and their effects cannot be foreseen with perfect clarity, policies will
need to be continually re-evaluated in the context of observed societal challenges and
evidence from fielded systems.
As this report documents, significant AI-related advances have already had
an impact on North American cities over thepast fifteen years, and even more
substantial developments will occur over the next fifteen. Recent advances are largely
due to the growth and analysis of large data sets enabled by the Internet, advances
in sensory technologies and, more recently, applications of deep learning. In the
coming years, as the public encounters new AI applications in domains such as
transportation and healthcare, they must be introduced in ways that build trust and
understanding, and respect human and civil rights. While encouraging innovation,
policies and processes should address ethical, privacy, and security implications, and
should work to ensure that the benefits of AI technologies will be spread broadly and
fairly. Doing so will be critical if Artificial Intelligence research and its applications
are to exert a positive influence on North American urban life in 2030 and beyond.

49
 APPENDIX I: A SHORT HISTORY OF AI
This Appendix is based primarily on Nilssons book140 and written from the prevalent
current perspective, which focuses on data intensive methods and big data. However
important, this focus has not yet shown itself to be the solution to all problems. A
complete and fully balanced history of the field is beyond the scope of this document.
The field of Artificial Intelligence (AI) was officially born and christened at a 1956
workshop organized by John McCarthy at the Dartmouth Summer Research Project
on Artificial Intelligence. The goal was to investigate ways in which machines could
The field of Artificial be made to simulate aspects of intelligencethe essential idea that has continued to
Intelligence (AI) was drive the field forward. McCarthy is credited with the first use of the term artificial
intelligence in the proposal he co-authored for the workshop with Marvin Minsky,
officially born and Nathaniel Rochester, and Claude Shannon.141 Many of the people who attended soon
led significant projects under the banner of AI, including Arthur Samuel, Oliver
christened at a 1956 Selfridge, Ray Solomonoff, Allen Newell, and Herbert Simon.
workshop. The goal was to Although the Dartmouth workshop created a unified identity for the field and
a dedicated research community, many of the technical ideas that have come to
investigate ways in which characterize AI existed much earlier. In the eighteenth century, Thomas Bayes
provided a framework for reasoning about the probability of events.142 In the
machines could be made nineteenth century, George Boole showed that logical reasoningdating back to
Aristotlecould be performed systematically in the same manner as solving a system
to simulate aspects of of equations.143 By the turn of the twentieth century, progress in the experimental
intelligencethe essential sciences had led to the emergence of the field of statistics,144 which enables
inferences to be drawn rigorously from data. The idea of physically engineering a
idea that has continued to machine to execute sequences of instructions, which had captured the imagination
of pioneers such as Charles Babbage, had matured by the 1950s, and resulted in the
drive the field forward. construction of the first electronic computers.145 Primitive robots, which could
sense and act autonomously, had also been built by that time.146
The most influential ideas underpinning computer science came from Alan Turing,
who proposed a formal model of computing. Turings classic essay, Computing Machinery
and Intelligence,147 imagines the possibility of computers created for simulating intelligence
and explores many of the ingredients now associated with AI, including how intelligence
might be tested, and how machines might automatically learn. Though these ideas
inspired AI, Turing did not have access to the computing resources needed to
translate his ideas into action.
Several focal areas in the quest for AI emerged between the 1950s and the 1970s.148
140 Nilsson, The Question for Artificial Intelligence.
141 J. McCarthy, Marvin L. Minsky, Nathaniel Rochester, and Claude E. Shannon, A Proposal
for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955, accessed
August 1, 2016, http://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html.
142 Thomas Bayes, An Essay towards Solving a Problem in the Doctrine of Chances,
Philosophical Transactions of the Royal Society of London 53 (January 1, 1763): 370418, accessed August 1,
2016, http://rstl.royalsocietypublishing.org/search?fulltext=an+essay+towards+solving&submit=yes
&andorexactfulltext=and&x=0&y=0.
143 George Boole, An Investigation of the Laws of Thought on Which are Founded the Mathematical Theories
of Logic and Probabilities, (Macmillan, 1854, reprinted with corrections, Dover Publications, New
York, NY, 1958, and reissued by Cambridge University Press, 2009), accessed August 1, 2016,
http://ebooks.cambridge.org/ebook.jsf ?bid=CBO9780511693090.
144 History of statistics, Wikipedia, Last modified June 3, 2016, accessed August 1, 2016,
https://en.wikipedia.org/wiki/History_of_statistics.
145 Joel N. Shurkin, Engines of the Mind: The Evolution of the Computer from Mainframes to Microprocessors
(New York: W. W. Norton & Company, 1996).
146 William Grey Walter, An Electromechanical Animal, Dialectica 4 (1950): 4249.
147 A. M. Turing, Computing Machinery and Intelligence, Mind 59, no. 236 (1950): 433460.
148 Marvin Minsky, Steps toward Artificial Intelligence, MIT Media Laboratory, October 24,
50 1960, accessed August 1, 2016, http://web.media.mit.edu/~minsky/papers/steps.html.
Newell and Simon pioneered the foray into heuristic search, an efficient procedure
for finding solutions in large, combinatorial spaces. In particular, they applied this
idea to construct proofs of mathematical theorems, first through their Logic Theorist
program, and then through the General Problem Solver.149 In the area of computer
vision, early work in character recognition by Selfridge and colleagues150 laid the
basis for more complex applications such as face recognition.151 By the late sixties,
work had also begun on natural language processing.152 Shakey, a wheeled
robot built at SRI International, launched the field of mobile robotics. Samuels
Checkers-playing program, which improved itself through self-play, was one of the Although the separation
first working instances of a machine learning system.153 Rosenblatts Perceptron,154
a computational model based on biological neurons, became the basis for the field of AI into sub-fields
of artificial neural networks. Feigenbaum and others advocated 155the case for
has enabled deep
building expert systemsknowledge repositories tailored for specialized domains
such as chemistry and medical diagnosis.156 technical progress along
Early conceptual progress assumed the existence of a symbolic system that could
be reasoned about and built upon. But by the 1980s, despite this promising headway several different fronts,
made into different aspects of artificial intelligence, the field still could boast no
significant practical successes. This gap between theory and practice arose in part from
synthesizing intelligence
an insufficient emphasis within the AI community on grounding systems physically, with at any reasonable scale
direct access to environmental signals and data. There was also an overemphasis on
Boolean (True/False) logic, overlooking the need to quantify uncertainty. The field invariably requires many
was forced to take cognizance of these shortcomings in the mid-1980s, since interest
in AI began to drop, and funding dried up. Nilsson calls this period the AI winter. different ideas to be
A much needed resurgence in the nineties built upon the idea that Good Old-
integrated.
Fashioned AI157 was inadequate as an end-to-end approach to building intelligent
systems. Rather, intelligent systems needed to be built from the ground up, at all times
solving the task at hand, albeit with different degrees of proficiency.158 Technological
progress had also made the task of building systems driven by real-world data more
feasible. Cheaper and more reliable hardware for sensing and actuation made robots
easier to build. Further, the Internets capacity for gathering large amounts of data,
and the availability of computing power and storage to process that data, enabled
statistical techniques that, by design, derive solutions from data. These developments
have allowed AI to emerge in the past two decades as a profound influence on our
daily lives, as detailed in Section II.
149 Allen Newell, John C Shaw, and Herbert A Simon, Report on a general problem-solving
program, Proceedings of the International Conference on Information Processing, UNESCO, Paris 15-20
June 1959 (Unesco/Oldenbourg/Butterworths, 1960), 256264.
150 O. G. Selfridge, Pandemonium: A paradigm for learning, Proceedings of the Symposium on
Mechanization of Thought Processes (London: H. M. Stationary Office, 1959): 511531.
151 Woodrow W. Bledsoe and Helen Chan, A Man-Machine Facial Recognition System: Some
Preliminary Results, Technical Report PRI 19A (Palo Alto, California: Panoramic Research, Inc., 1965).
152 D. Raj Reddy, Speech Recognition by Machine: A Review, Proceedings of the IEEE 64, no.4
(April 1976), 501531.
153 Arthur Samuel, Some Studies in Machine Learning Using the Game of Checkers, IBM
Journal of Research and Development 3, no. 3 (1959): 210229.
154 Frank Rosenblatt, The PerceptronA Perceiving and Recognizing Automaton,
Report 85-460-1, (Buffalo, New York: Cornell Aeronautical Laboratory, 1957).
155 Shakey the robot, Wikipedia, last modified July 11, 2016, accessed August 1, 2016,
https://en.wikipedia.org/wiki/Shakey_the_robot.
156 Edward A. Feigenbaum and Bruce G. Buchanan, DENDRAL and Meta-DENDRAL:
Roots of Knowledge Systems and Expert System Applications, Artificial Intelligence 59, no. 1-2
(1993), 233240.
157 John Haugeland, Artificial Intelligence: The Very Idea, (Cambridge, Massachusetts: MIT
Press, 1985).
158 Rodney A. Brooks, Elephants Dont Play Chess, Robotics and Autonomous Systems 6, no. 1-2
(June 1990): 315. 51
 In summary, following is a list of some of the traditional sub-areas of AI. As described
in Section II, some of them are currently hotter than others for various reasons.
But that is neither to minimize the historical importance of the others, nor to say that
they may not re-emerge as hot areas in the future.
Search and Planning deal with reasoning about goal-directed behavior. Search
plays a key role, for example, in chess-playing programs such as Deep Blue, in
deciding which move (behavior) will ultimately lead to a win (goal).
The area of Knowledge Representation and Reasoning involves processing
information (typically when in large amounts) into a structured form that can
be queried more reliably and efficiently. IBMs Watson program, which beat
human contenders to win the Jeopardy challenge in 2011, was largely based on
an efficient scheme for organizing, indexing, and retrieving large amounts of
information gathered from various sources.159
Machine Learning is a paradigm that enables systems to automatically improve
their performance at a task by observing relevant data. Indeed, machine
learning has been the key contributor to the AI surge in the past few decades,
ranging from search and product recommendation engines, to systems for
speech recognition, fraud detection, image understanding, and countless other
tasks that once relied on human skill and judgment. The automation of these
tasks has enabled the scaling up of services such as e-commerce.
As more and more intelligent systems get built, a natural question to consider
is how such systems will interact with each other. The field of Multi-Agent
Systems considers this question, which is becoming increasingly important in
on-line marketplaces and transportation systems.
From its early days, AI has taken up the design and construction of systems that
are embodied in the real world. The area of Robotics investigates fundamental
aspects of sensing and actingand especially their integrationthat enable
a robot to behave effectively. Since robots and other computer systems share
the living world with human beings, the specialized subject of Human Robot
Interaction has also become prominent in recent decades.
Machine perception has always played a central role in AI, partly in developing
robotics, but also as a completely independent area of study. The most commonly
studied perception modalities are Computer Vision and Natural Language
Processing, each of which is attended to by large and vibrant communities.
Several other focus areas within AI today are consequences of the growth of the
Internet. Social Network Analysis investigates the effect of neighborhood relations
in influencing the behavior of individuals and communities. Crowdsourcing is
yet another innovative problem-solving technique, which relies on harnessing
human intelligence (typically from thousands of humans) to solve hard
computational problems.
Although the separation of AI into sub-fields has enabled deep technical progress
along several different fronts, synthesizing intelligence at any reasonable scale invariably
requires many different ideas to be integrated. For example, the AlphaGo program160 161
that recently defeated the current human champion at the game of Go used multiple
machine learning algorithms for training itself, and also used a sophisticated search
procedure while playing the game.
159 David A. Ferrucci, Introduction to This is Watson, IBM Journal of Research and Development,
56, no. 3-4 (2012): 1.
160 David Silver et al., Mastering the Game of Go with Deep Neural Networks and Tree Search.
161 Steven Borowiec and Tracey Lien, AlphaGo beats human Go champ in milestone for
artificial intelligence, Los Angeles Times, March 12, 2016, accessed August 1, 2016, http://www.
latimes.com/world/asia/la-fg-korea-alphago-20160312-story.html.
52