Download roadmapping a us-german agenda for artificial intelligence policy

LEVI TILLEMANN AND COLIN MCCORMICK
ROADMAPPING A
U.S.-GERMAN AGENDA
FOR ARTIFICIAL
INTELLIGENCE POLICY
MARCH 2017
About the Authors
Acknowledgements
Dr. Levi Tillemann is Managing Partner at Valence
Strategic and a fellow at New America. Dr. Tillemann is
also the author of The Great Race: The Global Quest for
the Car of the Future (Simon & Schuster 2015) a book on
the intersection of policy and innovation in the global
auto industry. His current research focus is on robotics
and AI — with an emphasis on self-driving cars and
transportation. Previously he served as Special Advisor
for Policy and International Affairs at the U.S. Department
of Energy and was CEO of IRIS Engines, a company he
founded to develop a smaller, more efficient and more
powerful combustion engine (based on a design for
which he holds multiple patents). Tillemann was also
an associate director at IHS CERA. He has published in
both academic journals and general interest publications
including the Washington Post, the Wall Street Journal,
Fortune, The New Yorker and is a regular guest on radio
and TV programs including NPR’s Diane Rehm Show
and Marketplace. He is proficient in Chinese, Japanese,
Spanish and Portuguese. Tillemann is a graduate of Yale
University, was a Fulbright scholar in Taiwan and China,
and has a Ph.D. in Japan and China studies from Johns
Hopkins School of Advanced International Studies.
New America would like to thank the German Embassy
in Washington, D.C. for sponsoring the workshop
“Roadmapping a US-German Agenda for Artificial
Intelligence Policy.” In addition to helping conceptualize
this workshop, embassy officials helped bring together a
group of world-class experts in robotics, machine learning
and manufacturing as well as industry representatives
and officials from various government ministries, industry
organizations and think tanks. The resulting discussion
pointed toward a rich landscape for future inquiry and
many opportunities and challenges that lie ahead as the
United States and Germany embrace a new generation
of information technologies and robotics. New America
would also like to acknowledge Valence Strategic for its
role in conceptualizing and executing the workshop and
aggregating the resulting policy research agenda.
Dr. Colin McCormick is Partner and Chief Technologist
at Valence Strategic. As Senior Advisor for R&D at the
U.S. Department of Energy McCormick served as primary
science advisor to the U.S. Department of Energy for
energy innovation from 2010 to 2013. He has an extensive
background in technology analysis in energy, robotics,
and computer vision. His current research focus is on the
implications of artificial intelligence in a range of fields,
particularly transportation and satellite imagery analysis.
He also conducts computer vision research relating
to change detection in aerial imagery. McCormick has
published in a range of venues including both academic
journals and general interest publications such as
Fortune, Slate, and The New Yorker. Previously, he was
a professional staff member for the House Science &
Technology Committee. McCormick has a Ph.D. in atomic
and optical physics from UC Berkeley and conducted
post-doctoral research in quantum optics at the National
Institute of Standards and Technology (NIST) in the group
of 1997 Physics Nobel Laureate Bill Phillips.
About New America
New America is committed to renewing American politics,
prosperity, and purpose in the Digital Age. We generate big
ideas, bridge the gap between technology and policy, and
curate broad public conversation. We combine the best of
a policy research institute, technology laboratory, public
forum, media platform, and a venture capital fund for
ideas. We are a distinctive community of thinkers, writers,
researchers, technologists, and community activists who
believe deeply in the possibility of American renewal.
Find out more at newamerica.org/our-story.
Contents
Introduction
2
Safety (Physical Safety, Cybersecurity, Etc.)
3
Legal/Policy
4
Privacy/Data Sharing
5
Environment
6
Ethics
7
Distribution (Labor, Wages, Distribution, Taxation, Welfare, Etc.)
8
Notes
10
INTRODUCTION
2
Artificial Intelligence (AI) is poised to transform
society with impacts as far-reaching and profound
as electricity or the internal combustion engine.
Policymakers have a limited window within which
to grapple with foundational policy challenges
posed by the first general purpose technology of the
21st century. This task is complicated by the fact that
AI will impact almost every facet of our economy
and society—including areas as diverse as financial
planning, policing, elections, manufacturing, and
transportation.
into new AI-based frameworks. The utilization of AIbased technologies will also differ based on factors
ranging from demographics to economic structure.
AI’s role in our economy and society will cross
borders. It is therefore necessary to advance
a policy discussion that extends beyond
individual countries’ domestic policy concerns.
This policy research roadmap examines these
cross-border dynamics in the context of the U.S.German technology, trade, social, and economic
relationship. Both the U.S. and Germany are highly
industrialized countries striving to understand
the impact of digitalization and new forms of data
analytics. The emergence of AI as an integral part
of the twenty-first century economy will no doubt
have significant implications for the U.S.-German
cultural, political and economic relationship.
Policymakers in both countries will confront these
challenges within distinct social, cultural, and
legal settings. Under U.S. and German law there
will be differing rules and expectations regarding
aggregation, storage, and utilization of data feeding
To better understand areas in which transatlantic
policymakers can collaborate on understanding key
policy issues in the field of AI, New America and the
German Embassy in Washington, D.C. held a joint
workshop to roadmap critical research areas related
to artificial intelligence. The workshop brought
together a diverse range of participants from
academia, government, industry representatives,
and think tanks to discuss questions related to the
policy, social, and technological implications of AI
across society. The scope was policies that provide
a framework to enable socially beneficial uses of AI
between now and 2035.
This policy research roadmap
examines cross-border dynamics
in the context of the U.S.-German
technology, trade, social,and
economic relationship.
Some of the key insights and questions highlighted
by participants are summarized below.
NEW AMERICA
SAFETY (PHYSICAL SAFETY,
CYBERSECURITY, ETC.)
Allowing computers to make decisions that
would previously have been overseen by human
beings poses a number of challenges—none more
fundamental than ensuring basic safety. Not only
does AI impact the security of communications
and financial systems, but integration between AI
and robotics will increasingly lead to new safety
challenges in the physical world. Some key topics
and questions for research are included below.
Are there any areas/activities in which AI should
be prohibited or restricted?
There are precedents for emerging technologies that
pose particular dangers to be restricted, including
nuclear/chemical/biological weapons and human
cloning. Does AI pose dangers in any specific areas
that would justify such restrictions?2
Can we create fail-safe safety mechanisms for AI?
How do we classify the threat levels associated
with various AI technologies?
What approach should we use when trying to
understand the relative dangers/threat levels
associated with AI? Should threats be separated into
physical, financial, privacy, and other categories?
Will it be sufficient to extend existing cybersecurity
risk assessment frameworks (such as the NIST
approach to Industrial Control System Security1) or
is an entirely new framework required? An example
of a similar schema is the U.S. Department of
Transportation automation levels for autonomous
vehicles or the “Strategy for Autonomous and
Connected Driving” of the German Federal
Government.
Air gaps provide a physical mechanism to isolate
sensitive information technology systems from
hacking. Would similar hardware solutions be
effective for ensuring the safety of AI and intelligent
robotics (e.g. physical dongles)? If so, should they
be adopted as best practice?
How should geo-fencing be utilized by regulators?
As a response to airspace restrictions, some drone
manufacturers have implemented “geo-fencing,”
such that their products cannot operate in certain
geographical locations. Is this a viable safety
measure for other forms of AI that control movable
systems (cars, trucks, etc.)? How reliable can geofenced systems be made in practice? (See Sen.
Roadmapping a U.S.-German Agenda for Artificial Intelligence Policy
3
Schumer’s support for Federal Aviation Agency
(FAA) efforts to mandate geo-fencing of aerial
drones.)3
Can the United States and Germany develop a
common agenda around testing and validation for
autonomous robots?
Autonomous vehicles are one example of an
emerging AI-governed robotics technology that will
need to be regulated. How can the United States and
Germany collaborate on setting standards for this
and other AI-enabled robotics technologies as they
arise? What would harmonized testing methods be
based on?4 One potential platform for collaboration
is Germany’s “Digital Testbed Autobahn,” which the
German federal Government has established on the
A9 autobahn as a testing ground for autonomous
vehicles and other emerging technologies.
LEGAL/POLICY
Legal frameworks that govern commerce, criminal
law and liability in and between advanced
economies have developed over the course of
millennia and are enormously complex. These rules
generally assume the presence of a human being
who can be held responsible for actions that impact
society. The advent of AI-based decision-making
poses a host of questions for this approach to law
and policy including the following:
How do the distinct approaches to technological
deployment and development affect U.S. and
German needs for AI regulation?
In the United States there appears to be a decided
preference toward first implementing emerging
technologies and then developing regulatory
frameworks for them. In Germany, regulation often
precedes implementation. What are the implications
of this difference for industry development?
4
What are the different challenges associated
with the industrial structures of Germany and the
United States?
Germany’s heavy economic emphasis on
manufacturing implies that its focus should be on
preparing industries and workers for incorporating
AI into existing manufacturing systems. In the
United States, large-scale deindustrialization
implies the possibility of recapturing jobs
from low wage countries. In some cases such
reindustrialization is occurring in Germany as well
(e.g. creation of local production sites by Adidas).
How can we use existing regulatory function to
encourage positive path dependencies?
Like politics, the ultimate shape of a technology is
often dictated by historical chance. Historians and
political scientists call these developments “path
dependences.” Are there positive path dependencies
NEW AMERICA
in AI that we can encourage through existing
institutional structures?
is desirable that countries establish standards that
promote shared values, economic, and social goals.
Do we require new regulatory structures to deal
with regulating AI-enabled industries?
Compare U.S. and German standards in
autonomous vehicles
Existing regulatory frameworks were not designed
to deal with the advent of AI. For instance, decisions
made by AI systems are difficult to trace back to
individuals, making responsibility for the outcomes
unclear. AI’s technical complexity also makes
it imperative that regulatory institutions have
sufficient technical capacity to understand key
issues. What are the possible avenues for addressing
this shortfall in institutional capacity? Do we need
new institutions?5
In developed countries, mass deployment of
autonomous vehicles is likely to commence within a
decade. Develop a simple comparison of regulatory
and safety standards between U.S. and German (EU,
Japanese, other) autonomous vehicle regulations.
How to engage multi-country coalitions for AI
development?
What is the relationship between trade and
standards for AI?
AI technical standards have the potential to be
deployed as a barrier to various forms of trade.
Many of the concerns surrounding AI and trade may
be legitimate. However, others may be excuses for
various forms of protectionism.
AI will have profound implications for safety,
ethics, environment, and distribution of wealth. It
PRIVACY/DATA SHARING
In a hyper-connected world, the ubiquity of
data presents opportunities to build a society
that is more just, efficient, and human centered.
Simultaneously, there is the potential for this
data to be misused. In many cases the ethics of a
particular use of data may be ambiguous. Questions
surrounding AI and data include the following:
How should countries coordinate on data and
privacy issues associated with AI development?
Laws and customs regarding data gathering,
accumulation, utilization and privacy differ between
the United States and Germany. How should issues
of data localization and incorporation of local data
be harmonized?6
Roadmapping a U.S.-German Agenda for Artificial Intelligence Policy
5
Will AI inevitably lead to an erosion of personal
data privacy, or can we improve the framework
for granting/revoking permission for data sharing
and monitoring how personal data is being used?
Our current framework for privacy of personal
data relies on the theory of informed consent, and
generally does not allow more than binary control
(i.e. share vs. don’t share). However, informed
consent is often unrealistic (privacy and licensing
agreements often cover many pages and are not
actually read or understood) and having more
fine-grained control over when, how, and for what
purposes personal data is shared could be valuable.
ENVIRONMENT
AI-based industries have the potential to eliminate
an enormous amount of waste, delivering products
we need, when we need them and reducing
inefficiencies within transportation, manufacturing,
and logistics. At the same time, improvements in
one realm (e.g. speed, efficiency, cost, etc.) may
lead to challenges in another (e.g. employment,
emissions, etc.). Questions surrounding AI and the
environment include the following:
Can AI be used to optimize the efficiency and
performance of networks of machines?
One example might be scheduling the timing of
when electricity is drawn from the grid to minimize
emissions, or managing fleets of vehicles to reduce
traffic. More broadly, how should AI be deployed
in the context of the internet of things/connected
machines?7
What are the areas in which AI may result in
conflicting social goods/costs?
In certain areas, AI is likely to enable technologies
and business models that have profound social
benefits, while simultaneously exerting social costs.
One example is self-driving cars. On the one hand,
self-driving cars may result in wide-scale carpooling
(environmentally beneficial) and dramatically
reduce the number of road deaths (socially and
economically beneficial). On the other, they may
result in the elimination of hundreds of thousands
of jobs (socially destructive) or a dramatic
expansion in vehicle miles traveled (VMT).
AI-based industries have the potential to eliminate an
enormous amount of waste.
6
NEW AMERICA
ETHICS
AI has the potential to improve our utilization
of data in ways that eliminate social biases. But
sometimes objectively true data can also lead
to poor social outcomes. This may result from
mislabeling, categorization, or poor sampling.
The German Federal Ministry of Transport and
Digital Infrastructure has established an Ethics
Commission in order to deal with ethical aspects of
AI. Some of the questions associated with the ethics
of data and AI include:
How to deal with racial, gender, and other forms
of discrimination promoted by AI-based systems?
Should businesses using AI be required to divulge
the legal parameters/requirements of their
decision basis?
Increasingly, AI may be used to make decisions
that affect a person’s finances, livelihood, housing,
medical care, and other important aspects of
their lives. Should there be a method to challenge
or review highly consequential decisions taken
by AI systems? Given the complex nature of AI
processing/decision-making, is this technically
feasible?
Does AI pose a threat to democracy?
There are numerous ways in which AI systems may
lead to discriminatory outcomes. For instance,
ratings systems for customers/users of a specific
service might socially disadvantage one group
(e.g. women, men, the elderly, minorities, etc.). In
some cases, this may be the result of incentives
that align with a particular set of data but don’t
adequately address certain social needs. It is easy
to foresee a situation in which certain Western
biases are prioritized over cultural values from other
societies. It is just as easy to imagine a situation in
which a non-Western country (e.g. Russia, China,
etc.) generates AI that does not adhere to values
that many Westerners hold to be universal (e.g.
democracy).8
Democratic elections appear to be more vulnerable
to information warfare than was previously
understood. Would AI exacerbate this situation?
Could AI systems be used to manipulate information
flow or directly alter voting counts to affect political
outcomes? Could AI forge information, such as
images, documents, or sound recordings, to cause
political damage?
How to improve transparency in AI-animated
fields?
Part of the power of machine learning is the ability
to deploy data and complex algorithms to achieve
technological advances that would be challenging
Roadmapping a U.S.-German Agenda for Artificial Intelligence Policy
7
or impossible for human programmers to achieve.
Unfortunately, the complexity of these algorithms
often makes them inscrutable. How do we create
systems that increase the transparency of these
powerful, complex computational tools?
Can there be a science of “AI forensics”?
In some cases it might be necessary to trace the
decisions of an AI to humans. If this is impossible,
who should be held responsible for AI decisions
that cause harm but cannot be directly traced to
human developers or operators?
DISTRIBUTION (LABOR, WAGES,
DISTRIBUTION, TAXATION, WELFARE,
ETC.)
AI is bound to improve the productivity of many
sectors of the economy and reduce human error in
areas such as medicine and finance. At the same
time, there are serious questions regarding the labor
market implications of these technologies and how
to reimagine the social contract in a world where
an increasing proportion of economic productivity
is accounted for by machines and computers. Some
critical questions include the following:
Can the economic premise of compensation for
human production and labor survive in a world
where AI increasingly performs productive and
even creative tasks, with less and less human
input?
income, since we believe more productive people
should get more wealth. Who is being “productive”
when a job function is automated and no longer
needs human input?9
Could a Guaranteed Basic Income (GBI) be
implemented within the U.S. and/or German
political and legal systems?
The idea of a GBI has been introduced as a response
to the increasing automation and job displacement
portended by AI. What social impact would a GBI
have on dimensions such as happiness and life
satisfaction, health, and economic growth?
What are the limits of reskilling?
Our legal and economic systems are fundamentally
based on a linkage between productivity and
8
Are there areas in which reskilling efforts should be
NEW AMERICA
focused? Are there certain characteristics that we
can isolate that make a worker more/less likely to be
successfully reskilled?
Is there a vulnerability index for automation?
Can age, education, or other factors predict the
vulnerability of certain labor demographics to
automation?
to risk potential failure because of an extensive
familial safety net. Is expanding the social safety
net through tools like guaranteed basic income
a means of increasing entrepreneurship and
economic growth?
What are the prospects for reskilling labor
displaced by AI? How can we learn from existing
efforts within technical schools and successful
reskilling programs?
How do we improve skills-needs forecasting?
The time lag between demand for labor in emerging
fields and the training of workers capable of
fulfilling those jobs is a persistent and disruptive
source of friction within the labor market and
society.
How do we nurture an entrepreneurial mindset as
opposed to identity occupation?
For many people, their identity is densely
intertwined with occupation. Many families have
been working in a specific industry (e.g. assembly
line, mining coal, or serving in the military) for
generations. Is it possible to create a less fixed,
more entrepreneurial mindset that will help citizens
adapt?
What is the relationship between safety nets and
entrepreneurship?
Germany’s vocational training programs have
demonstrated enormous success at preparing
workers for employment in technical fields, and
thus served as an example for U.S. policymakers
and institutions. However, successfully reskilling
workers displaced by automation and technological
change remains a challenge for Germany and the
United States.
What are the areas of employment that are least
susceptible to being replaced by AI? Are these the
same across borders? Or do they differ because of
culture or regulation?
From a purely technical perspective, it should be
possible to make broad assumptions about which
jobs are susceptible to being replaced by algorithms
and robots. However, there may be cultural or
regulatory factors that impact the likelihood of jobs
being replaced by AI in the near future (to 2035).
Most of Silicon Valley’s most successful
entrepreneurs enjoyed the luxury of being able
There are serious questions regarding the labor market
implications of these technologies and how to reimagine
the social contract in a world where an increasing
proportion of economic productivity is accounted for by
machines and computers.
Roadmapping a U.S.-German Agenda for Artificial Intelligence Policy
9
Notes
Stouffer, Keith et al. Guide to Industrial Control Systems
(ICS) Security: Supervisory Control and Data Acquisition
(SCADA) systems, Distributed Control Systems (DCS), and
other control system configurations such as Programmable
Logic Controllers (PLC) (National Institute for Standards
and Controls 2015)
1
2
Autonomous Weapons: an Open Letter from AI & Robotics
Researchers (The Future of Life Institute 2015)
See, for example, Schumers Urging Drone Geo-fencing
Proposal Passes the Senate New Technology Could
Prohibit Drones from Flying Near Airports Sensitive Areas
Sporting Events and Above 500ft (Office of Charles E.
Schumer 2016)
3
Preparing for the Future of Artificial Intelligence
Recommendations 20, 21 (US National Science
Technology Council 2016); Federal Automated Vehicles
Policy (US Department of Transportation 2016): https://
obamawhitehouse.archives.gov/blog/2016/10/12/
administrations-report-future-artificial-intelligence,
https://one.nhtsa.gov/nhtsa/av/av-policy.html.
4
See, for example, Scherer, Matthew. Regulating Artificial
Intelligence Systems: Risks, Challenges, Competencies,
and Strategies. (Harvard Journal of Law and Technology,
Spring 2016): https://papers.ssrn.com/sol3/papers2.
cfm?abstract_id=2609777
5
See, for example, Artificial Intelligence, Robotics, Privacy
and Data Protection (38th International Conference of
Data Protection and Privacy Commissioners 2016)
6
See, for example, DeepMind AI Reduces Google Data
Center Cooling Bill by 40% (DeepMind 2016).
7
See, for example, Big Data: A Tool for Inclusion or
Exclusion? (Federal Trade Commission 2016).
8
See, for example, Artificial Intelligence and Life in 2030:
One Hundred Year Study on Artificial Intelligence (Stanford
University 2016).
9
10
NEW AMERICA
This report carries a Creative Commons Attribution 4.0 International license, which
permits re-use of New America content when proper attribution is provided. This
means you are free to share and adapt New America’s work, or include our content
in derivative works, under the following conditions:
• Attribution. You must give appropriate credit, provide a link to the license, and
indicate if changes were made. You may do so in any reasonable manner, but
not in any way that suggests the licensor endorses you or your use.
For the full legal code of this Creative Commons license, please visit
creativecommons.org.
If you have any questions about citing or reusing New America content, please visit
www.newamerica.org.
All photos in this report are supplied by, and licensed to, shutterstock.com unless
otherwise stated. Photos from federal government sources are used under section
105 of the Copyright Act.