Recommendations for an AI Strategy in

Switzerland
A white paper organised by
the SATW topical platform on Artificial Intelligence

Table of Content
Preamble 3

A Safe Harbour for Data 5

A Verification Body for AI 9

Increase Human Trust in AI 12

AI in Higher Education 15

Enable AI for SMEs 18

Contributors

Alessandro Curioni, Editor, IBM Research – Zurich

Lukas Czornomaz, Editor, IBM Research – Zurich
Joachim Buhmann, ETH Zurich
Ernst Hafen, ETH Zurich, DatenundGesundheit Assoc., MIDATA Genossenschaft
Manuel Kugler, SATW
Hervé Bourlard, Idiap Research Institute and EPFL
Jana Koehler, Deutsches Forschungszentrum für Künstliche Intelligenz (DFKI)
Matthias Kaiserswerth, Hasler Stiftung
Anika Schumann, IBM Research – Zurich

This document reflects a consensus between the contributors, but the responsibilities lie with the
authors of each part.
The positions and opinions in this document are those of the contributors and don't necessarily
represent IBM's positions, strategies or opinions.
This white paper has not undergone SATW’s usual review process, and SATW therefore assumes no
responsibility for the content or the quality of the document. Responsibility for these lies solely with
the authors.

Cover photo: freepik.com

2
Preamble1
Digital transformation is radically reshaping almost every Given the penetration of AI across most
aspect of our society. The explosion of artificial intelligence industries, its potential impact on GDP
(AI) and big data analytics applications is enabled by the promises to be very high2. In Switzer-
extreme availability of data in combination with the sub- land, AI is already reshaping industries
stantial computing power of modern highly distributed such as banking, insurance, pharmaceu-
computing infrastructures connected by high-speed net- ticals and manufacturing. Furthermore,
works. Machine learning technologies can be trained to per- Switzerland is the European country that
form specific tasks with an efficiency and an accuracy that has the highest number of AI start-ups
can supplement and, in some cases, outperform that of hu- per citizen3, with more than 100 start-
mans. These systems provide deep insights by learning from ups4. Many leading countries are heavily
data and interactions with users, which is already leading to investing in AI development strategies
a profound transformation of numerous industries, profes- and the establishment of technology
sions and society at large. The current state of AI is, howev- transfer centres in this field5,6,7,8,9,10,11,12,13
er, still far from delivering truly intelligent behaviour that is
comparable to human intelligence. An AI research strategy
should therefore carefully analyse AI’s history with its vari- 2
Chui, M., et al., (2018). NOTES FROM THE AI
ous waves of large promises and conceptual shortcomings. FRONTIER –INSIGHTS FROM HUNDREDS OF USE
CASES. McKinsey Global Institute. Retrieved from
Recent advancements in machine learning have enabled AI www.mckinsey.com/featured-insights/artificial-
intelligence/notes-from-the-ai-frontier-
technologies to become extremely successful. Speech applications-and-value-of-deep-learning
recognition, natural language interaction with machines and 3
The European Artificial Intelligence Landscape |
More than 400 AI companies built in Europe. me-
facial recognition based on deep learning are now com-
dium.com/cityai/the-european-artificial-
modities that have changed the way people interact. The intelligence-landscape-more-than-400-ai-
machine learning strategy of emulating human performance companies-build-in-europe-bd17a3d499b
4
100 start-ups on the Swiss Artificial Intelligence
by learning from human experience promises a solution to
Start-up Map, www.startupticker.ch/en/news/july-
the knowledge extraction problem. However, the automated 2017/100-startups-on-the-swiss-artificial-
reasoning process is as opaque as human decision making. intelligence-startup-map
5
Overview of national AI strategies. medi-
Evolution has enabled humans to collectively reason and act um.com/politics-ai/an-overview-of-national-ai-
on our collective experience, though other humans are of- strategies-2a70ec6edfd
6
US to endorse new OECD principles on artificial
ten black boxes. Today, we are confronted with computa- intelligence. www.politico.eu/article/u-s-to-
tional artefacts that are adapted to complex human decision endorse-new-oecd-principles-on-artificial-
intelligence/amp/
making and, thereby, have inherited a similar “black box” 7
Ethics guidelines for trustworthy AI.
behaviour. ec.europa.eu/digital-single-market/en/news/ethics-
guidelines-trustworthy-ai
8
Eckpunkte der Bundesregierung für eine Strategie
Künstliche Intelligenz.
www.bmwi.de/Redaktion/DE/Downloads/E/eckpun
ktepapier-ki.pdf?__blob=publicationFile&v=4
9
« La stratégie IA, pour faire de la France un acteur
majeur de l'intelligence artificielle ».
www.enseignementsup-
1
Preamble – lead author: Lukas Czornomaz
recherche.gouv.fr/cid128618/la-strategie-ia-pour-

3
To date, Switzerland has not developed a dedicated AI To improve the image of AI, we need to
strategy. AI is one of many topics covered in the strategy increase human trust in AI. A positive
“Digitale Schweiz”. An interdepartmental working group on message about how AI empowers hu-
AI which should ensure knowledge exchange in the domain mans to become more efficient, makes
of AI within the federal administration and coordinate Swit- knowledge more accessible and im-
zerland's positions in international bodies, is mandated to proves quality of life needs to be com-
submit a report to the Federal Council by September 2019. municated. The profound impact of data
Furthermore, an interdisciplinary study on behalf of TA- science on society requires additional
SWISS is evaluating the opportunities and risks of AI on the educational efforts to enable the popula-
basis of various focal points: work, education, media, con- tion to meaningfully use this AI technol-
sumption and administration. The publication of that study ogy. Corresponding challenges lead to a
is planned for the end of 2019. demand for AI in higher education but
also a profound revision of the topic cat-
The contributors to this publication are convinced that Swit-
alogues in grammar schools. The posi-
zerland can be a frontrunner in selected fields of AI services
tive effect of AI on the Swiss economy
and applications. However, to attain this position, a struc-
could be strongly accelerated if we were
tured approach and common direction for efforts in AI re-
to create an appropriate framework
search and technology development are required. This doc-
through which to ease access and ena-
ument formulates recommendations for an AI strategy for
ble AI for SMEs. AI empowers SMEs to
Switzerland. The goal is to position Switzerland as a leading
compete better with larger market actors
AI country and to amplify and accelerate the positive impact
and to accelerate their growth.
of AI on the Swiss economy. The recommendations are de-
livered in the context of five important aspects of AI.

Switzerland should pioneer a model for an open market for

data, stimulating the valuation and exchange of data while
ensuring privacy, security and trust. Thus, we propose lever-
aging Switzerland’s unique historical geopolitical reputation
and trust to promote itself as a safe harbour for data stor-
age. Furthermore, Switzerland should take the lead in the
definition of the general requirements that AI systems
should follow to enhance their acceptability by businesses
and society. As AI is being more widely deployed, it be-
comes crucial to ensure that these models meet high ethical
standards and provide safety. A verification body for AI
could provide requirements for trusted ethical and safe AI
systems.

faire-de-la-france-un-acteur-majeur-de-l-intelligence-
artificielle.html?menu=4
10
Pan-Canadian Artificial Intelligence Strategy. www.cifar.ca/ai/pan-
canadian-artificial-intelligence-strategy
11
United Kingdom AI Sector Deals. www.gov.uk/government/
publications/artificial-intelligence-sector-deal
12
US President’s approach to AI. www.whitehouse.gov/wp-
content/uploads/2018/05/Summary-Report-of-White-House-AI-
Summit.pdf
13
New Generation Artificial Intelligence Development Plan.
www.gov.cn/zhengce/content/2017-07/20/content_5211996.htm

4
A Safe Harbour for Data
The status quo and Switzerland’s future in a global AI-
dependent digital society14

Individuals are a major source of data for AI solutions in have been at the forefront of cloud com-
various fields, including healthcare, education, urban plan- puting and machine learning, the tech-
ning, energy, and economics. This chapter focuses on how nology and knowledge gap between
personal data can best be harnessed in a fair and sustaina- large multinational cloud storage and AI
ble way that protects individuals’ privacy and digital self- providers and local companies is rapidly
determination and how Switzerland can play a leading role expanding.
in the concomitant democratization of the personal data
In the European Union, the rights of in-
economy.
dividuals to control their personal data
Personal data are a new asset class , and its market value is
15
have been strengthened with the enact-
estimated to reach more than one trillion Euros in 2020 . 16
ment of the European General Data Pro-
Currently, personal data are stored in incompatible silos and tection Regulation (GDPR) in May 2018.
are increasingly fuelling the profits of a few multinational da- In particular, the right to data portability
ta companies that offer free services and smartphone apps in gives EU citizens the opportunity to ob-
exchange for personal data. This approach has led to a rapid tain a digital copy of all their personal
increase in socioeconomic asymmetry in the concentration of data. In the revision of the Swiss data
and control over personal data . With the increasing use of
17
protection law, which is currently being
AI in healthcare, education and other disciplines, this asym- discussed in parliament, the federal
metry will increase dramatically since the companies and in- council has omitted the data portability
stitutions with the largest amount of data will be able to de- provision on the grounds that it de-
velop the best algorithms . This asymmetry threatens to
18
creases the competitiveness of small and
leave little opportunity for smaller institutions and countries medium enterprises, a view that is also
to catch up or compete, particularly countries such as Swit- shared by Economiesuisse19. Thus, in
zerland. Although academic research groups in Switzerland Switzerland, individuals currently possess
significantly less autonomy in the control
of their own data than EU citizens. In a
14
A Safe Harbour for Data – lead author: Ernst Hafen
future digital global society that is heavi-
15
World Economic Forum. (2011). Personal Data: The Emergence of a New
Asset Class (pp. 1–40). World Economic Forum. ly dependent on AI solutions involving
The Boston Consulting Group. (2012). The Value of Our Digital Identity personal data, individuals will have to
16

(pp. 1–65). Retrieved from www.libertyglobal.com/PDF/public-policy/The-

Value-of-Our-Digital-Identity.pdf play a central role as the aggregators
17
Haynes, P., & Nguyen, C. M.-H. (2013). Rebalancing Socioeconomic
Asymmetry in a Data-Driven Economy. In B. Bilbao-Osorio, S. Dutta, & B.
Lanvin (Eds.), The Global Information Technology Report (pp. 67–72). 19
Djonova, I., & Herzog, E. (n.d.). Eine Datenpolitik
18
Lee, K.-F. (2017, June 24). The real threat of artificial intelligence. New des Vertrauens für Fortschritt und Innovation (No.
York Times. Retrieved from # 03/2018). Retrieved from
www.nytimes.com/2017/06/24/opinion/sunday/artificial-intelligence- www.economiesuisse.ch/de/entityprint/node/4542
economic-inequality.html 5

5
and access controllers of their personal data. In the current Medical data stored in incompatible silos
financial economy, most citizens possess a bank account or on paper has hindered systematic
and decide how to invest or spend their money. It is the in- outcomes research and medical research
dividual consumer who drives the economy. in general. The only datasets that are
also in demand internationally are those
The role of citizens20,21 in a data-driven society and econo-
of cohorts23 funded by the Swiss
my is more central than in the current consumer economy,
National Science Foundation, which
which is still dominated by physical goods, because three
provide a longitudinal dataset of AIDS
unique features of personal data set them apart from goods
patients and healthy individuals and
and financial assets. First, personal data are a nonrivalrous
typically contain data from several
good. In contrast to money and other physical assets, data
thousand participants.
are not consumed; they can be copied and reused. This fact
makes the article on Data Portability in the EU GDPR a true A trust-promoting framework that
innovation since it legally entitles data subjects to copies of supports the fair and active participation
all their personal data22. Second, personal data are a new of citizens relies on several of the qualities
asset class that is equally distributed among individuals. In for which Switzerland is recognized
contrast to the uneven distribution of financial assets, our internationally. The one person one vote
common biology implies that each individual possesses a principle of direct democracy fits well
genome that contains six billion base pairs, has a heart that with the equal distribution of personal
beats with a similar frequency, consumes similar numbers of data. Historically, Switzerland has been a
meals (even though caloric intake differs), etc. Third, and forerunner in democratically controlled
most importantly, individuals are the maximal aggregators cooperatives. Some of Switzerland’s
of their personal data. Only individuals have the potential alpine farming cooperatives are several
and the legal right to aggregate medical, social media, con- hundred years old24. As a repository and
sumer, and genome data. AI solutions in areas such as platform for sharing personal data,
healthcare, education or urban planning are increasingly cooperatives offer two advantages over
dependent on the aggregation of such different data types other organizational forms. Cooperatives
from millions of people. Thus, the individual, as an active belong to their members. Much like Swiss
producer and aggregator of personal data and consumer of citizens who control their government,
data services, will play an even more important role in a fu- cooperative members manage their data
ture AI-supported society. For this to happen, however, and control cooperative governance and
there is need for a new trust-promoting framework in which how revenues are invested.
citizens play an active role in the personal data economy.

20
Toffler, A. (1980). The third wave.
23
21
Tapscott, D. (1995). The Digital Economy. Example 1: The Swiss Digital Health Cohort,
22
De Hert, P., Papakonstantinou, V., Malgieri, G., Beslay, L., & Sanchez, I. https://www.satw.ch/fileadmin/user_upload/docum
(2017). The right to data portability in the GDPR: Towards user-centric in- ents/02_Themen/08_Kuenstliche-Intelligenz/SATW-
teroperability of digital services. Computer Law & Security Review: The Swiss_AI_Strategy-Example1.pdf
International Journal of Technology Law and Practice, 34(2), 1–11. 24
Ostrom, E. (2015). Governing the Commons (Re-
doi.org/10.1016/j.clsr.2017.10.003 issued edition). Cambridge.

6
Recommendations for networked personal data coop-
eratives

The Swiss MIDATA cooperative25, with its non-profit and education. Early projects and platforms
ethical governance principle, aims to be the founding ex- by CERN and the University of Geneva26
ample of networked personal data cooperatives in other as well as ETH Zurich and the University
countries, thus promoting the democratization of the per- of Zurich27 to promote and strengthen
sonal data economy. Moreover, Switzerland’s trusted role as the participation of citizens in science
a safe harbour and provider of financial services could be are encouraging.
extended to the management of personal data. Finally, Swit-
Two problems remain that the Swiss na-
zerland’s neutrality and its credibility in international organi-
tional parliament must address. First and
zations will help promote the democratization of the per-
foremost, it is essential that Switzerland
sonal data economy.
adopt the data portability framework
Given that personal data is a non-rivalrous good, coopera- of the EU GDPR. Switzerland should do
tives in which citizens control the aggregation of and access so, not only because it is the only coun-
to their personal data do not replace but extend the existing try in Europe that does not grant its citi-
global personal data ecosystem. All existing providers of zens the fundamental right to a digital
data and AI solutions will benefit from such an extension, copy of their personal data but more
since they obtain access to new data aggregates. Democrat- importantly because it lays the founda-
ically governed cooperatives acting as the fiduciaries of their tion for a citizen-controlled extension of
member’s data can ensure that the socioeconomic asym- personal data and AI ecosystem.
metry in a global data and AI-dependent society is re-
Second, the rapid introduction of an
balanced and that the economic benefits of these data are
electronic identity (eID), which is cur-
also returned to society at large.
rently being discussed in parliament, is
Switzerland possesses an active and internationally com- essential. Countries, such as Estonia, that
petitive academic research community in machine learning, possess such an eID have seen a trans-
AI, cloud computing and data security. Recent investments formative shift towards a digital society,
in the Swiss Data Science Center and the Swiss Personal with many start-ups and international
Health Network initiative are further steps in strengthening companies offering new services. More-
academic research in the areas of AI. As outlined above, it is over, politicians and the media need to
essential to empower and enable citizens to become ac- realize that, while it is important to talk
tive data aggregators and participants in all areas that rely about privacy and data protection, it is
on the availability of personal data, including healthcare and equally important to talk about the po-

25
Example 2: Box – MIDATA Personal Data Cooperative, 26
The Citizen Cyberlab, www.citizencyberlab.org
https://www.satw.ch/fileadmin/user_upload/documents/02_Themen/08_Ku 27
The Citizen Science Center Zürich, citizen-
enstliche-Intelligenz/SATW-Swiss_AI_Strategy-Example2.pdf science.ch

7
tential benefit of open, transparent and fair data sharing. smartphone sensors and medical records
Many studies have shown that people are willing to share willingly and knowingly provided directly
data for the benefit of medical research and for personal- by people will be orders of magnitude
ized information , . 28 29
greater.

Finally, financial support for the effective initiation of the However, the path to establishing such a
transformation to a fair democratically controlled per- democratically controlled data economy
sonal data ecosystem is essential. Once established, a new cannot be financed by classical invest-
democratically controlled personal data economy that in- ment strategies that rely on venture cap-
volves hundreds of millions of people across the globe will ital equity investments. Cooperatives
generate substantial benefits because of the data aggrega- belong to their members, one vote at a
tion power of citizens and because they will not only con- time, and thus cannot accept equity in-
tribute data but also their human intelligence. Run by citi- vestors. Therefore, there is a need for
zen-owned non-profit cooperatives, these benefits will be foundations, philanthropists and crowd
returned to society. Examples include real-world patient- funding to bridge this initial gap to
reported outcomes via smartphone apps and sensors from achieve the breakeven point of data co-
medical treatments, drug efficacy or the active recruitment operatives.
of people for clinical trials. Today, pharmaceutical compa-
nies are making deals with data companies for hundreds of
millions to billions of dollars to access digital health rec-
ords30,31. The market value of real-time health data from ogy. (2018, June 19). Roche and Foundation Medi-
cine reach definitive merger agreement to acceler-
ate broad availability of comprehensive genomic
28
Vayena, E., Ineichen, C., Stoupka, E., & Hafen, E. (2014). Playing a part in profiling in oncology.
research? University students' attitudes to direct-to-consumer genomics. 31
GSK and 23andMe sign agreement to leverage
Public Health Genomics, 17(3), 158–168. doi.org/10.1159/000360257
genetic insights for the development of novel
29
Mooser V, Currat C. The Lausanne Institutional Biobank: a new resource medicines | GSK. (2018). Retrieved from
to catalyse research in personalised medicine and pharmaceutical sciences. www.gsk.com/en-gb/media/press-releases/gsk-
Swiss Med Wkly. 2014;. doi.org/10.4414/smw.2014.14033
and-23andme-sign-agreement-to-leverage-
30
Roche and Foundation Medicine reach definitive merger agreement to genetic-insights-for-the-development-of-novel-
accelerate broad availability of comprehensive genomic profiling in oncol- medicines/

8
A Verification Body for AI
Towards ethical and safe AI32

New risks arise with the use of AI and decision-making algo- to foster ethics in AI33,34,35,36,37. However,
rithms. Aspects such as bias or interpretability are nontrivial there are few concrete projects establish-
and require a thorough examination. AI systems should follow ing a holistic view for addressing general
general ethical and safety requirements that enhance their ac- ethics and safety requirements across in-
ceptability by businesses and society. To date, no clear gener- dustries and applications other than some
ally accepted guidelines for the implementation of AI systems limited projects with specific applications,
exist. Therefore, there is a need to define the essential re- e.g., autonomous cars38.
quirements that must be fulfilled by an AI system and the as-
In different countries and cities, laws are
sociated verification processes. The definition of such ethical
implemented to guide the application of
and safety requirements and their verification implies a pro-
machine learning and the accountability of
found technical expertise but also a more societal, political and
algorithms39,40. Currently, we do not know
philosophical point of view regarding how democracies can
of any concrete measures for the imple-
adapt to coexist with AI and benefit most from it.
mentation of an interdisciplinary approach
The topic of the requirements, validation and verification of AI to specify the sufficient and practical re-
systems is related to software liability, where providers typically quirements for the safety of self-learning
are not liable for bugs in their systems or consequential dam- systems and to address the validation and
age caused by a failure of their systems. Medical software tools verification of AI systems in Switzerland.
or apps also have similar challenges, and their effectiveness
must be provable to be accredited as a medical product.
33
The partnership on AI is a multi-stakeholder
Providing broadly applicable requirements and verification
organization to better understand AI’s impacts.
processes could advance the development and dissemination www.partnershiponai.org
of AI systems in Switzerland and thereby support and
34
Informatics Europe recommends to establish
means, measures and standards to assure fair ADM
strengthen the local economy. The global trust in its institu- systems. www.informatics-europe.org
tions and democratic system could help to position Switzer- High-Level Expert Group on Artificial Intelligence
35

supports implementation of European AI strategy.

land as a global role model that defines widely acceptable ec.europa.eu/digital-single-market/en/high-level-
guidelines and procedures for AI systems. group-artificial-intelligence
36
MIT Turing Box as a certification environment.
turingbox.mit.edu
There are many activities in the field of trust in AI; increasing 37
The IEEE Global Initiative on Ethics of Autono-
the reliability of algorithms is a prerequisite for their deploy- mous and Intelligent Systems. stand-
ment in critical applications. Various experts and initiatives aim ards.ieee.org/industry-
connections/ec/autonomous-systems.html
38
TÜV SÜD and DFKI to develop “TÜV for Artificial
Intelligence”. www.tuv-sud.com
39
European General Data Protection Regulation.
32
A Verification Body for AI – lead author: Manuel Kugler eugdpr.org
40
New York City Council, Fairness in computing.
The authors acknowledge Markus Christen (Managing Director Digital
legistar.council.nyc.gov/LegislationDetail.aspx?ID=
Society Initiative, University of Zurich) and Christian Westermann (Leader
3137815&GUID=437A6A6D-62E1-47E2-9C42-
Data & Analytics, PwC Switzerland) for their valuable inputs to this section.
461253F9C6D0

9
Essential general requirements of AI systems

General requirements and associated verification and vali- data. Bias and explainability are essential
dation procedures should cover cross-sectorial industries, requirements linked to ethics and liabil-
be technologically independent and apply to systems of ity.
different providers. They should differentiate AI systems
Security aspects are crucial, particularly
based on whether they are truly self-learning, adaptive or
in cybersecurity. A successfully tested AI
fixed. The general guidelines should also provide a frame-
system should be verified as safe against
work to determine whether or not a human is required in
known adversarial attacks. The perfor-
the loop depending on the criticality of the application and
mance of the algorithms and their adap-
severity of the consequences of a failure. Decision-making
tation needs to be verifiable while main-
systems that are truly self-learning pose the greatest chal-
taining individuals’ data privacy and the
lenges. In such cases, a decoupling of decision making and
intellectual property rights of the system
autonomous action-taking might be required.
developers. The verification process
Another key requirement is to demonstrate the system’s should be as transparent and explainable
flexibility in reacting to new situations. The algorithms as possible. Thus, the key components of
should be robust and reliable when facing shifts in data or the verification and validation proce-
manipulations of processed data, i.e., the algorithms have to dures should be based on open source
generalise. The algorithms should be able to detect when systems.
the input data are outside of the trained capabilities and
There are numerous incentives for Swit-
defined boundaries. Learning algorithms must be enhanced
zerland to establish a place to define
by an automonitoring capability to assess the uncertainty in
requirements and verify AI systems. Giv-
the algorithm’s knowledge of the domain/world. Often, it is
en its neutrality and reputation as a me-
essential to know what one does not know. To minimize the
diator, Switzerland is in a very favourable
risk of a system failure in such an event, a domain-specific
position. Being known worldwide for
process is required, e.g., a feedback loop so humans can
excellence in product quality and ser-
remain in control of the system’s response.
vices, trust in Switzerland as a reliable
The control of bias in data or AI systems must be consid- state with profound ethical standards is
ered. AI systems should be fair, and they should not fall prey high and could also apply to Swiss-
to unintended discrimination. There must be a strong ethi- verified AI systems. Being small and in-
cal definition of what an AI system should and should not dependent, Switzerland is more agile
decide. In particular, the responsibility for a system failure than other European countries. In addi-
needs to be clarified. Along with such design principles, all tion, Switzerland has internationally
AI systems should be able to explain the reasons, at an ade- known research institutes in the field of
quate level of detail, of why they made certain decision and AI, and various large technology compa-
what the determining factors were in the input and training nies that are leaders in AI have research
facilities here. Bringing together industry,

10
research institutes and technology companies as well as
governmental institutions to set up a requirements and veri-
fication process is demanding but crucial and in the interest
of all the involved parties.

Based on the statements above, we recommend defining an

institution at the governmental level in Switzerland to
assume responsibility for the requirements and verification
of AI systems.

- The institution should provide clear but flexible

general guidelines for testing algorithms.

- The guidelines should be extended by industry-

specific standards and include different levels of
requirements depending on the criticality of the use
cases and the severity of the consequences in case
of failure.

- Standards and guidelines can be verified by existing

industry-specific service providers and testing bod-
ies.

- The proposed body could be organised as an extra-

parliamentary committee41 or existing verification
bodies42.

An expert committee consisting of academic and industry

specialists in AI core technology, industry use cases, securi-
ty, societal implications and ethics should be installed to
develop an implementation plan for such a requirements
and verification institution. SATW would be willing to organ-
ise and coordinate such a committee in collaboration with a
government body.

41
www.admin.ch/gov/de/start/bundesrecht/ausserparlamentarische-
kommissionen.html
42
National Institute of Standards and Technology, U.S. Department of
Commerce. www.nist.gov/topics/artificial-intelligence

11
Increase Human Trust in AI
An ambitious but necessary objective43

The availability of affordable, highly parallel computing, vast or that the resulting systems will be ro-
amounts of training data, and economic interest from large bust to adversarial attacks, as is expected
corporate actors in “big data” has led to large-scale AI sys- from any critical software? Security is a
tems capable of solving problems that were, until recently, central topic in practice, as an AI system
considered unsolvable for at least a decade, if ever. remains a complex piece of software
running in a failure-prone and vulnerable
It is predicted that these novel and complex AI technologies
physical device.
will impact virtually every human activity, from health and
education to manufacturing and warfare, and will become Trust in the impact of AI on society.
central in the lives of citizens and consumers. Our common sense has been shaped
over thousands of years to handle the
In this context, it will be critical to improve the AI image in
risks and benefits of dealing with either
public discourse; hence, we are also focusing on “human
objects devoid of cognition or intelligent
trust in AI” by studying how to improve the multifaceted
human beings. AI creates a new category
trust between human users and “AI systems” by making the
to which our most fundamental behav-
systems more robust, more intelligible to human minds, and
iours have not (yet) been adapted. How
more aligned with society’s needs. This ambitious goal
can we prepare citizens for pervasive
should develop along three complementary axes:
intelligent systems that will influence

Trust in AI as an abstract object. A clear quantification of them in all their decisions, including per-

AI’s technical performance is fundamental to allowing hu- sonal and societal decisions? How can

mans to anticipate its failures and conversely to trust it we ensure that AI follows basic ethical

when it operates in a properly functioning regime, just as principles and adapts to societal chang-
one constantly infers the behaviour of other humans to es? How can we forecast the disruptions

safely interact with them. Research and development should AI may cause to the economy?

place particular emphasis on the challenge of determining

Additionally, we believe that reproduci-
the uncertainty in the learned knowledge of AI agents. How
bility and openness in the context of ex-
can we specify AI’s expected performance? How can we en-
tremely complex AI models and large-
sure that this level of performance is reached? Additionally,
scale data corpora are critical and re-
at a more fundamental level, how can we create a common
quire a strong commitment to and in-
understanding between humans and AI?
vestment in “open science” tools and

Trust in AI as an operational system. Can we trust that practices that will be central to the pro-

there will be proper handling of the required mass of data, ject.

particularly regarding privacy laws and intellectual property,

43
Raise Human Trust in AI – lead author: Hervé Bourlard

The author acknowledges Dr. François Fleuret from the Idiap Research
Institute for his inputs to this section.

12
How to improve human trust and AI in publics

It is clear that Switzerland has reached a critical mass in re- Content creation and information access:
gard to the technical expertise required to improve AI and AI will play a fundamental role in content
make it trustworthy and safe. However, at the national level, creation, enabling easier access to all
there is no initiative that brings the different players togeth- kinds of data for everyone. As a conse-
er in a common forum that could act as a force multiplier. quence, knowledge will be “democra-
Even more problematic is the scattered expertise and the tized”, potentially also improving the
lack of coordinated research on the implications that AI will enforcement of data privacy.
have for various aspects of our society, ranging from the
Positive changes in the quality of life:
economy to law to ethics, which is a situation that, never-
There are a growing number of AI appli-
theless, reflects what is typical of AI research worldwide.
cations that actively improve people’s
In this context, the expert committee proposed for the veri- lives and create positive changes in the
fication of AI systems would have the right to take a leading world. For instance, AI systems will help
role in increasing human trust in and public awareness of AI. improve clinical diagnoses (e.g., applying
Efforts should be focused on sharing positive messages AI to various types of healthcare data,
about the impact of AI on society, increasing the transpar- cancer detection), medical prevention, or
ency of AI solutions by generally requiring explainability in the diagnosis of faults in physical sys-
software, and finally encouraging societal debates about tems. AI also has the potential to im-
wider social consequences of AI technologies. prove farming and food production pro-
cesses. Finally, the security of people
Share positive AI messages44:
(including children’s safety), systems

Empowering humans: AI will empower humans by realizing (communication), and criminal activity

technologies that benefit humanity instead of destroying prevention will be improved.

and intruding on the human rights of privacy and freedom

More efficient society: Through AI, socie-
to access information.
ty will have the potential to become

Open source information45: Open source information and AI more efficient, for instance, by optimiz-
collections will provide opportunities for new technological ing traffic and transportation systems,

progress and global technological parity. reducing waste, and improving ecologi-
cal behaviours. AI has great potential to
support human society in facing the
challenge of climate change and other
44 global threats.
14 Ways AI Will Benefit or Harm Society, Forbes Technology Council,
www.forbes.com/sites/forbestechcouncil/2018/03/01/14-ways-ai-will-
benefit-or-harm-society
45
Top 8 Open Source AI Technologies in Machine Learning, open-
source.com, opensource.com/article/18/5/top-8-open-source-ai-
technologies-machine-learning

13
Increase transparency and encourage societal debates46: enhance the quality of life and wellbeing
of individuals and society rather than in
Embed transparency in software: Sophisticated AI often
ways that damage lives.
takes complex and sometimes obscure routes (software,
data, etc.) in its methodologies, data mining, and algo- Focused studies, monitoring and analy-
rithms. We should encourage ways to embed transparency sis47: AI guidance should be developed
and clarity into the software (as well as hardware). For in- through focused and multidisciplinary
stance, when AI makes life-changing judgments around studies, monitoring, and analysis. Max-
sentencing, welfare, and medical decisions, it is fair to as- imizing the impact of AI developments
sume that those affected need to know how these decisions and their acceptability by society will
were made. indeed require engagement with cross-
disciplinary groups, including computer
Address multifaceted AI concerns: Encourage debates
scientists, social scientists, psychologists,
across society to ensure that we use the various AI technol-
economists, and lawyer.
ogies (which are changing the way we behave, work and
interact with others) in ways that are ultimately beneficial Increase transparency and data privacy:
and that consider many kinds of concerns. In that direction, the EU General Data
Protection Regulation (GDPR)48 is proba-
Clearly identify long-term impacts of AI developments: The
bly the most important change in data
technology industry must recognise the long-term impacts
privacy regulation in 20 years.
of AI developments and assume responsibility for the wider
social consequences of its work. This goal requires policy
makers to start thinking seriously about the challenges that
may arise from the advancement and increasing application
of AI across all industrial sectors.

Engage business and citizens in discussions: Businesses and

citizens need to become more broadly engaged in the dis-
cussions, recognising their common stakes in defining the
future. Thanks to AI, we have the opportunity to redesign
many aspects of our world and to make the most of the op-
portunities provided by new technologies. We must focus
on making these changes in ways that we all trust and that

47
AI, People, and Society, Eric Horvitz, Science, July
46
Could transparency make AI safe and reduce public fears ? TechWorld, 2017: Vol. 357, Issue 6346, sci-
October 2017, www.techworld.com/data/could-transparency-make-ai-safe- ence.sciencemag.org/content/357/6346/7
48
reduce-public-fears-3665487/ eugdpr.org

14
AI in Higher Education
A core research topic49

Machine learning and AI are currently fundamentally chang-

ing the way we think about computing at large, especially
how we approach algorithm design, software engineering
and computer systems engineering. Modern intelligent de-
cision-making systems process and interpret high volume
data streams for computer vision, natural language and
speech processing, robotics, and health data analysis and,
more generally, for flexible data-centric model building in
engineering, natural science and, for the first time, in the
humanities. Such systems commonly process heterogene-
ous data with substantial uncertainty caused by measure-
ments. Classical concepts of how to establish the correct-
ness of planning and decision-making algorithms when they
process uncertain data often rely on worst-case guarantees
that do not seem to be adequate and predictive enough for
many real-world applications. Self-adapting AI algorithms
must generalize well over fluctuations and model misspeci-
fications; they should be resilient by proper regularization to
avoid overfitting, i.e., “reading the tea leaves”. On the other
hand, versatile intelligent systems should also be sufficiently
flexible to exploit the signal in data to its full extent for pre-
diction. Intelligent decision-making systems should provide
(provable) guarantees to deliver typical answers with a high
predictive value; however, scientifically, it is still mostly un-
known what that means for algorithm design, and we need
large research programs to make progress on these ques-
tions beyond the standard paradigm of supervised machine
learning. Research on these fundamental questions will also
strengthen public trust in this technology that is about to be
deployed for autonomous transport, for health care applica-
tions, for societal infrastructure management and for public
and private services, including entertainment.

49
AI in Higher Education – lead author: Joachim Buhmann

15
AI education should be modernized

Where does higher and continuous education stand in this Current educational programs in higher
scientific revolution? Computational thinking has clearly education have to evolve to strengthen
amended the scientific method, the unforeseeably success- probabilistic computational thinking.
ful methodology of industrialized societies, by adding a Such a mindset will be vital for our future
third pillar to empiricism – experimentation, theory building society since it will largely determine the
and computational modelling. In light of the big data chal- value generation of future products and
lenge, we must revisit all aspects of computational thinking service. Are we prepared for these chal-
and fuse it with probabilistic reasoning. Lessons from the lenges in higher education? Current cur-
history of AI indicate that the reduction of intelligent behav- ricula in computer science emphasize a
iour to logical calculus delivers successful expert systems scientific methodology for algorithms
but can also be interpreted as one of the causes of the AI and systems design and their analysis,
winter since neither the scientific community nor the gen- which mostly focuses on the efficient
eral public has seen its high expectations of AI fulfilled. Hu- usage of computational resources such
man intelligence clearly employs rational logical calculus, as memory requirements and processing
but it seems to be even more governed by our subcon- time. Data science, as a necessary pre-
scious abilities to imitate intelligent behaviours and, more requisite for intelligent reasoning, plan-
importantly, by our capacity to teach such behaviours to ning and acting in the real world, de-
other human beings across generations. Deep learning, with mands resilient algorithms and systems
its inherent connectionism, emulates this style of thinking as design in computation. These two scien-
opposed to the symbolic reasoning strategy and connects tific cultures must be fused into a coher-
to the age of cybernetics before symbolic AI. We see predic- ent picture of computational reasoning
tive patterns in data apparently without using a consciously and inference. Resource efficient compu-
accessible, “rationally accessible” theory; for example, medi- tation should be balanced with the pre-
cal doctors are top experts in using their experience to dictive power of computational solu-
achieve successful benefits in health care with often very tions. In many applications, this trade-off
little theoretical understanding that would explain the suc- leads to a win-win situation in which su-
cess in diagnosis, prognosis and therapy. perior predictions can be computed in a
more efficient way than with traditional
algorithmic approaches.

16
Getting our students and workforce ready for AI

Are our students in higher education well prepared for this ticular efforts to empower all students in
epochal paradigm shift? The answers to this question depend higher education to participate in these
on their specialization of studies and the role of model build- societal transformations.
ing by computation in the respective field. Required
What are the needs of the workforce in
knowledge and skills could range from the developer level,
the private and public sectors? AI tech-
where novel AI solutions are created and deployed, to various
nologies have enriched the start-up cul-
grades of the expert user level, where there is creative usage of
ture with novel models of collaboration
AI tools in various application domains. Science, technology,
and communication. Work teams are now
engineering and mathematics (STEM) students mostly acquire
forming over virtual platforms, and con-
a solid basis for this paradigm shift due to rigorous mathemat-
tinuous education might be provided
ical training.
through such platforms. AI tools affect
Future curricula reforms should be assessed according to crite- the job profile of intellectuals and highly
ria that strengthen these educational goals. The long tradition skilled brain workers. The critical thinking
of deterministic computational thinking has to be comple- skills of such employees will be essential
mented by a rigorous education in probabilistic modelling and in this transformation when algorithms
the design of probabilistic algorithms that can deliver typical start to replace experts in their decision-
solutions and readily adapt to changing input data properties, making roles. Imagine a radiologist inter-
demonstrating at least traces of intelligence. In addition, a acting with an automated radiology scor-
knowledge of systems theory and (optimal) control should also ing program: for various tasks, the algo-
be substantially required in disciplines outside of engineering. rithm already delivers superior perfor-
mance in scoring radiology images due to
The evolution of the scientific method by probabilistic compu-
its speed and enormous storage capacity.
tational thinking has to influence curricula in all scientific disci-
The humans in the loop have to be able
plines, ranging from physics to theology. Digitalization and
to evaluate the machine’s decisions, and
intelligent information processing affect nearly all areas of hu-
such a competence for quality assurance
man thought and will leave a profound footprint on all disci-
requires profound computational training.
plines. Novel ways to conduct research are expected to emerge
Continuous educational programs will
in the humanities when intelligent search algorithms can scan
serve such an educational need. In partic-
large digital libraries orders of magnitude faster than human
ular, aging societies should allocate the
investigators; sociological experiments can be conducted at
necessary resources for continuous edu-
the scale of societies. Research strategies from the natural sci-
cation to remain competitive with socie-
ences and engineering might substantially enrich the method
ties of developing countries that are usu-
catalogue of the humanities, and conversely, a novel common
ally characterized by much younger
language will also enrich the STEM disciplines with advances in
populations than the highly industrialized
computational humanities. As a society, we should make par-
world.

17
Enable AI for SMEs
Create an appropriate environment50

AI has not been an important subject in higher education in access to AI-related technological know-
Switzerland. Few educational courses have been available in how. Such efforts are rarely beneficial to
the last 20-30 years, and those only showed single aspects SMEs, which have limited budgets.
of AI in light of a particular approach. Furthermore, comput-
er science has not been viewed as a strong and important
field of its own but often been subsumed under other sci-
ences. Only very recently has this approach started to
change, for example in the “Lehrplan 21”.

Consequently, we do not find much representation of com-

puter science skills among the management of SMEs in
Switzerland. Notably, a deep understanding of algorithms,
modelling, or software architecture is a rare skill that is
highly in demand. This lack of skills makes it very difficult for
SMEs to assess what AI can truly do to bring innovation to
their businesses and what type of problems can be solved
by the current state of AI technology. Many industry repre-
sentatives are deeply influenced by the deep learning hype
and believe that this technology is all that AI is about and
that it will solve all problems. Furthermore, the hysteria
about AI as a job killer encourages many companies to pre-
fer to look into optimization and automation scenarios but
less into innovation. Becoming more efficient and flexible is
important; however, such a mindset does not necessarily
lead to innovative products and services.

Numerous companies struggle to understand the impact of

digitalization and novel technologies on their business.
These companies need to link technology (computer sci-
ence, AI and others) to the transformation scenarios that
may turn out to be highly relevant for them. For example,
consider booking.com (or uber.com), where brokers have
taken over a market and created a difficult situation for the
local Swiss hotel industry. A similar effort is currently un-
derway in Switzerland to establish a brokering service for

50
Enable AI for SMEs – lead author: Jana Koehler

18
Concrete recommendations to enable AI for SMEs

- Offer courses for postgraduate education that teach AI as - Link to international sources to make it
part of computer science and offer a broad and adequate easier to learn about AI applications
picture of the technology. created by SMEs in other countries.

- Adequately inform SMEs about AI safety, technological - Create a network of experts and
limitations and risks. knowledge resources that SMEs can
access at low cost without paying high
- Show examples51 of innovative and deployed applications
membership or brokering fees.
of AI to broaden the understanding of the opportunities
that AI can offer (do not oversell current laboratory proto- - Create one strong industrial represen-
types that are present in the media, e.g., in medical diag- tation in IT to support the exchange of
nosis). knowledge and networking as op-
posed to the current situation with
- Understand how AI, digitalization, industry 4.0, and novel
scattered IT industry associations.
business models interact for SMEs52,53.

- Link AI with economics and show transformation scenari-

os that help SMEs to reposition themselves if necessary.

- Describe and illustrate how AI and other technologies will

impact typical professions in Switzerland and show what
new professions will be created; link this information to
educational offers.

- Start with key industries in Switzerland to drive the trans-

formation of the Swiss economy through AI.

- Help SMEs look beyond the hype and provide methods

that enable them to evaluate technology based on their
needs and opportunities.

51
Industry 4.0 – Germany Market Report and Outlook.
www.gtai.de/GTAI/Navigation/EN/Invest/Service/Publications/business-
information,t=industrie-40--germany-market-report-and-
outlook,did=917080.html
52
Statusreport "Arbeitswelt Industrie 4.0". shop.vde.com/de/statusreport-
arbeitswelt-industrie-40-download

Smart Service Welt – Internetbasierte Dienste für die Wirtschaft.

53

Deutsche Akademie Der Technikwissenschaften.

www.acatech.de/Projekt/smart-service-welt/

19