General Principles For The Use of Artificial Intelligence in The Financial Sector

General principles for the
use of Artificial Intelligence

in the financial sector
General principles for the use of Artificial Intelligence in the financial sector
©2019 De Nederlandsche Bank
Author: Joost van der Burgt

Contents
Practical information 5
Summary6
1 Introduction9
2 AI: past, present and future 11

2.1 Past (1950-2009): 60 years of anticipation 11
2.2 Present (2010-2019): The coming of age of AI 15
2.3 Future (2020-beyond): Ubiquitous AI 21
3 AI in finance 25
3.1 Current applications 25
3.2 Potential applications in the near future 27
3.3 What makes finance special 30
4 General principles for the use of AI by financial institutions (‘SAFEST’) 33

4.1 Soundness34
4.2 Accountability35
4.3 Fairness36
4.4 Ethics37
4.5 Skills37
4.6 Transparency38
References41
Appendix A Relevant articles in the Financial supervision Act

(unofficial translation) 48
Practical information
Artificial Intelligence (AI) is increasingly DNB, therefore, welcomes comments on this 5
being applied in the financial sector and discussion paper. Comments can be sent to
is likely to become more important in [email protected]. Please note that the deadline for the
the years to come. How the adoption submission of comments is 18 October 2019.
of AI will come to shape the financial Comments submitted after this deadline, or
sector precisely, however, is not yet submitted via other means may not be processed.
clear. Nor is it clear what form future
regulatory responses will take. In order All contributions received will be published following
to facilitate a dialogue on this important the close of the consultation, unless you request
topic, De Nederlandsche Bank (DNB) otherwise. If you request that your response is
presents its preliminary view on a set treated as confidential, it will not be published on
of possible principles in this discussion DNB’s website, or shared with any third parties.
paper. We believe that the issues and
ideas outlined in this paper will benefit DNB will use this discussion paper, and the
from a broader discussion. Particularly comments received, to engage in a dialogue with
relevant questions regard the need for, as the Dutch financial sector over the coming months.
well as the scope and form of a potential We will report on the outcome of this process in the
regulatory response. Furthermore, this course of 2020.
paper invites a discussion on what market
participants can already do themselves in
order to ensure compliance with existing
regulations.
Summary
6 Artificial Intelligence (AI), defined as research reports project a 1,000 to 2,000 percent
“the theory and development of increase in AI-related revenue in the years 2018-
computer systems able to perform tasks 2025 and financial services is consistently named as
that traditionally have required human one of the sectors that invests most heavily in AI.
intelligence”, is becoming progressively
more commonplace in finance. This paper Current applications of AI in the financial sector are
provides basic background information manifold and widespread, both in front-end and in
on AI, as well as its current and future back-end business processes. Examples of current
use in the Dutch financial sector. It AI applications are advanced chatbots, identity
concludes with a set of general principles verification in client onboarding, transaction data
for a responsible application of AI in the analysis, fraud detection in claims management,
financial sector. pricing in bond trading, AML monitoring, price
differentiation in car insurance, automated analysis
AI, as a research area, has been around since 1950. of legal documents, customer relation management,
Driven by an exponential increase of computing risk management, portfolio management, trading
performance, AI has continuously evolved over execution and investment operations. Taking
time. The period between 1950 and 2010 can be a forward-looking perspective we can expect
described as 60 years of anticipation, during which further AI-driven innovations in various financial
steady progress was made, but real breakthroughs domains, ranging from leaner and faster operations
– like the 1997 defeat of Garry Kasparov by Deep (“doing the same thing better”) to completely
Blue – were few and infrequent. Since 2010, AI new value propositions (“doing something
seems to have come of age, and milestones have radically different”). A recent report by the World
been achieved almost on a yearly basis. In the last Economic Forum provides a few dozen examples
decade, AIs have bested humans in increasingly of potential AI innovations in deposits & lending,
difficult tasks such as playing complex games – ‘Go’, insurance, payments, investment management,
poker, and real-time strategy games – and driving capital markets and market infrastructure. From
cars. Nowadays AI consists of a blooming and an AI-perspective, finance is special in a number
diverse landscape of cutting-edge research areas, of important ways that warrant an adequate
technologies in various stages of development, and regulatory and supervisory response. First of all,
commercial real-world applications. In the near the financial sector is commonly held to a higher
future, AI is expected to become more advanced and societal standard than many other industries, and
increasingly ubiquitous, driven by the continuous incidents with AI could have serious reputation
increase of computer power, data availability and effects for the financial system. Second, incidents
the advance of the Internet of Things. Various could have a substantial impact on financial stability.
Given the inherent interconnectivity of the financial and systemic risks might arise. Firms should also be 7
system, the rise of AI also has a strong international accountable for their use of AI, as AI applications
dimension. Furthermore, the specific data may not always function as intended and can result
environment of the financial sector is a challenge in damages for the firm itself, its customers and/
for certain AI applications. Last but not least, the or other relevant stakeholders. Although fairness is
progress of AI on the one hand, and its increasing primarily a conduct risk issue, it is vital for society’s
importance in financial sector business models on trust in the financial sector that financial firms’
the other hand, invites us to rethink our traditional AI applications – individually or collectively – do
supervisory paradigms. not inadvertently disadvantage certain groups of
customers. Financial firms should therefore be
Based on the observations in this document, and able to define what their conception of fairness is
building on the previous work by other regulatory and demonstrate how they ensure that their AI
bodies, we have formulated a number of general applications behave accordingly. As AI applications
principles regarding the use of AI in the financial take on tasks that previously required human
sector. The principles are divided over six key intelligence, ethics becomes increasingly important
aspects of responsible use of AI, namely and financial firms should ensure that their
(i) soundness, (ii) accountability, (iii) fairness, (iv) customers, as well as other stakeholders, can trust
ethics, (v) skills, and (vi) transparency (or ‘SAFEST’). that they are not mistreated or harmed – directly
or indirectly – because of the firm’s deployment
From a prudential perspective, soundness is the of AI. When it comes to skills, from the work floor
aspect of AI that is of our primary concern. AI to the board room, people should have a sufficient
applications in the financial sector should be reliable understanding of the strengths and limitations
and accurate, behave predictably, and operate of the AI-enabled systems they work with.
within in the boundaries of applicable rules and Transparency, finally, means that financial firms
regulations. This aspect becomes particularly should be able to explain how and why they use AI
important when financial firms start to apply in their business processes and (where reasonably
identical (or relatively similar) AI-driven solutions appropriate) how these applications function.
8
1 Introduction
Artificial Intelligence (AI) is becoming This document should be seen as a first supervisory 9
more and more commonplace in step into this phenomenon. It aims to provide basic
finance. Financial firms increasingly use background information on AI (chapter 2), as well as
AI to enhance their business processes its current and potential future use in the financial
and improve their product and service sector (chapter 3). Finally, chapter 4 provides a set of
offerings. The advent of AI has often general principles for financial firms that apply AI in
been described as the Fourth Industrial their business processes.
Revolution.1 Whether or not AI will live
up to this claim, at the very least, it
should be considered as an impactful
development that requires an adequate
response from regulators and supervisors.
It should therefore not be a surprise that
an increasing number of supervisors and
international organisations are publishing
their takes on AI and its ramifications for
financial sector supervision.
But what do we mean exactly when we talk about

AI? Definitions vary widely and have changed over
the years. Most financial sector studies by now have
adopted the broad definition put forward by the
Financial Stability Board (FSB) of AI as “the theory
and development of computer systems able to
perform tasks that traditionally have required human
intelligence”.2 Depending on the nature and objective
of future studies, other definitions may be adopted,
but for this paper the FSB definition will suffice.
1 See for instance WEF (2019).

2 FSB (2017, p. 35).
10
2 AI: past, present and future

This chapter provides a brief overview of The first electronic digital computers, developed 11
the history of AI and projections regarding in the late 1930s were little more than gargantuan
its development in the near future, and calculators operating though vacuum tubes. With the
is meant to provide some background replacement of vacuum tubes by transistors in 1947
and context for the subsequent chapters. and the subsequent introduction of the integrated
Paragraph 2.1 describes the period of 1950- circuit (‘microchip’) in 1963, computers soon
2009, during which milestones in AI were developed into smaller and more powerful machines
mainly conceptual in nature, but promises that started to decorate our desktops in the 1980s.
remained largely unfulfilled. Paragraph 2.2 In Figure 1, we can clearly see the exponential growth
details the coming of age of AI from 2010 of computing power since 1945 from the world’s first
to the present day, and paragraph 2.3 supercomputer ‘ENIAC’ (capable of 500 floating point
provides a forward-looking perspective in operations per second – or ‘flops’ – see box 1) to
which AI will steadily become ubiquitous today’s fastest supercomputer ‘Summit’ (capable of
in our everyday lives. a staggering 200 petaflops).4,5 For context, the graph
also shows the development of computing power
of consumer video cards (graphics processing units)
2.1 Past (1950-2009): 60 years of and high-end smartphones.6 Through the second
anticipation half of the 20th century, this exponential growth of
computing power gave rise to evermore advanced
Although AI is a hot topic that seems to have only and capable AI applications.
entered the scene recently, the concept (in its
modern form) has been around for at least seven
decades. Already in 1950, Alan Turing posited that
“machines will eventually compete with men in all
purely intellectual fields”.3
3 Turing (1950, p. 460).

4 Chodosh (2017).
5 See TOP500.org (2019). This record is expected to soon be shattered as AMD and Cray are developing
a new supercomputer with 1.5 exaflops (or: 1,500 petaflops) of processing power for the US Department
of Energy, to be operational in 2021 (Vincent, 2019).
6 It is often mentioned that today’s iPhone has more computing power than NASA’s computers that
were used to put a man on the moon in 1969. To put this in perspective, as can be seen in figure 1 today’s
fastest high-end smartphone has about 100,000 times the computing performance as the fastest
supercomputer available in 1969.
12 Box 1 FLOPS
Floating point operations per second – or ‘FLOPS’ – is the most common metric used to describe
computing power when it comes to AI. A floating point operation is any basic mathematical operation
that involves decimal numbers (such as ‘0.007’; as opposed to integers such as ‘3’). As performance
increases exponentially, typical shorthand notations used are kiloflops or KFLOPS (103 flops), megaflops (106),
gigaflops (109), teraflops (1012), petaflops (1015), exaflops (1018), and zettaflops (1021).
In the 1950s the first (very small scale) artificial algorithms and the first attempt at a chatbot called
neural networks were developed to emulate ‘ELIZA’.13,14 In 1968, the first chess playing program
the behaviour of a few dozen neurons.7 The first was developed that could compete at amateur level,
checkers playing program was also developed and the world’s first autonomous robot ‘Shakey’ was
during this decade.8 In 1956 the first academic demonstrated by the Stanford Research Institute.15,16
conference was held on, what organiser John 1968 was also the year that AI was popularised
McCarthy coined, ‘artificial intelligence’. 9 In 1959 the (albeit in a negative light) in the persona of HAL
same McCarthy co-founded the world’s first AI lab 9000, the murderous AI featuring in Arthur C.
at MIT.10 Clark’s ‘2001: A Space Odyssey’.
During the 1960s academic researchers further In the 1970s a lot of work was devoted to further
developed the conceptual framework for AI, develop the conceptual frameworks for artificial
amongst others by laying down the foundations of neural networks, natural language processing and
a mathematical theory for AI and introducing the visual perception.17 In 1977, the first backgammon
idea of semantic nets. 11,12
The 1960s also saw the light playing program was developed that could
of the first computer-based pattern recognition compete at a professional level and in 1979 Stanford
7 Yadav, et al. (2015).

8 Samuel (1959).
9 The conference is colloquially called the ‘Dartmouth Workshop’ (Sarangi & Sharma, 2019).
10 CSAIL (n.d.).
11 Solomonoff (1960).
12 Quillian (1968).
13 Uhr & Vossler (1961).
14 Weizenbaum (1966).
15 Greenblatt, Eastlake, III & Crocker (1967).
16 Stanford Research Institute (n.d.).
17 Driven by, among others, Linnainmaa’s work on algorithmic probability (1970), Winograd’s SHRDLU
program (1971) and the University of Edinburgh work on ‘Freddy the Robot’ (Artificial Intelligence
Applications Institute, 2005).
Figure 1 Computing performance over time 13
1018
Floating point operations per second (FLOPS)
1015
1012
109
106
103
1
1945 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020
Fastest supercomputers
Fastest consumer GPUs
Fastest high-end smartphones
University students built the self-navigating precursor of today’s machine learning (see
‘Stanford Cart’, which can be considered as the box 2) – was introduced to allow a computer to
ancestor of today’s self-driving cars. 18,19
During create a set of rules based on training data.21 Later
the 1970s expert systems were also developed, that decade, ‘NETtalk’ was the first successful
in which knowledge was added to the system by attempt at using artificial neural networks for
programmers but without any ground-breaking learning natural language pronunciation, and in
results (leading to the first “AI winter” in which 1989 Tim Berners Lee invented the World Wide
progress slowed down and funding decreased).20 Web (which would become instrumental in the
In the 1980s Explanation Based Learning – the development of AI in the 2000s).22,23
18 Berliner (1977).
19 Kubota (2019).
20 Examples of such expert systems were Shortliffe’s MYCIN program (1977) and so-called LISP-machines
(Bawden, et al., 1977).
21 Milton, et al. (1989).
22 Rosenberg & Sejnowski (1986).
23 World Wide Web Foundation (n.d.).
14 Box 2 Machine Learning
Machine learning comes in various sizes and shapes. The most important categories are listed below.
Supervised Learning
Supervised learning algorithms construct models for predicting outcomes based on past observations.
Once the model is trained by being fed example input-output data, it can take inputs only and make output
predictions for the purpose of classification or regression. Supervised ML is far more commonly used than
unsupervised ML. Using transaction history and patterns to detect fraud, or using loan data and history for
credit scoring are examples of supervised learning.
Unsupervised Learning
In unsupervised learning, the model finds patterns based only on input data. Most unsupervised learning
implements cluster analysis techniques that split data items into groups, based on the similarity of their
features. A common example of unsupervised learning is anomaly detection (eg. transaction monitoring for
AML). Another common example is market research, in which unsupervised learning is used to partition the
general population of consumers into market segments based on data such as region, spending patterns, or
social media activity.
Reinforcement learning
Reinforcement learning refers to goal-oriented algorithms, which learn how to attain a complex objective
(goal) or maximise along a particular dimension over many steps; for example, maximise the points won in
a game over many moves. Such an AI can start from a blank slate, and under the right conditions, like with
DeepMind’s ‘AlphaGo’ (see §2.2), can achieve superhuman performance.
Deep Learning
When there are many hidden layers in an artificial neural network, mimicking the working of a human
brain, it is considered a deep learning model. With more layers comes the ability to solve more complex
problems, albeit at more computational expense. Deep learning is able to achieve extremely high levels of
accuracy, but is also at the root of the “black box” problem in which the human who constructed the model
can no longer really understand what the model is doing or how it is making its decisions.
Following the continuing disenchantment with The 1990s was also the decade in which AIs started
knowledge-based systems, resulting in another to outperform humans in certain areas. In 1994, the
“AI winter” in the late 1980s, the focus shifted to reigning checkers world champion Marion Tinsley
data-driven machine learning and reinforcement resigned in a match against the computer program
learning in the 1990s. From there, the first ‘Chinook’.25 Only three years later, in 1997, the world
companies started to integrate machine learning in witnessed the defeat of chess grandmaster Garry
their business processes (like Mellon Bank, which Kasparov by IBM’s ‘Deep Blue’.26 The 1990s also
used machine learning to detect credit card fraud).24 welcomed Google and Amazon on the digital stage,
24 OECD (1997).
25 Madrigal (2017).
26 Weber (1997).
which would become two of the most prominent of data to feed data-hungry AI systems. As with 15
drivers of today’s AI developments. computing performance, the annual data production
continues to grow at an exponential rate.
Driven by increasingly powerful computer systems,
most notably in the form of graphics processing units As can be seen in Figure 2, in 2019 we will create
(GPU), AI finally started to get real traction in the more than 40 zettabytes of data.29,30 To put this
2000s. This progress coincided with the development in perspective: if we would write all this data to
of deep learning as a subfield of machine learning.27 CD-ROMs, and stack them up neatly in a pile, that
The introduction of Sony’s PlayStation 3 (PS3) in 2006 pile would be 80,000,000 kilometres high. That is
also provided a boost for AI research, as the PS3 about 1.5 times the distance from Earth to Mars.
design and programming allowed researchers to link
multiple of these gaming systems together, enabling The first important milestone for this decade
them to build their own low-cost supercomputers. 28
came in 2010 when Microsoft launched Kinect
A year later, Apple introduced the iPhone, setting the for Xbox 360: an array of sensors that could
stage for mobile internet. Finally, after 60 years of track human body movement accurately in three
anticipation, the promise of AI would start to come dimensions and in real-time.31 A year later in 2011,
to fruition in the 2010s. IBM’s ‘Watson’ defeated two top contestants in
the popular TV show ‘Jeopardy!’.32 That same year,
Apple launched ‘Siri’, the first widely available
2.2 Present (2010-2019): The coming conversational AI. This was later followed by
of age of AI ‘Google Now’ in 2012 and Amazon’s ‘Alexa’ and
Microsoft’s ‘Cortana’ in 2014. In 2012, Google built
Since 2010, milestones in the world of AI have a neural network of 16,000 computer processors
been achieved almost on a yearly basis. Besides that successfully taught itself to identify cats in
the ever increasing computer power available to YouTube videos using deep learning.33 That same
AI researchers, the rise of the World Wide Web in year, ‘AlexNet’ – a convolutional neural network
the 1990s and 2000s has created massive amounts – made headlines at the ImageNet Large Scale
27 Driven by, among others, the work of Geoffrey Hinton (deeplearning.ai, 2018).
28 Moss (2018).
29 A zettabyte is 1021 bytes.
30 Reinsel, Gantz, & Rydning (2018).
31 Not only was the Kinect’s software developed utilizing machine learning (Microsoft, 2011), the device
also facilitated AI research in human motion recognition and lower-level computer vision tasks such as
segmentation and object detection (Lun & Zhao, 2015).
32 Markoff (2011).
33 Le, et al. (2012).
16
Figure 2 Annually created data world-wide 17

Zettabytes (1021 bytes)
50
40
30
20
10
0
2000 2002 2004 2006 2008 2010 2012 2014 2016 2018 2020
Visual Recognition Challenge, by defeating the the use of Google’s Tensor Processing Units, which
competition by a large margin. 34,35
are specifically developed for AI applications.
Subsequently, its successor ‘AlphaZero’ successfully
In 2016, another leap in AI research was achieved applied transfer learning to excel in two other
when Google DeepMind’s ‘AlphaGo’ spectacularly board games: it became the world’s best chess
defeated world champion Lee Sedol in a 5-round player in only 4 hours and the world’s best Shogi
game of ‘Go’. 36 Go is known as a game that is player in a mere 2 hours.38 Also in 2017, the poker-
significantly more complex and difficult to master playing AI called ‘Libratus’ managed to defeat top
than chess. A year later in 2017, DeepMind’s professional players in heads-up games of Texas
‘AlphaGo Zero’ unceremoniously defeated its Hold’em poker.39 Liberatus became the first AI to
predecessor ‘AlphaGo’ after only 24 hours of learning achieve a superhuman level of proficiency in a game
the game by playing against itself; the score was of imperfect information (with games like chess
100-0. This accomplishment was facilitated by
37
and ‘Go’, the information available to the players is
34 A convolutional neural network is a class of deep neural networks.

35 Krizhevsky, Sutskever, & Hinton (2012).
36 Mundy (2016).
37 Silver, et al. (2017).
38 Silver, et al. (2018).
39 Brown N. & Sandholm (2018).
18 identical or ‘perfect’, but with a game like poker a The last decade also saw the introduction of relatively
player does not have all the information as he does easy to use AI-specific software packages and low-
not know the other player’s cards). In 2019, Google’s cost cloud-based solutions that enabled researchers,
DeepMind achieved another milestone when its corporations and even SMEs to develop their own
‘AlphaStar’ defeated a pro player in ‘StarCraft II’. 40 AI-driven applications.44 This proliferation of AI
This computer game has long been considered one capabilities, however, also brought to light certain
of the most challenging real-time strategy games risks associated with AI. As Bill Gates once said: “it is
(as opposed to computer games where players take fine to celebrate success but it is more important to
turns to plan and execute their actions). StarCraft heed the lessons of failure”. Box 3 below therefore
II was considered an exceptionally difficult game expounds some of AI’s cautionary tales.
to master for an AI: it is an imperfect information
game, that requires the balancing of short and Today the world of AI consists of a blooming and
long-term goals, managing both the macro diverse landscape of cutting-edge research areas,
(economy) and micro (individual unit control) level, technologies in various stages of development, and
and the need to continuously adapt to unexpected commercial real-world applications. Figure 3 below
situations; all in real-time. presents an overview of various AI technologies
and research areas. An in-depth discussion of all
In May 2019, Tesla presented its ‘Full Self-Driving’ technologies and research areas mentioned falls
package for autonomous driving.41 In 2017, Tesla outside the scope of this paper, but the overview
already enabled the ‘Enhanced Auto Pilot’ mode does show that the field of AI and its potential
via a firmware update, making the Tesla the future applications in the financial sector entail a lot
first commercially available (semi) self-driving more than ‘just’ predictive analytics and machine
car.42 Google’s Waymo, Uber, and traditional car learning. The next paragraph will provide a few
manufacturers like Mercedes-Benz and General reflections as to what we may expect regarding the
Motors are also working hard to bring their versions development of AI in the coming decade.
of a self-driving car to market. 43
40 DeepMind (2019).
41 Rapier & Bastone (2019)
42 Brown, M. (2017).
43 Welch & Behrmann (2018).
44 Most notably the AI-related services available on Microsoft Azure Machine Learning Studio, Google
Cloud Machine Learning, and Amazon Web Services SageMaker.
Box 3 When AI gets it wrong: errors, accidents and disasters

American Economist Leo Cherne is attributed with saying that “The computer is incredibly fast, accurate, 19
and stupid”. Despite the reference to ‘intelligence’ in AI, this holds true for today’s AI as well. Current AI
applications are exceptionally strong at performing narrow, specific tasks. However, they lack the general
intelligence to reflect on their performance, and typically fail to catch errors that are plain obvious to
humans. Some of these errors are merely funny, others can be painful, damaging, or even catastrophic.
One reason AI fails, is because it has not been properly trained for all eventualities. In November 2018, an AI
system in the southern port city of Ningbo, China, falsely classified a photo of Chinese billionaire Mingzhu
Dong on an ad on a passing bus as a jaywalker. This embarrassing mistake went viral on Chinese social
media and Ningbo police were quick to apologise.45 In June 2016, a Tesla driver was killed while the car was
in auto pilot mode, when it crashed into a truck. After careful examination of the logs, Tesla stated that
“neither Autopilot nor the driver noticed the white side of the tractor trailer against a brightly lit sky, so the
brake was not applied”.46
Another reason for AI failure is poor training data. A 2016 article from the University of Washington
demonstrates how an artificial neural network that was trained to distinguish between wolves and
huskies failed as a result of poor training data. Because all of the training images of wolves had snow in the
background, and those of the huskies did not, the AI did not distinguish between features of the animals
themselves but simply used the background as the discriminating feature.47 Such accidents, however, are
not confined to academic experiments. In 2015, Google got embarrassingly called out when its Photos
app labelled black people as “gorillas”.48 Just a month before this incident, Yahoo’s Flickr app misclassified
pictures of a concentration camp as “jungle gym” and “sport”.49
Sometimes AIs get it wrong, simply because they adopt societal biases. In 2018, Amazon had to scrap its
AI-enabled recruiting software because it was biased against women. This bias was a direct reflection of the
male dominance in the tech industry, and even after several attempts to exclude gender as a feature, the
AI kept finding proxies for gender to discriminate on.50 Another well-known example is how popular online
translation engines struggle to be gender-neutral. For example, if you would translate “she is a doctor”
from English to Turkish, the result is “o bir doctor”. Translate it back, however, and because of gender bias in
English literature translation engines typically return “he is a doctor”.51
AIs are also prone to be corrupted by malicious intent. In 2016, The Verge headlined “Twitter taught
Microsoft’s AI chatbot to be a racist a**hole in less than a day”. Microsoft’s Twitter bot ‘Tay’ was caught
twittering very explicit racist and hateful messages within 24 hours after it went live.52
People, of course, are fallible and prone to failure too. However, when an AI fails, it fails diligently, at scale,
and often in unexpected ways. This in particular is the reason, at least for AI applications in finance, why it is
so important to get things right from the start.
45 Dodds (2018).
46 Boudette (2017).
47 Ribeiro, Singh & Guestrin (2016).
48 Grush (2015).
49 Hern (2015).
50 Dastin (2018).
51 Sonnad (2017).
52 James (2016).
20 Figure 3 Overview of AI technologies and research areas
Artificial
Artificial general
general intelligence
intelligence Intelligent
Intelligent agent
agent Robotics
Robotics
Transferlearning
Transfer learning Agent
Agent architecture
architecture Behavior-based
Behavior-based robotics
robotics
Cloudrobotics
Cloud robotics
Automation Knowledge management Cybernetics
Automation Knowledge management Cybernetics
Automated reasoning Concept mining Developmental robotics
Automated
Automaticplanning and scheduling
target recognition Concept
E-mailmining
spam filtering Developmental robotics
Epigenetic robotics
Automated
Automated planning and scheduling
reasoning Information
E-mail extraction
spam filtering Epigenetic robotics
Evolutionary robotics
Automatic target recognition Information
Information extraction
extraction Evolutionary robotics
Biologically inspired computing Named-entity
Information extraction
extraction
Speech generating device
Evolutionary computation Knowledge representation Natural language generation
Biologically inspired computing Named-entity extraction Speech generating device
Neuromorphic hardware Semantic web
Evolutionary computation Knowledge representation Natural
Various language generation
Computer audition
Neuromorphic hardware Machine
Semantic weblearning Affective computing
Speech recognition Constrained Conditional Models Argument technology
Various
Speaker recognition
Computer audition Supervised
Machine learning
learning Artificialcomputing
Affective creativity
Unsupervised learning Artificial life
Speech recognition
Computer vision Constrained Conditional Models Argument technology
Reinforcement learning AI diagnosis
Speaker recognition
Image processing Supervised learning
Deep learning Artificial creativity
Hybrid intelligent system
Intelligent word recognition Unsupervised learning fields
Neural modeling Artificial life control
Intelligent
Object recognition
Computer vision Reinforcement learning Nonlinear
AI diagnosis control
Optical mark recognition Natural language processing Strategic planning
Image processing Deep learning Hybrid intelligent system
Optical character recognition Chatbots Vehicle infrastructure integration
Intelligent word recognition Neural modeling fields Intelligent control
Facial recognition systems Language identification Virtual Intelligence
Object recognition Natural language user interface Nonlinear control
Expertmark
Optical system
recognition Natural
Naturallanguage processing
language understanding Virtual reality
Strategic planning
Decision
Optical support
character system
recognition Machine translation
Chatbots Augmented
Vehicle reality integration
infrastructure
Facial recognition systems Question
Language answering
identification Virtual Intelligence
Game artificial intelligence Semantic translation
Computer game bot Natural language user interface
Expert
Video system
game AI Natural language
Pattern understanding
recognition Virtual reality
Generalsupport
Decision game playing
system Optical
Machine character recognition
translation Augmented reality
Game theory Handwriting
Question recognition
answering
Game artificial intelligence Speechtranslation
Semantic recognition
Face recognition
Computer game bot
Video game AI Pattern recognition
General game playing Optical character recognition
Game theory Handwriting recognition
Speech recognition
Face recognition
2.3 Future (2020-beyond): further discussion regarding the future development 21

Ubiquitous AI of computing performance).
As with the past 70 years, the main driver for What is perhaps even more thought-provoking
future developments will be the further increase is that, in terms of raw computational power, we
of computing power. Figure 4 shows the projected are likely to achieve (and then quickly surpass) the
development of computing performance computational power equivalent of the human
from 2020-2035. Compared to the fastest brain in the coming years. Although our brains do
supercomputers, GPUs in home PCs have a lag of not perform floating point operations as computers
about 12 years in terms of computing performance. do, researchers estimate that the human brain has
Subsequently, mobile phones lag about 9 years a performance equivalence in the exa-zettaflops
behind home PCs. In other words, if the trend of range.53 Now, achieving human brain equivalent
the last few decades continues, the performance of computing power, in and of itself, does not mean
today’s fastest supercomputer will sit on your desk that we will suddenly have created a sentient entity
at home by 2030 and will fit comfortably in your with human-like general intelligence. The human
jeans’ pocket by the year 2039 (see box 4 for some brain is a very complex, and highly efficient piece
Figure 4 Projected computer performance

1024
Floating point operations per second (FLOPS)
1021
Estimated computing performance
required to emulate a human brain
1018
1015
1012
109
1990 1995 2000 2005 2010 2015 2020 2025 2030 2035
Fastest supercomputers
Fastest consumer GPUs
Fastest high-end smartphones
53 Bouchard, et al. (2018), Sanberg & Bostrom (2008).

22 of biological hardware, and neuroscience still has a driving finance’ in which AI will first augment, and
long way to go in fully understanding its workings. eventually autonomously manage, our finances.56
It does raise the question what kind of feats AI
applications may accomplish in the next decade, From an industry-perspective, former Google and
given that current AIs are already able to beat Baidu scientist Andrew Ng describes AI as the new
humans in increasingly complex cognitive tasks. electricity: “Just as electricity transformed almost
everything 100 years ago, today I actually have a
Another, perhaps more relevant, way of looking at hard time thinking of an industry that I don’t think
all this is the implication that currently available AI will transform in the next several years”.57 Ng
feats of AI will come in increasingly smaller regards financial technology as one of the sectors
packages. This means that AI applications that best suited for an AI-led transformation, and states
presently need a lot of expensive hardware or cloud- that the only factors that could slow down this
services to operate will become available to desktop development are the scarcity of talent, and the
environments and, eventually, in tiny devices, as challenge for companies to effectively harness
time progresses. The ‘Internet of Things’ is a direct their data.
consequence of this trend, which will gradually bring
AI into all aspects of our daily lives, quietly operating Various research reports project a 1,000 to 2,000
in the background and creating ever more data percent increase in AI-related revenue in the years
in the process.54 Mark Weiser, Chief Technologist 2018-2025.58 Financial services is consistently named
at Xerox PARC in the 1990s, said that “the most as one of the sectors that invests most heavily in AI,
profound technologies are those that disappear. and in AI-related big data and business analytics,
They weave themselves into the fabric of everyday banking is even named as the largest revenue-
life until they are indistinguishable from it”.55 This generating sector.59,60 The next chapter provides a
quote aptly captures what we may expect from AI more in-depth overview of current and potential
in the coming years from a consumer-perspective. future AI applications in the financial sector.
Following the self-driving car, we will likely see
the advent of other automated services, like ‘self-
54 Brunkhard (2018).
55 Weiser (1991, p. 94).
56 WEF (2018).
57 Lynch (2017).
58 Tractica (2019), Orbis Research (2017), Grand View Research (2017).
59 Statista (2016), Ibid. (2015).
60 Ibid. (2019).
23
Box 4 The end of Moore’s Law?
the promise of exponential growth of computing performance was more or less formalised with “Moore’s
Law” (named after Intel’s co-founder Gordon Moore) in 1965. Moore projected that the number of
transistors in a dense integrated circuit would continue to double about every two years. As we have seen
in paragraph 2.1, the exponential increase of computing performance has kept a steady pace during the last
70 years.
Moore’s Law, however, is not a law of nature, but rather a prediction about future technological
developments that has been a guiding principle for major chip-developers like Intel, NVIDIA and AMD. This
means that historical growth-rates do not guarantee that these rates will be sustainable in the future. In
the last few years chip-developers have struggled to make the step from a 10 to a 7 nanometre lithography
process in the production of computer chips, and in 2015, Intel admitted that it would no longer be able to
keep up with Moore’s Law. Instead of doubling transistors every 24 months, Intel said, this process would
advance every 30 months from then on.61 This prompted many commenters to proclaim the imminent end
of Moore’s Law. Although the challenge of creating ever-smaller transistors undeniably becomes harder,
there are good reasons to expect that computing performance will continue to grow exponentially in
decades to come.
First of all, computer performance has increased at a consistently higher pace than the growth rate that
characterises Moore’s Law, doubling every 18 months instead of every 24 months. This is because Moore’s
Law is about the number of transistors only, and does not take into account innovations that increase the
performance of the chips (such as higher clock speeds, multi-processors and multi-threading). In 2018,
NVIDIA CEO Jensen Huang underlined that the number of transistors is of diminishing importance and
that, if anything, performance increase is speeding up.62 Furthermore, technologies and methods are being
developed to circumvent the challenges of even lower nanoscale lithography processes on silicon, for
example by moving from a single layer (2D architecture) to multilayer (3D architecture) computer chips, the
adoption of carbon nanotubes technology to replace silicon, quantum computing, and bio-computing.63,
64, 65 ,66
Finally, even if performance increases would slow down at some point, this would not change the
exponential nature of this increase, but would only mean that new milestones, such as reaching zettaflops-
performance, will be achieved at a somewhat slower (but still exponential) rate.
61 Oreskovic (2016).
62 Perry (2018).
63 Giles (2018).
64 Malecek (2016).
65 Bachor & Simmons (2017).
66 Nicolau, et al. (2016).
24
3 AI in finance
This chapter will provide some insight into are material to customers (such as credit scoring, 25
how AI is currently already being used in advice, and underwriting). Areas of applications
finance, and what potential applications provided by other reports are chatbots, claim
we might expect in the near future. management, personalised pricing, financial
Paragraph 3.1 presents an overview of and operational risk management, compliance,
current use cases in the financial sector, anti-money laundering, cyber security, portfolio
paragraph 3.2 describes potential future management, and internal risk modelling.69
applications, and paragraph 3.3 analyses
specific AI-related issues from a financial A number of real-world cases, noting the use of AI
sector perspective. in both front-end and back-end business processes
of financial institutions, are detailed below. Note
that these examples are not exhaustive, their
3.1 Current applications inclusion neither implies that DNB endorses these
applications, nor have we verified that these
AI adoption in the financial sector already is quite examples adhere to the principles set out in this
widespread, even though we might not always paper. These cases merely intend to illustrate the
be aware of its presence. In their report “Artificial manifold contemporary applications of AI in finance.
Intelligence and machine learning in financial
services” the FSB distinguishes four different areas Customer interaction
for AI application: customer-focused applications, An increasing number of financial service providers
operations-focused uses, trading and portfolio employ AI-enabled chatbots that use natural
management, and regulatory compliance and language processing techniques to provide a
supervision.67 The French Autorité de Contrôle 24/7 first line of customer interaction. Intelligent
Prudentiel et de Résolution (ACPR) has observed that chatbots can save costs dramatically as they
the implementation of AI seems more advanced in reduce the need for customer service call centres.
the banking sector than in the insurance sector. 68
The Dutch insurance company ASR’s chatbot-
As stated in their report, most applications currently based website is a typical example of an intelligent
in production deal with operational processes chatbot.70 Increasingly, companies opt to integrate
(eg. using NLP to automate the processing of their chatbots into existing platforms such as
written documents, fraud detection), and so far Facebook Messenger or WhatsApp, as well as
AI is not yet commonly used for activities that
67 FSB (2017).
68 ACPR (2018).
69 FSB (2017), BaFin (2018), Monetary Authority of Singapore (2018), ACPR (2018), CSSF (2018).
70 ASR (n.d.).
26 speech-enabled platforms such as Google Assistant, Fraud detection
Apple Siri and Amazon Alexa. Dutch fintech company FRISS is an AI-enabled
service provider that is used by various Dutch
Client oboarding insurance companies for fraud detection in claims
Nowadays, when we want to open a new bank management.73 Their system uses expert knowledge,
account, we are likely to encounter AI during the proprietary data, and external data sources to
onboarding process. Rabobank, for instance, uses provide real time risk scores. FRISS also offers AI-
UK-based iProov’s deep-learning driven ‘Face driven risk assessment at underwriting and CDD
Verifier’ to match a selfie with the face image from solutions for onboarding new clients.
the customer’s identity document, including liveness
check. Rabobank uses this service in conjunction Investor services
with ‘ReadID’ to verify the ID document’s ING’s ‘Katana’ tool helps bond traders with their
authenticity, which also uses AI techniques such as pricing decisions. Katana employs predictive
optical character recognition.71 analytics, based on historic and real-time data,
to help traders decide what price to quote when
Account information services buying and selling bonds for their clients.74 This
AI is also commonly used in account information tool is a typical example of how AI can be used for
services, such as ABN AMRO’s ‘Grip’ app. Grip uses
72
augmentation (rather than for automation) as it
machine learning to automatically classify incoming analyses and visualises relevant information for the
and outgoing transactions and assigns them to user to decide on.
categories that are relevant to the app’s user.
The app also provides insight into possible saving Transaction monitoring
opportunities as well as more general budgeting BusinessForensics’ HQ Craft is an example of an
services. With the recent implementation of PSD2, AI-driven solution for real-time AML transaction
more such apps by third-party service providers are monitoring. This big data processing platform uses
expected to be introduced. data analysis of network transactions, forensic
analysis, fraud management and machine learning.
This Netherlands-based company provides its
services to various Dutch banks and insurance
companies.75
71 Wegdam (2018).
72 ABN AMRO (n.d.).
73 FRISS (n.d.).
74 Ibid. (2017).
75 BusinessForensics (n.d.).
Car insurance Various Dutch financial firms are among Salesforce’s 27

Fairzekering is a Dutch AI-enabled car insurance customers. 78
product. The insurance company analyses their

customers’ driving behaviour based on the data Portfolio management and risk analytics
collected by a smart module ‘Chipin’ that is installed BlackRock’s ‘Aladdin’ is an example of a financial
in their customers’ cars. Depending on your risk platform that increasingly applies AI.79 Aladdin is a
score, you get a 0-35% discount on your monthly widely used operating system that combines AI with
rate. Customers also have access to a dashboard sophisticated risk analytics, comprehensive portfolio
that provides them with additional insights in their management, trading execution and investment
driving behaviour.76 operations tools on a single platform. Today, Aladdin
is used to manage an estimated 18 trillion US dollars
Document analysis in assets globally.80 BlackRock’s platform is used by
JPMorgan Chase developed an AI-driven Contract numerous Dutch financial institutions, including
Intelligence platform called ‘COiN’ to analyse legal major banks and pension funds.
documents and extract important data points and
clauses. According to JPMorgan, the system was As the examples above show, the use of AI is already
used to review 12,000 annual commercial credit widespread in the financial sector, and we can
agreements, which would normally take 360,000 expect that AI will be used increasingly to augment
hours to review manually.77 A number of Dutch or automate existing business processes as well as
banks and insurance companies have indicated that to enable new financial products and services.
they are also implementing automated solutions to
analyse documents.
3.2 Potential applications in the near
Customer relationship management future
Salesforce is an AI-enabled CRM platform that offers
companies a wide variety of tools to support their Various reports provide a glimpse of what AI has
client interaction. One of their products, ‘Einstein’, to offer the financial sector.81 AI applications can
uses AI to provide lead scoring, activity capture, for instance be used to structure relatively complex
opportunity insights, and account insights services. middle and back office settlement processes
76 Fairzekering (n.d.).
77 JPMorgan Chase & Co (2017).
78 Salesforce (n.d.).
79 Kochansky (2018).
80 Massa (2018).
81 FSB (2017), BaFin (2018), Monetary Authority of Singapore (2018), ACPR (2018), CSSF (2018).
28 more efficiently or even automate them. AI can channels, (v) predict defaults with greater accuracy,
furthermore improve the customer experience (vi) automate and augment business credit
and compliance processes (especially in the decision-making, and (vii) improve client advisory
prevention of money laundering). It can increase by integrating into data streams for opportunity
the precision of risk assessment models and discovery.
generally bring efficiency and effectiveness gains
to financial firms’ core processes. AI can be used to Insurance
monetise customer data, resulting in new income Regarding insurance the WEF foresees AI
sources for financial firms. It can simplify and applications to (i) improve underwriting, pricing
speed up interactions with consumers and result efficiency and accuracy, (ii) increase capital
in faster processes in trading, sales and product efficiency through better risk modelling and real-
development. Finally, the application of AI may lead time risk monitoring, (iii) triage and grade claims
to highly individualised products and services. Last to increase adjudicator efficiency, (iv) increase the
year, the World Economic Forum (WEF) published capabilities of sales agents and advisors, (v) improve
a comprehensive report titled “The New Physics scale efficiencies to enter new markets, (vi) develop
of Financial Services”, which provides an insightful modern, mobile-first insurance offerings, (vii) reduce
forward-looking perspective.82 The report gives fraud using new tools and data, (viii) process
examples in various financial domains, ranging claims instantly, (ix) advise clients on prevention
from leaner and faster operations (“doing the same strategies to lower their risk profiles, (x) provide
thing better”), to completely new value propositions predictive analytics to clients that help them better
(“doing something radically different”). Although understand their risks, (xi) introduce new pricing
it is hard to predict how AI will exactly shape the and payment models, (xii) develop modularised
financial sector landscape in years to come, the policies, and (xiii) use proxy data to insure new risk
examples provided by the WEF report illustrate the categories.
wide variety of potential innovations nicely.
Payments
Deposits and lending The WEF report projects that the payments industry
According to the WEF report, AI can be expected will apply AI to (i) further automate compliance
to be increasingly used in deposits and lending and reporting, (ii) drive loyalty by offering bespoke
to (i) provide just-in time lending, (ii) miniaturise incentives and rewards, (iii) compete to become
unsecured lending to be use-specific, (iii) provide a provider of invisible payments infrastructure,
personalised practicable advice in real time, (iv) increase detection precision to eliminate
(iv) offer tailored ‘always-on’ experiences across false positives, (v) deploy real-time surveillance
82 WEF (2018).
capabilities, (vi) create an advisory capability for post-trade risk-management solutions, and (vi) use 29
macroeconomic trends, (vii) offer ‘prediction as a broader and better data to develop predictive risk
service’ to merchants, and (viii) act as the ultimate models that drive capital savings.
personal shopper for customers.
Market infrastructure
Investment management For the sixth and final sector, market infrastructure,
AI can be deployed in investment services in order the WEF expects AI to be used to (i) automate alert
to (i) offer seamless account setup and customer triage, investigation and reporting, (ii) automate
acquisition, (ii) control back-office compliance reconciliation and incident reporting, (iii) integrate
costs, (iii) digitise customer servicing, (iv) equip post-trade workflows to achieve straight-through
advisers with highly personalised insights, processing, (iv) improve trade speed and price using
(v) holistically understand investor preferences dynamic execution methods, (v) deploy new order
in real-time, (vi) offer outcome-based portfolio types to protect investors from risks proactively,
modelling, (vii) use existing but unused platforms (vi) deploy holistic market-surveillance services,
for distribution, (viii) mimic advanced strategies (vii) develop macroeconomic indicators using
for low costs, (ix) establish passive products that internal data, (viii) offer insights on market structure
track new datasets, (x) analyse vast quantities of and risk, and (ix) sell internal ‘analytics capabilities
data at scale, (xi) share more detailed economic as a service’.
insights, (xii) continuously source new and exclusive
datasets, (xiii) develop unique strategies and new It is hard, if not impossible, to predict whether and
investment products, (xiv) find new and unique how these global trends described by the WEF and
correlations between datasets, and (xv) enable users other reports will materialise in Europe and in the
to effectively manage their investments. Netherlands. To a large extent, this will also depend
on future regulatory responses.83 The new Payment
Capital markets Service Directive (PSD2), for instance, is expected
Applications in capital markets are foreseen to enable a number of AI-driven innovations, most
by the WEF to (i) automate pre-trade analysis, notably in deposits, lending and payments. The
(ii) automate investment monitoring and reporting, recently implemented General Data Protection
(iii) use predictive models to improve deal Regulation (GDPR), on the other hand, sets explicit
identification, pairing and sales activities, (iv) achieve and implicit constraints on potential innovations by
better investments performance by using new data regulating the manner in which companies may use
in opaque markets, (v) develop real-time pre- and personal data in their product and service offerings.
83 The European Commission is currently looking into AI, and the Commission’s High Level Expert Group
is expected to come up with policy proposals later in 2019.
30 Within Europe, the Dutch financial market seems a strictly legalistic perspective can result in material
to be well-suited to become a front-runner in the reputational damage if it does not correspond to the
adoption of AI. Amsterdam and Eindhoven are expectations of society at large.
positioning themselves as upcoming AI hubs and
ING has recently launched a new AI Fintech Lab in Another aspect that makes finance special is the
collaboration with Delft University of Technology.84, 85 potential impact of AI on financial stability. The
financial system is not impervious to shocks. As
experienced in the global financial crisis, problems
3.3 What makes finance special can propagate through the system rapidly. When
AI increasingly manages financial transactions,
The challenges provided by an increasing adoption relatively small incidents could have major
of AI are not exclusive to the financial sector, and consequences. Given the increasing importance
important questions regarding ethics, privacy of tech giants in providing AI-related services and
and security are relevant for financial and non- infrastructure, the concept of systemic importance
financial firms alike. Still, from an AI-perspective may also need to be extended to include these
finance differs from other industries in a number companies at some point.
of important ways that warrant an adequate
regulatory and supervisory response. Given the inherent interconnectivity of the financial
system, the rise of AI has a strong international
One such difference is that the financial sector dimension.86 An adequate policy response will
is commonly held to a higher societal standard require close international cooperation and clear
than many other industries, as trust in financial minimum standards and guidelines for the sector
institutions is considered essential for an effective to adhere to. Regulatory arbitrage in the area of AI
financial system. Furthermore, many consumers could have dangerous consequences and should be
regard financial institutions like banks, insurance prevented where possible.
companies, and pension funds to be providers of
a utility function. The data that these companies Another way in which finance can be different from
extract from their customers are therefore other industries is that for many financial services
regarded as a by-product of this function, instead we have a very specific data environment. Because
of a commercial asset for the financial sector to of cultural and legal differences data are, in many
monetise. Even data usage that is acceptable from cases, only representative for domestic markets.
84 ING (2019).
85 Municipality of Amsterdam (2018)
86 FSB (2017).
Good examples of this are data pertaining to Finally, the increasing importance of AI for the 31
credit, pensions and mortgages. This may provide financial sector also invites us to rethink our
a challenge for the development of data-hungry traditional supervisory paradigms. When it comes
AI systems, especially for relatively small markets to AI, strong operational controls could be more
as that of the Netherlands. Furthermore, because important than capital and liquidity buffers. We will
of continuous changes to the financial regulatory increasingly come across business models that are
framework, historic data quickly becomes less driven primarily by the monetization of data, rather
representative and unusable for training than by product fees and interest margins.
AI-enabled systems.
The progress of AI in beating humans in increasingly

difficult games should also give us pause. In its
design, the stock market is set up as a perfect
information game, in which all participants have
access to the same information.87 Banking and
insurance can arguably be regarded as imperfect
information games: customers have information
regarding their financial position and risk profile
that is unavailable to banks and insurance
companies.88 As we have seen in the previous
chapter, AIs like AlphaGo and Liberatus are making
huge strides in beating humans in both perfect and
imperfect information games. Now that computer
performance is progressing to the point where
it equals the raw performance capabilities of the
human brain, we should consider the possibility that
at some point – perhaps sooner than we think – AIs
could surpass human agents in trading, banking and
insurance as well.89
87 This property of the stock market is commonly referred to as the efficient-market hypothesis.
88 This is also referred to as the ‘lemon problem’ (Akerlof, 1970).
89 Algorithmic trading already accounts for about 70% of total orders in some markets. These systems,
however, do not currently outperform human traders in the quality of their trade decisions per se, but
primarily in response and execution times.
32
4 General principles for

the use of AI by financial
institutions (‘SAFEST’)
As the previous chapters illustrate, AI of the direction and desirability of future regulatory 33
is a ubiquitous phenomenon. On the developments. It goes without saying that the
one hand, this implies that many of the formulated principles do not preclude that other,
prerequisites for a responsible application more stringent, regulations and/or regulatory
of AI are already covered by existing practices may already apply to specific regulated
regulations. In particular the requirements activities; in these cases the most stringent
regarding sound and controlled business regulations take precedence.
operations (see Financial Supervision
Act articles 3:10, 3:17 and 3:18 for legal The principles in this chapter should be seen in
context) cover many aspects.90 On the context of controlled and sound business
the other hand it is challenging to get operations. Proportionality applies to these
a comprehensive overview of the AI- principles, and their applicability should be
specific challenges faced by financial firms. considered in light of the scale, complexity and
Moreover, loopholes in regulation may materiality of an organisation’s AI applications. The
exist and societal expectations may go applicability of these principles is also determined
beyond what current regulations require. by the role of an AI application in the organisation’s
Consequently, various legislators and
supervisory authorities, including the Figure 5 Applicability of principles in
European Commission and the European light of materiality and purpose of
Banking Authority, are considering the the AI
development of AI-specific policies and/or
regulations.91 We cannot predict, however, automation integral
how future regulatory frameworks
regarding the application of AI in the prescriptive
financial sector will evolve, or the time
Purpose
it will take for these regulatory responses predictive

to take shape.
diagnostic
The principles presented in this chapter aim to
provide a framework that helps financial firms to descriptive limited
assess how responsible their application of AI is and

also provide a way to facilitate our own assessment
low materiality for business high
continuity / customers
90 Appendix ‘A’ provides an unofficial translation for the relevant articles.

91 European Commission (n.d.), High-Level Expert Group on Artificial Intelligence (2019), EBA (2019).
34 decision making process; this means whether 1) Ensure general compliance with regulatory
the AI application serves a descriptive, diagnostic, obligations regarding AI applications.
predictive, prescriptive, or automation purpose. We 2) Mitigate financial (and other relevant
can visualise this proportionality in a ‘heat map’, as prudential) risks in the development and use
can be seen in Figure 5. of AI applications.
3) Pay special attention to the mitigation of
The principles are divided over six key aspects of model risk for material AI applications.
responsible use of AI, namely soundness (§4.1), 4) Safeguard and improve the quality of data
accountability (§4.2), fairness (§4.3), ethics (§4.4), used by AI applications.
skills (§4.5), and transparency (§4.6).92 For each 5) Be in control of (the correct functioning of)
principle suggestions are provided on how the procured and/or outsourced AI applications.
principle can be operationalised in an organisation.
Ad 1: Where possible and appropriate, compliance

4.1 Soundness with applicable regulatory frameworks is
implemented in the design of the AI applications
From a prudential perspective, soundness is the itself (“compliance-by-design”). In case of
aspect of AI that is of DNB’s primary concern. malfunctioning AI systems, fall-back plans are
AI applications in the financial sector should first available for core business processes to ensure
and foremost be reliable and accurate, behave continuity of operations and regulatory compliance.
predictably, and operate within in the boundaries
of applicable rules and regulations, including non- Ad 2: Domain experts are actively involved in the
financial regulations, such as the GDPR. This aspect development and implementation of AI applications
becomes particularly important when financial firms (eg. regarding an AI application to better predict
start to apply identical (or relatively similar) AI driven probability of default levels, experts in credit risk
solutions and systemic risks might arise. Financial and special asset management can be considered
firms applying AI in their business processes should relevant domain experts). Where appropriate,
demonstrate that they have taken all necessary boundaries are set to constrain model outcomes,
measures to ensure their business continuity. serving either as floors/ceilings or as a mechanism
to delegate further decision-making to a domain
expert. The choice for and definition of relevant
evaluation metrics (such as accuracy, precision,
92 These key aspects are partly inspired by the ‘FEAT principles’ from the Monetary Authority of Singapore
(2018).
recall and specificity) are well substantiated and safeguards are in place to maintain and improve 35
documented. Models, where appropriate, are data integrity and security during the process
periodically retrained, recalibrated and assessed, of data collection, data preparation, and data
especially in the event of significant changes in management. Issues with data integrity and bias,
the input data, relevant external factors, and/ both in development and production, are evaluated
or in the legal or economic environment. Criteria and documented in a structural manner for future
for significant change as well as other fail criteria reference. Original data sets used to (re)train and
are well documented for each AI application, and (re)calibrate models are systematically archived.
become more stringent as the model’s materiality
increases. Procedures are in place for personnel to Ad 5: The organisation’s AI policy (and associated
address issues with the choice of models, used data standards and requirements) is applied consistently
sets, and/or model outcomes. to both applications developed in-house as well as
applications developed by third parties. Financial
Ad 3: Criteria for model choices include firms need to understand and be in control of the
considerations other than quantitative evaluation risks associated with their use of AI applications,
metrics, such as the explainability, simplicity, and whether these applications are procured or
reliability of the chosen models. Consistency should outsourced, developed in a partnership, or have any
be maintained, where reasonable, between models other material involvement from third parties.
used for similar analyses (eg. between models used
for loss estimates and models used for pricing).
The impact of incorrect model outcomes for the 4.2 Accountability
organisation are periodically assessed. Material AI
enabled systems are designed with ‘human-in-the- The workings of AI applications can be complex and
loop’ and/or ‘human-on-the-loop’ review processes. difficult to understand. Furthermore, AI applications
The correct workings of material AI applications is may not always function as intended and can result
regularly assessed by evaluating model outcomes in damages for the firm itself, its customers and/
against conventional (non-AI enabled) models. or other relevant stakeholders. Especially when
AI applications become more material, financial
Ad 4: Minimal requirements regarding data quality firms should demonstrate unequivocally that they
are defined. Efforts are made on a continuous understand their responsibility for AI applications
basis to ensure that data are correct, complete, and that they have operationalised accountability
and representative. Special attention is paid to for these applications throughout their organisation.
missing or incorrect data-points, potential sources Model complexity or third party reliance should
of bias in data, features and inference results (such never be used as arguments for limiting the
as selection and survival bias). Procedures and organisation’s accountability.
36 6) Assign final accountability for AI applications 4.3 Fairness
and the management of associated risks
clearly at the board of directors level. Although fairness is primarily a conduct risk issue,
7) Integrate accountability in the organisation’s it is vital for society’s trust in the financial sector
risk management framework. that financial firms’ AI applications – individually
8) Operationalise accountability with regard to or collectively – do not inadvertently disadvantage
external stakeholders. certain groups of customers. Whether we are
talking about an AI that monitors transactions to
detect money laundering, or one that facilitates loan
Ad 6: The board of directors understands their decisions, financial firms should be able to define
responsibilities regarding the adoption and use of their concept of fairness and demonstrate how they
AI, and takes genuine responsibility for the risks ensure that their AI applications behave accordingly.
associated with their AI applications. Operational
accountability is explicitly assigned at all relevant 9) Define and operationalise the concept of
levels of the organisation, and final accountability fairness in relation to your AI applications.
for AI applications and their outcomes (both for the 10) Review (the outcomes of) AI applications for
organisation and their customers) is assigned to one unintentional bias.
(or more) board member(s). The board of directors
accountability explicitly extends to externally
developed and/or sourced AI applications. Ad 9: Fairness is taken into account in the design
and training of the AI application, both in the
Ad 7: The adoption and use of AI is integrated in the selection of input parameters (especially when these
organisation’s risk management framework. Clear contain personal attributes) and potential sources
roles and responsibilities are assigned throughout of unintentional bias (“process fairness”; preventing
the organisation to ensure the responsible use, disparate treatment). Fairness is qualitatively
management and auditability of AI applications. defined in terms of group fairness and/or individual
fairness. Based on the definition, it is subsequently
Ad 8: On request (eg. by a customer or an auditor) operationalised in evaluation metrics, for example
and where appropriate, specific AI driven decisions in terms of false positive and/or false negative rates
and/or model outcomes are reviewed by a domain (“outcome fairness”; preventing disparate impact).
expert in order to verify (or adjust) and explain Trade-offs between process fairness, outcome
these decisions/outcomes. Verified and relevant fairness and accuracy (soundness) are substantiated,
supplementary data provided by customers are well documented and available for review.
taken into account when performing a review of
AI driven decisions.
Ad 10: Material AI enabled systems (in terms of includes criteria that inform decision making on 37
potential impact for customers) are designed with the use (or disuse) of specific models, methods, and
‘human-in-the-loop’ and/or ‘human-on-the- data in AI applications. Procedures are in place that
loop’ review processes to detect and minimise channel ethical and other material concerns that
unintentional bias. Procedures are also in place may arise from the development, implementation,
for ‘after the fact’ reviews, on instigation of the and use of AI to an ethics commission or other
customer and/or other stakeholders. suitable organisational body.
Ad 12: Criteria for the fitness of model outcomes

4.4 Ethics are determined and documented a priori, and are
informed both by legal requirements and by the
As AI applications take on tasks that previously institution’s values and principles. In the design and
required human intelligence, financial firms should approval of customer-oriented AI applications, due
ensure that the outcomes of these systems do consideration is given to the customer’s interest.
not violate the firm’s ethical standards. This moral AI applications do not (inadvertently) exploit
obligation goes above and beyond compliance consumer’s behavioural patterns or psychological
with applicable legal requirements. Financial firms biases in a way that is harmful to their (financial)
should ensure that their customers, as well as other well-being.
stakeholders, can trust that they are not mistreated
or harmed – directly or indirectly – because of the
firm’s deployment of AI, even if such applications 4.5 Skills
operate within the boundaries of applicable laws
and regulations. As most AI applications augment human tasks,
financial firms’ employees and management will
11) Specify objectives, standards, and increasingly come to rely on AI applications to
requirements in an ethical code, to guide the support them in their work. Like with many other
adoption and application of AI. tools, wrong use can result in accidents and it is
12) Align the (outcome of) AI applications with the organisation’s responsibility to ensure that
your organisation’s legal obligations, values their (senior) management, risk management and
and principles. compliance functions have an adequate level of
expertise. From the work floor to the board room,
a sufficient understanding of the strengths and
Ad 11: The organisation's policy includes criteria limitations of the organisation’s AI-enabled systems
that inform consistent decision making on whether is vital.
specific processes and functions are (or are not)
considered suitable for deployment of AI. The policy
38 13) Ensure that senior management has a 4.6 Transparency
suitable understanding of AI (in relation to
their roles and responsibilities). Transparency means that financial firms should be
14) Train risk management and compliance able to explain how they use AI in their business
personnel in AI. processes, and (where reasonably appropriate)
15) Develop awareness and understanding of how these applications function. Adherence to
AI within your organisation. this principle enables adequate risk management
and internal audits. This also means effective
supervision of the correct workings of the firm’s AI
Ad 13: The board of directors’ competences and applications to ensure stable operations. When the
expertise include relevant and up to date knowledge materiality of AI applications increases, so should
and/or experience enabling the board to sufficiently the organisation’s efforts in understanding and
understand the risks associated with the application controlling the applications' functioning.
of AI to the organisation’s core business processes.
Managers responsible for AI applications are trained 16) Be transparent about your policy and
to understand AI-specific challenges, issues and risks. decisions regarding the adoption and use of
AI internally.
Ad 14: Risk management and compliance personnel 17) Advance traceability and explainability of
that is responsible for the assessment and audit AI driven decisions and model outcomes.
of AI-enabled systems is trained to understand
AI-specific challenges, issues, and risks. Their
competence and expertise is of a sufficient level to Ad 16: The organisation’s policy regarding AI is
adequately perform their second line function. clearly communicated throughout the organisation.
Material decisions regarding AI applications, and
Ad 15: Personnel working with AI applications their underlying models and data, are systematically
are trained to responsibly use these applications, documented and sufficiently justified. The
and sufficiently understand their strengths and limitations of adopted models and data sets are
limitations. Educational programs are used to create well documented and communicated, as well as the
and improve awareness regarding the use of AI circumstances under which the use of a particular
applications and their associated risks throughout model or data set should be discontinued. The
the organisation. reasons for specific model choices for AI applications
are systematically documented and available
for review. Decisions that favour accuracy over 39

traceability and explainability should be
well-motivated, documented, and approved at
the appropriate level.
Ad 17: Where reasonable, efforts are made to

improve the reproducibility of operations and
outcomes of AI systems, in order to facilitate review
processes. AI model outcomes are sufficiently
traceable and explainable. This means that, where
feasible, it can be reasonably demonstrated how
(sets of) individual input parameters contribute to
the model’s outcomes on an aggregate level (global
explanation) and/or how individual outcomes tend
to respond to changes in the input variables (local
explanation). In cases where the relation between
input parameters and model outcomes is obscure
because of model choices (eg. when deploying
advanced convolutional neural networks), domain
experts periodically evaluate model outcomes to
improve the organisation’s understanding of the
model’s workings. A continuous effort is made to
develop processes, tools and interfaces to further
improve traceability and explainability of AI
applications to ensure continuing alignment of their
performance with the organisation’s objectives.
Procedures are in place for ‘after the fact’ reviews,
on instigation of the customer and/or other
stakeholders.
References
ABN AMRO. (n.d.). Use the Grip app and manage your Berliner, H. J. (1977). Experiences in evaluation with 41
money. Retrieved from ABN AMRO: BKG-A program that plays backgammon. Proceedings
https://www.abnamro.nl/en/personal/apps/grip/ of the 5th international joint conference on Artificial
index.html intelligence - Volume 1 (pp. 428-33). San Francisco:
Morgan Kaufmann Publishers Inc.
ACPR. (2018). Artificial Intelligence: challenges for the
financial sector. Paris: Autorité de Contrôle Prudentiel Bouchard, K. E., Aimone, J. B., Chun, M., Dean, T.,
et de Résolution . Denker, M., Diesmann, M., . . . Prabhat. (2018).
International Neuroscience Initiatives through the
Akerlof, G. A. (1970). The market for “lemons”: Lens of High-Performance Computing. Computer,
quality uncertainty and the market mechanism. 51(4), 50-9.
The Quearterly Journal of Economcs, 84(3), 488-500.
Boudette, N. E. (2017, January 19). Tesla’s self-driving
Artificial Intelligence Applications Institute. (2005). system cleared in deadly crash. New York Times.
Edinburgh Freddy Robot (Mid 1960s to 1981). Retrieved
from http://www.aiai.ed.ac.uk/project/freddy/ Brown, M. (2017, January 2). Tesla just launched its
futuristic “Enhanced Autopilot” update. Retrieved from
ASR. (n.d.). ASR homepage. Retrieved from Inverse: https://www.inverse.com/article/25848-
https://www.asr.nl/ elon-musk-confirm-enhanced-autopilot-tesla
Bachor, H.-A., & Simmons, M. (2017, November Brown, N., & Sandholm, T. (2018, January 26).
32). Quantum computing. Retrieved from Australian Superhuman AI for heads-up no-limit poker:
Academy of Science: https://www.science.org.au/ Libratus beats top professionals. Science, 359(6374),
curious/technology-future/quantum-computing pp. 418-24.
BaFin. (2018). Big data meets artificial intelligence: Brunkhard, P. (2018, September 3). The future of IOT
challenges and implications for the supervision and is AI. Retrieved from TechUK: https://www.techuk.
regulation of financial services. Frankfurt am Main: org/insights/opinions/item/13827-the-future-of-iot-
Bundesanstalt für Finanzdienstleistungsaufsicht. is-ai
Bawden, A., Greenblatt, R., Holloway, J., Knight, T., BusinessForensics. (n.d.). References. Retrieved from
Moon, D., & Weinreb, D. (1977). LSIP machine progress https://businessforensics.nl/en/references/#clients
report. Cambridge: Massachusetts Institute of
Technology.
42 Chodosh, S. (2017, April 22). The incredible evolution of European Commission. (n.d.). Artificial intelligence.
supercomputers’ powers, from 1946 to today. Retrieved Retrieved from European Commission:
from Popular science: https://www.popsci.com/ https://ec.europa.eu/digital-single-market/en/
supercomputers-then-and-now/ artificial-intelligence
CSAIL. (n.d.). Mission & History. Retrieved from Fairzekering. (n.d.). Zo werkt onze Autofairzekering.
https://www.csail.mit.edu/about/mission-history Retrieved from https://fairzekering.nl/hoe-werkt-
het/
CSSF. (2018). Artificial intelligence: opportunities,
risks and recommendations for the financial sector. FRISS. (n.d.). FRISS fraud detection for insurance claims
Luxembourg: Commission de Surveillance du and lower loss ratios. Retrieved from FRISS:
Secteur Financier . https://www.friss.com/products/fraud-detection-
at-claims/
Dastin, J. (2018, October 10). Amazon scraps secret
AI recruiting tool that showed bias against women. FSB. (2017). Artificial intelligence and machine learning
Reuters. in financial services: Market developments and financial
stability implications. Basel: Financial Stability Board.
deeplearning.ai. (2018, August 25). Heroes of Deep
Learning: Geoffrey Hinton. Retrieved from Giles, M. (2018, December 12). Intel’s new 3D chip
https://www.deeplearning.ai/blog/hodl-geoffrey- technology may help prolong Moore’s Law. Retrieved
hinton/ from MIT Technology Review: https://www.
technologyreview.com/f/612587/intels-new-3d-
DeepMind. (2019). AlphaStar: Mastering the Real-Time chip-technology-may-help-prolong-moores-law/
Strategy Game StarCraft II. Retrieved from
https://deepmind.com/blog/alphastar-mastering- Grand View Research. (2017). Artificial Intelligence
real-time-strategy-game-starcraft-ii/ Market Analysis By Solution (Hardware, Software,
Services), By Technology (Deep Learning, Machine
Dodds, L. (2018, November 25). Chinese business Learning, Natural Language Processing, Machine Vision),
woman accused of jaywalking after AI camera spots By End-use, By Region, and Segment Forecasts, 2018 -
her face on an advert. The Telegraph. 2025. Grand View Research.
EBA. (2019). Consultation paper: draft guidelines on

loan origination and monitoring. European Banking
Authority.
Greenblatt, R. D., Eastlake, III, D. E., & Crocker, S. D. ING. (2019, March 27). ING takes academic approach 43
(1967). The Greenblatt chess program. Proceeding to AI. Retrieved from https://www.ing.com/
AFIPS ‘67 (Fall) Proceedings of the November 14-16, 1967, Newsroom/All-news/ING-takes-academic-
fall joint computer conference (pp. 801-10). Anaheim: approach-to-AI.htm
ACM.
JPMorgan Chase & Co. (2017). Annual report 2016.
Grush, L. (2015, July 1). Google engineer apologizes JPMorgan Chase & Co.
after Photos app tags two black people as gorillas.
Retrieved from The Verge: https://www.theverge. Kochansky, J. (2018, March 8). The promise of artificial
com/2015/7/1/8880363/google-apologizes-photos- intelligence and what it means to BlackRock. Retrieved
app-tags-two-black-people-gorillas from BlackRock Blog: https://www.blackrockblog.
com/2018/03/08/artificial-intelligence-blackrock/
Hern, A. (2015, May 20). Flickr faces complaints over
‘offensive’ auto-tagging for photos. The Guardian. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012).
ImageNet classification with deep convolutional
High-Level Expert Group on Artificial Intelligence. neural networks. Advances in neural information
(2019). Ethics guidelines for trustworthy AI. Brussels: processing systems, (pp. 1097-105).
European Commission.
Kubota, T. (2019, January 16). Stanford’s robotics
ING. (2015, October 21). ING to start strategic legacy. Retrieved from Stanford News: https://news.
partnership and launch pilot with fintech Kabbage. stanford.edu/2019/01/16/stanfords-robotics-legacy/
Retrieved from https://www.ing.com/Newsroom/
All-news/Press-releases/ING-to-start-strategic- Le, Q. V., Ranzato, M., Monga, R., Devin, M., Chen, K.,
partnership-and-launch-pilot-with-fintech- Corrado, G. S., . . . Ng, A. Y. (2012). Building high-level
Kabbage.htm features using large scale unsupervised learning.
Proceedings of the 29th International Conference on
ING. (2017, December 12). Katana gives bond traders a Machine Learning. Edinburgh, Scotland, UK.
cutting edge. Retrieved from https://www.ing.com/
Newsroom/All-news/Katana-gives-bond-traders- Linnainmaa, S. (1970). The representation of the
a-cutting-edge.htm cumulative rounding error of an algorithm as a Taylor
expansion of the local rounding errors. Univ. Helsinki.
ING. (2018, December 17). Using AI to assess credit risk.
Retrieved from https://www.ing.com/Newsroom/ Lun, R., & Zhao, W. (2015). A survey of applications
All-news/Using-AI-to-assess-credit-risk.htm and human motion recognition with Microsoft
Kinect. International Journal of Pattern Recognition and
Artificial Intelligence, 29(05).
44 Lynch, S. (2017, Macrh 14). Andrew Ng: Why AI is Monetary Authority of Singapore. (2018). Principles
the new electricity. Retrieved from Stanford News: to Promote Fairness, Ethics, Accountability and
https://news.stanford.edu/thedish/2017/03/14/ Transparency (FEAT) in the Use of Artificial Intelligence
andrew-ng-why-ai-is-the-new-electricity/ and Data Analytics in Singapore’s Financial Sector.
Singapore: MAS.
Madrigal, A. C. (2017, July 19). How checkers was
solved. Retrieved from The Atlantic: https://www. Moss, S. (2018, June 8). The PlayStation Supercomputer.
theatlantic.com/technology/archive/2017/07/ Retrieved from DatacenterDynamics: https://www.
marion-tinsley-checkers/534111/ datacenterdynamics.com/analysis/the-playstation-
supercomputer/
Malecek, A. (2016, September 2). For first time, carbon
nanotube transistors outperform silicon. Retrieved from Mundy, S. (2016, March 15). AlphaGo conquers
Phys.org: https://phys.org/news/2016-09-carbon- Korean grandmaster Lee Se-dol. Financial Times.
nanotube-transistors-outperform-silicon.html
Municipality of Amsterdam. (2018, October 9).
Markoff, J. (2011, February 17). Computer wins on AI report shows Dutch companies are ahead in Europe.
‘Jeopardy!’: trivial, it’s not. The New York Times, p. 1. Retrieved from I amsterdam: https://www.
iamsterdam.com/en/business/news-and-insights/
Massa, A. (2018, August 30). BlackRock’s Decade: How news/2018/ai-report-shows-dutch-companies-are-
the Crash Forged a $6.3 Trillion Giant. Retrieved from ahead-in-europe
Bloomberg: https://www.bloomberg.com/news/
articles/2018-08-30/blackrock-s-decade-how-the- Nicolau, D. V., Lard, M., Korten, T., van Delft, F. C.,
crash-forged-a-6-3-trillion-giant Persson, M., Bengtsson, E., . . . Nicolau, D. V. (2016).
Parallel computation with molecular-motor-
Microsoft. (2011, January 25). Human pose estimation propelled agents in nanofabricated networks. PNAS,
for Kinect. Retrieved from Microsoft: https://www. 113(10), 2591-6.
microsoft.com/en-us/research/project/human-
pose-estimation-for-kinect/ OECD. (1997). Information Technology Outlook. Paris:
OECD.
Milton, S., Carbonell, J. G., Knoblock, C. A., Kuokka,
D. R., Etzioni, O., & Gil, Y. (1989). Explanation-based Orbis Research. (2017, April 25). Artificial Intelligence-
learning: A problem solving perspective. Artificial Global Market Outlook (2016-2022). Dallas: Orbis
Intelligence, 40(1-3), 63-118. Research. Retrieved from Reuters:
https://www.reuters.com/brandfeatures/venture-
capital/article?id=4999
Oreskovic, A. (2016, March 23). Intel admits the Rosenberg, C. R., & Sejnowski, T. J. (1986). The 45
engine behind the tech industry’s amazing 50-year spacing effect on NETtalk, a massively-parallel
run is slowing down. Retrieved from Business network. Proceedings of the Eighth Annual Conference
Insider: https://www.businessinsider.com/intel- of the Cognitive Science Society (pp. 72-89). New
acknowledges-slowdown-to-moores-law-2016- Jersey: Erlbaum Hillsdale.
3?international=true&r=US&IR=T
Salesforce. (n.d.). Customer success stories. Retrieved
Perry, T. S. (2018, April 2). Move Over, Moore’s Law: from https://www.salesforce.com/nl/customer-
Make Way for Huang’s Law. Retrieved from IEEE success-stories/#!page=1
Spectrum: https://spectrum.ieee.org/view-from-
the-valley/computing/hardware/move-over- Samuel, A. (1959). Some Studies in Machine Learning
moores-law-make-way-for-huangs-law Using the Game of Checkers. IBM Journal of Research
and Development, 3(3), 210-29.
Quillian, M. R. (1968). Semantic networks. In M. L.
Minsky, Semantic Information Processing. MIT Press. Sanberg, A., & Bostrom, N. (2008). Whole brain
emulation: a roadmap. Technical report #2008-3. Future
Rapier, G., & Bastone, N. (2019, April 23). Tesla unveils of Humanity Institute, Oxford University.
the ‘best chip in the world’ for self-driving cars at its
autonomy day event. Retrieved from Business Insider: Sarangi, S., & Sharma, P. (2019). Artificial Intelligence:
https://www.businessinsider.nl/tesla-autonomy- evolution, ethics and public policy. New York:
investor-day-live-updates-event-details-elon- Routledge. Retrieved from https://www.dartmouth.
musk-self-driving-2019-4/ edu/~ai50/homepage.html
Reinsel, D., Gantz, J., & Rydning, J. (2018). The Shortliffe, E. H. (1977). Mycin: A Knowledge-Based
digitization of the world: from edge to core. IDC. Computer Program Applied to Infectious Diseases.
Proceedings of the Annual Symposium on Computer
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). See Application in Medical Care 1977, (pp. 66-9).
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016, August).
Why should i trust you? Explaining the predictions Sicular, S., & Brant, K. (2018). Hype cycle for Artificial
of any classifier. Proceedings of the 22nd ACM SIGKDD Intelligence, 2018. Gartner, Inc.
international conference on knowledge discovery and data
mining (pp. 1135-44). ACM. Silver, D., Hubert, T., Schrittweiser, J., Antonoglou,
I., Lai, M., Guez, A., . . . Hassabis, D. (2018, December
7). A general reinforcement learning algorithm that
masters chess, shogi, and Go through self-play.
Science, 362(6419), pp. 1140-4.
46 Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, Statista. (2019, April 5). Share of big data and business
I., Huang, A., Guez, A., . . . Hassabis, D. (2017, October analytics revenues worldwide in 2019, by industry.
19). Mastering the game of Go without human Retrieved from Statista: https://www.statista.com/
knowledge. Nature, pp. 354-359. statistics/616225/worldwide-big-data-business-
analytics-revenue/
Solomonoff, R. J. (1960). A preliminary report on a
general theory of inductive inference. TOP500.org. (2019, June). TOP500 List. Retrieved
from https://www.top500.org/lists/2019/06/
Sonnad, N. (2017, November 29). Google Translate’s
gender bias pairs “he” with “hardworking” and “she” with Tractica. (2019). Artificial Intelligence Market Forecasts.
lazy, and other examples. Retrieved from Quartz: Tractica.
https://qz.com/1141122/google-translates-gender-
bias-pairs-he-with-hardworking-and-she-with- Turing, A. (1950). Computer machinery and
lazy-and-other-examples/ intelligence. Mind, 49(236), 433-60.
Stanford Research Institute. (n.d.). Shakey Uhr, L., & Vossler, C. (1961). A pattern recognition
the Robot. Retrieved from SRI International: program that generates, evaluates, and adjusts
https://www.sri.com/work/timeline- its own operators. Papers presented at the May 9-11,
innovation/timeline.php?timeline=computing- western joint IRE-AIEE-ACM computer conference.
digital#!&innovation=shakey-the-robot New York: ACM.
Statista. (2015, December 16). Robotics and artificial Vincent, J. (2016, March 24). Twitter taught Microsoft’s
intelligence (AI) worldwide market size estimates, based AI chatbot to be a racist asshole in less than a day.
on 2018 to 2030 forecasts, by segment (in billion U.S. Retrieved from The Verge: https://www.theverge.
dollars). Retrieved from Statista: https://www. com/2016/3/24/11297050/tay-microsoft-chatbot-
statista.com/statistics/621656/worldwide-artificial- racist
intelligence-robotics-segment-estimates/
Vincent, J. (2019, May 7). World’s fastest supercomputer
Statista. (2016, September 30). Share of companies will be built by AMD and Cray for US government.
investing in artificial intelligence (AI) worldwide, by Retrieved from The Verge: https://www.theverge.
industry, as of 2016. Retrieved from Statista: https:// com/2019/5/7/18535078/worlds-fastest-exascale-
www.statista.com/statistics/621544/worldwide- supercomputer-frontier-amd-cray-doe-oak-ridge-
artificial-intelligence-investment-by-industry/ national-laboratory
Weber, B. (1997, May 12). Swift and slashing, Winograd, T. (1971). Procedures as a representation for 47
computer topples Kasparov. The New York Times, p. 1. data in a computer program for understanding natural
language . Cambridge: Massachusetts Institute of
WEF. (2018). The New Physics of Financial Services: Technology.
Understanding how artificial intelligence is transforming
the financial ecosystem. World Economic Forum. World Wide Web Foundation. (n.d.). History of the
Web. Retrieved from https://webfoundation.org/
WEF. (2019). Fourth Industrial Revolution. Retrieved about/vision/history-of-the-web/
from World Economic Forum: https://www.
weforum.org/focus/fourth-industrial-revolution Yadav, N., Yadav, A., & Kumar, M. (2015). An
introduction to neural network methods for differential
Wegdam, M. (2018, February 4). Rabobank starts equations. Dordrecht: Springer.
mobile identification. Retrieved from InnoValor:
https://innovalor.nl/en/Blogs/Rabobank-starts-
with-mobile-identication
Weiser, M. (1991). The computer for the 21st century.

Scientific American, 265(3), 94-105.
Weizenbaum, J. (1966). ELIZA---a computer program

for the study of natural language communication
between man and machine. Communications of the
ACM, 9(1), 36-45.
Welch, D., & Behrmann, E. (2018, May 7). Who’s

Winning the Self-Driving Car Race? Retrieved from
Bloomberg: https://www.bloomberg.com/news/
features/2018-05-07/who-s-winning-the-self-
driving-car-race
Appendix A
Relevant articles in the Financial
supervision Act (unofficial
translation)
48 Financial Supervision Act section 3:10 2. Rules shall be laid down by or pursuant to a
Decree with regard to Subsection (1). These rules
1. A clearing institution, credit institution or insurer shall concern:
having its registered office in the Netherlands a. control of business processes and
shall pursue an adequate policy that safeguards business risks;
controlled and sound business operations. This b. integrity, which is understood to mean the
shall mean that: prevention of:
a. measures are taken to prevent conflicts of i. conflicts of interest;
interest; ii. offences or other transgressions of the law
b. measures are taken to prevent the financial committed by the financial enterprise or its
enterprise or its employees from committing employees that could damage confidence
offences or other transgressions of the law in the financial enterprise or in the financial
that could damage confidence in the financial markets;
enterprise or in the financial markets; iii. relations with clients that could damage
c. measures are taken to prevent confidence confidence in the financial enterprise or in
in the financial enterprise or in the financial the financial markets; and
markets from being damaged because of its iv. other acts performed by the financial
clients; and enterprise or its employees that are so
d. measures are taken to prevent the financial contrary to generally accepted standards
enterprise or its employees from performing as to seriously damage confidence in the
other acts that are so contrary to generally financial enterprise or in the financial
accepted standards as to seriously damage markets;
confidence in the financial enterprise or in the c. the soundness of the financial enterprise,
financial markets. which shall be understood to mean:
i. control of financial risks;
ii. control of other risks that may affect the
Financial Supervision Act section 3:17 soundness of the financial enterprise;
iii. ensuring the maintenance of the required
1. A clearing institution, credit institution or insurer financial safeguards; and
having its registered office in the Netherlands iv. other matters, to be specified by Decree.
shall organise its operations in such a way as 3. Without prejudice to Section 4:14, Subsection (2),
to safeguard controlled and sound business opening words and under (c) shall apply
operations. mutatis mutandis to management companies
of a collective investment scheme having its 3. By or pursuant to a Decree: 49

registered office in the Netherlands that offers a. rules shall be laid down, for the purpose of
units in the Netherlands, collective investment supervision of compliance with the provisions
schemes having their registered office in the arising from this part, with regard to the
Netherlands that offer units in the Netherlands, delegation of activities by financial enterprises;
investment firms having their registered office b. rules shall be laid down with regard to the
in the Netherlands that provide investment control of risks entailed by the delegation
services or perform investment activities in of activities by clearing institutions, credit
the Netherlands, and depositaries associated institutions and insurers; and
with a collective investment scheme having its c. rules shall be laid down with regard to the
registered office in the Netherlands that offers agreement to be concluded between a
units in the Netherlands. clearing institution, credit institution or insurer
and the third party relating to the delegation
of activities.
Financial Supervision Act section 3:18
1. If a financial enterprise having its registered

office in the Netherlands delegates activities
to a third party, the financial enterprise shall
ensure that such third party complies with the
rules applicable under this part to the delegating
financial enterprise with regard to those
activities.
2. A clearing institution, credit institution or insurer
shall not delegate activities to be designated by
Decree.
De Nederlandsche Bank N.V.
PO Box 98, 1000 AB Amsterdam
+31 20 524 91 11
dnb.nl

General Principles For The Use of Artificial Intelligence in The Financial Sector

Uploaded by

Copyright:

Available Formats

General Principles For The Use of Artificial Intelligence in The Financial Sector

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

General Principles For The Use of Artificial Intelligence in The Financial Sector

Uploaded by

Copyright:

Available Formats

General principles for the

use of Artificial Intelligence

©2019 De Nederlandsche Bank

Author: Joost van der Burgt

2 AI: past, present and future 11

4 General principles for the use of AI by financial institutions (‘SAFEST’) 33

Appendix A Relevant articles in the Financial supervision Act

But what do we mean exactly when we talk about

1 See for instance WEF (2019).

2 AI: past, present and future

3 Turing (1950, p. 460).

7 Yadav, et al. (2015).

Figure 1 Computing performance over time 13

Figure 2 Annually created data world-wide 17

34 A convolutional neural network is a class of deep neural networks.

Box 3 When AI gets it wrong: errors, accidents and disasters

2.3 Future (2020-beyond): further discussion regarding the future development 21

Figure 4 Projected computer performance

53 Bouchard, et al. (2018), Sanberg & Bostrom (2008).

Car insurance Various Dutch financial firms are among Salesforce’s 27

product. The insurance company analyses their

The progress of AI in beating humans in increasingly

4 General principles for

it will take for these regulatory responses predictive

assess how responsible their application of AI is and

90 Appendix ‘A’ provides an unofficial translation for the relevant articles.

Ad 1: Where possible and appropriate, compliance

Ad 12: Criteria for the fitness of model outcomes

for review. Decisions that favour accuracy over 39

Ad 17: Where reasonable, efforts are made to

EBA. (2019). Consultation paper: draft guidelines on

Weiser, M. (1991). The computer for the 21st century.

Weizenbaum, J. (1966). ELIZA---a computer program

Welch, D., & Behrmann, E. (2018, May 7). Who’s

of a collective investment scheme having its 3. By or pursuant to a Decree: 49

1. If a financial enterprise having its registered

You might also like

2 AI: past, present and future 11

4 General principles for the use of AI by financial institutions (‘SAFEST’) 33