many contemporary debates.

Industry Fraud detection: Banks use artificial

investments in AI are rapidly increasing intelligence to determine if there is
Artificial Intelligence and Machine
[3], and governments are trying to strange activity on your account.
Learning: Policy Paper
understand what the technology could Unexpected activity, such as foreign
mean for their citizens. [4] transactions, could be flagged by the
Artificial intelligence is a technology that algorithm.
is already impacting how users interact
Speech recognition: Applications use
with, and are affected by the Internet. In The collection of “Big Data” and the
artificial intelligence to optimize speech
the near future, its impact is likely to expansion of the Internet of Things
recognition functions. Examples include
only continue to grow. AI has the (IoT), has made a perfect environment
intelligent personal assistants, e.g.
potential to vastly change the way that for new AI applications and services to
Amazon’s “Alexa” or Apple’s “Siri”.
humans interact, not only with the grow. Applications based on AI are
digital world, but also with each other, already visible in healthcare diagnostics, machine learning, a specific approach to
through their work and through other targeted treatment, transportation, AI and the driving force behind recent
socioeconomic institutions – for better public safety, service robots, education developments.
or for worse. and entertainment, but will be applied in
Artificial Intelligence – What it’s all
more fields in the coming years.
about
Together with the Internet, AI changes
Executive Summary the way we experience the world and Artificial intelligence (AI) traditionally
has the potential to be a new engine for refers to an artificial creation of human-
Artificial Intelligence (AI) is a rapidly economic growth. like intelligence that can learn, reason,
advancing technology, made possible by
plan, perceive, or process natural
the Internet, that may soon have Current Uses of AI:
language.
significant impacts on our everyday
Although artificial intelligence evokes
lives. AI traditionally refers to an Artificial intelligence is further defined
thoughts of science fiction, artificial
artificial creation of human-like as “narrow AI” or “general AI”. Narrow
intelligence already has many uses
intelligence that can learn, reason, plan, AI, which we interact with today, is
today, for example:
perceive, or process natural language designed to perform specific tasks
Email filtering: Email services use within a domain (e.g. language
artificial intelligence to filter incoming translation). General AI is hypothetical
Introduction emails. Users can train their spam filters and not domain specific, but can learn
by marking emails as “spam”. and perform tasks anywhere. This is
Artificial intelligence (AI) has received outside the scope of this paper. This
increased attention in recent years. Personalization: Online services use
paper focuses on advances in narrow AI,
Innovation, made possible through the artificial intelligence to personalize your
particularly on the development of new
Internet, has brought AI closer to our experience. Services, like Amazon or
algorithms and models in a field of
everyday lives. These advances, Netflix, “learn” from your previous
computer science referred to as machine
alongside interest in the technology’s purchases and the purchases of other
learning.
potential socio-economic and ethical users in order to recommend relevant
impacts, brings AI to the forefront of content for you.
Machine learning – Algorithms that Unsupervised learning: The data given layered neural networks – also known as
generate Algorithms to the learning algorithm is unlabeled, “deep learning” – have inspired new
and the algorithm is asked to identify services, but is also spurring
Algorithms are a sequence of
patterns in the input data. For example, investments and research in other parts
instructions used to solve a problem.
the recommendation system of an e- of the field.
Algorithms, developed by programmers
commerce website where the learning
to instruct computers in new tasks, are KEY CONSIDERATIONS
algorithm discovers similar items often
the building blocks of the advanced
bought together. As machine learning algorithms are used
digital world we see today. Computer
in more and more products and services,
algorithms organize enormous amounts Reinforcement learning: The
there are some serious factors must be
of data into information and services, algorithm interacts with a dynamic
considered when addressing AI,
based on certain instructions and rules. environment that provides feedback in
particularly in the context of people’s
It’s an important concept to understand, terms of rewards and punishments. For
trust in the Internet:
because in machine learning, learning example, self-driving cars being
algorithms – not computer programmers rewarded to stay on the road.1 Socio-economic impacts. The new
– create the rules. functions and services of AI are expected
to have significant socio-economic
The basic process of machine learning is
The current growth in AI and machine impacts. The ability of machines to
to give training data to a learning
learning is tied to developments in three exhibit advanced cognitive skills to
algorithm. The learning algorithm then
important areas: process natural language, to learn, to
generates a new set of rules, based on
plan and to perceive, makes it possible
inferences from the data. This is in Data availability: Just over 3 billion
for new tasks to be performed by
essence generating a new algorithm, people are online with an estimated 17
intelligent systems, sometimes with
formally referred to as the machine billion connected devices or sensors. [9]
more success than humans
learning model. That generates a large amount of data
which, combined with decreasing costs Transparency, bias and
of data storage, is easily available for accountability. AI-made decisions can
How machines learn: use. Machine learning can use this as have serious impacts in people’s lives. AI
training data for learning algorithms, may discriminate against some
Although a machine learning model may
developing new rules to perform individuals or make errors due to biased
apply a mix of different techniques, the
increasingly complex tasks. training data.
methods for learning can typically be
categorized as three general types: Computing power: Powerful computers New uses for data. Machine learning
and the ability to connect remote algorithms have proved efficient in
Supervised learning: The learning
processing power through the Internet analyzing and identifying patterns in
algorithm is given labeled data and the
make it possible for machine-learning large amounts of data, commonly
desired output. For example, pictures of
techniques that process enormous referred to as “Big Data”. Big Data is
dogs labeled “dog” will help the
amounts of data. [10] used to train learning algorithms to
algorithm identify the rules to classify
increase their performance. This
pictures of dogs. Algorithmic innovation: New machine
generates an increasing demand for
learning techniques, specifically in
data, encouraging data collection and Data Quality and Bias. In machine Accountability. The strength and
raising risks of oversharing of learning, the model’s algorithm will only efficiency of learning algorithms is based
information at the expense of user be as good as the data it trains on – on their ability to generate rules without
privacy. commonly described as “garbage in, step-by–step instructions. While the
garbage out”. This means biased data technique has proved efficient in
Security and safety. Advancements in
will result in biased decisions. accomplishing complex tasks such as
AI and its use will also create new
facerecognition or interpreting natural
security and safety challenges. These Safety and Security. As the AI agent
language, it is also one of the sources of
include unpredictable and harmful learns and interacts with its
concern.
behavior of the AI agent, but also environment, there are many challenges
adversarial learning by malicious actors. related to its safe deployment. They can Social and Economic Impact. It is
stem from unpredictable and harmful predicted that AI technologies will bring
Ethics. AI may make choices that could
behavior, including indifference to the economic changes through increases in
be deemed unethical, yet also be a
impact of its actions productivity. This includes machines
logical outcome of the algorithm,
being able to perform new tasks, such as
emphasizing the importance to build in In reinforcement learning this stems
self-driving cars, advanced robots or
ethical considerations into AI systems from the so-called exploration/exploit
smart assistants to support people in
and algorithms. dilemma. This means an AI agent may
their daily lives
depart from a successful strategy of
New ecosystems. Like the impact of
solving a problem in order to explore Governance. The institutions, processes
mobile Internet, AI makes new
other options that could generate a and organizations involved in the
applications, services, and new means of
higher payoff. governance of AI are still in the early
interacting with the network possible.
stages. To a great extent, the ecosystem
There is also a risk that autonomous
Challenges overlaps with subjects related to
systems are exploited by malicious
Internet governance and policy. Privacy
Many factors contribute to the actors trying to manipulate the
and data laws are one example.
challenges faced by stakeholders with algorithm. The case of “Tay”, a chatbot
the development of AI, including: deployed on Twitter to learn from Ethical Considerations in Deployment
interactions with other users, is a good and Design
Decision-making: transparency and example. It was manipulated through a
“interpretability”. With artificial Principle: AI system designers and
coordinated attack by Twitter users,
intelligence performing tasks ranging builders need to apply a user-centric
training it to engage in racist behavior.
from self-driving cars to managing approach to the technology. They need
[17] Other examples of so-called
insurance payouts, it’s critical we to consider their collective responsibility
“adversarial learning” include attacks
understand decisions made by an AI [25] in building AI systems that will not
that try to influence the training data of
agent. But transparency around pose security risks to the Internet and
spam filters or systems for abnormal
algorithmic decisions is sometimes Internet users.
network traffic detection, so as to
limited by things like corporate or state mislead the learning algorithm for
secrecy or technical literacy subsequent exploitation
Recommendations: Make safety a priority: Any
deployment of an autonomous system
Adopt ethical standards: Adherence to Recommendations:
should be extensively tested beforehand
the principles and standards of ethical
“Algorithmic Literacy” must be a basic to ensure the AI agent’s safe interaction
considerations in the design of artificial
skill: Whether it is the curating of with its environment (digital or
intelligence [26], should guide
information in social media platforms or physical) and that it functions as
researchers and industry going forward.
self-driving cars, users need to be aware intended.
Promote ethical considerations in and have a basic understanding of the
Privacy is key: AI systems must be data
innovation policies: Innovation policies role of algorithms and autonomous
responsible. They should use only what
should require adherence to ethical decision-making.
they need and delete it when it is no
standards as a pre-requisite for things
Provide the public with information: longer needed (“data minimization”).
like funding.
While full transparency around a They should encrypt data in transit and
Ensure “Interpretability” of AI service’s machine learning techniques at rest, and restrict access to authorized
systems and training data is generally not persons (“access control”). AI systems
advisable due to the security risk, the should only collect, use, share and store
Principle: Decisions made by an AI agent data in accordance with privacy and
public should be provided with enough
should be possible to understand, personal data laws and best practices.
information to make it possible for
especially if those decisions have
people to question its outcomes.
implications for public safety, or result Think before you act: Careful thought
in discriminatory practices. Responsible Deployment should be given to the instructions and
data provided to AI systems. AI systems
Recommendations: Principle: The capacity of an AI agent to should not be trained with data that is
act autonomously, and to adapt its biased, inaccurate, incomplete or
Ensure Human Interpretability of
behavior over time without human misleading.
Algorithmic Decisions: AI systems
direction, calls for significant safety
must be designed with the minimum If they are connected, they must be
checks before deployment, and ongoing
requirement that the designer can secured: AI systems that are connected
monitoring.
account for an AI agent’s behaviors. to the Internet should be secured not
Recommendations: only for their protection, but also to
Empower Users: Providers of services
that utilize AI need to incorporate the Humans must be in control: Any protect the Internet from malfunctioning
ability for the user to request and autonomous system must allow for a or malware-infected AI systems that
receive basic explanations as to why a human to interrupt an activity or could become the next-generation of
decision was made. shutdown the system (an “off-switch”). botnets. High standards of device,
There may also be a need to incorporate system and network security should be
Public Empowerment applied.
human checks on new decision-making
Principle: The public’s ability to strategies in AI system design, especially Responsible disclosure: Security
understand AI-enabled services, and where the risk to human life and safety researchers acting in good faith should
how they work, is key to ensuring trust is great. be able to responsibly test the security
in the technology. of AI systems without fear of
prosecution or other legal action. At the Social and Economic Impacts
same time, researchers and others who
Principle: Stakeholders should shape an
discover security vulnerabilities or other
environment where AI provides socio-
design flaws should responsibly disclose
economic opportunities for all.
their findings to those who are in the
best position to fix the problem. Recommendations:
Ensuring Accountability All stakeholders should engage in an
ongoing dialogue to determine the
Principle: Legal accountability has to be
strategies needed to seize upon artificial
ensured when human agency is replaced
intelligence’s vast socio-economic
by decisions of AI agents.
opportunities for all, while mitigating its
Recommendations: potential negative impacts. A dialogue
could address related issues such as
Ensure legal certainty: Governments
educational reform, universal income,
should ensure legal certainty on how
and a review of social services.
existing laws and policies apply to
algorithmic decision-making and the use Open Governance
of autonomous systems to ensure a
Principle: The ability of various
predictable legal environment. This
stakeholders, whether civil society,
includes working with experts from all
government, private sector or academia
disciplines to identify potential gaps and
and the technical community, to inform
run legal scenarios
and participate in the governance of AI is
Put users first: Policymakers need to crucial for its safe deployment.
ensure that any laws applicable to AI
Recommendations:
systems and their use put users’
interests at the center. This must include Promote Multistakeholder
the ability for users to challenge Governance: Organizations, institutions
autonomous decisions that adversely and processes related to the governance
affect their interests. of AI need to adopt an open, transparent
and inclusive approach. It should be
Assign liability up-front: Governments
based on four key attributes:
working with all stakeholders need to
Inclusiveness and transparency;
make some difficult decisions now about
Collective responsibility; Effective
who will be liable in the event that
decision making and implementation
something goes wrong with an AI
and Collaboration through distributed
system, and how any harm suffered will
and interoperable governance
be remedied.