ARTIFICIAL INTELLIGENCE (AI)

ETHICS GUIDE
2 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

This Guide is a resource for local-level policymakers across sectors seeking to understand the challenges,
opportunities, and risks of AI adoption.

The Guide includes a definition of AI, a discussion of ethical issues related to AI, and proposals for how these issues
can be addressed.

It also includes examples to illustrate AI-related ethical issues, approaches for addressing these issues, and relevant
reflection questions. Specifically, this Guide uses the education sector as an example to explore elements of ethical AI,
but the lessons are generalizable to other sectors as well.

Finally, the Guide includes a Glossary and an Annex with additional details of case studies highlighted in the document.
3 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

WHAT IS ARTIFICIAL INTELLIGENCE (AI)?

Artificial intelligence, or AI, is a concept referring to computer algorithms that solve problems using
techniques associated with human intelligence: logical reasoning, knowledge representation and recall,
language processing, and pattern recognition. AI is currently used to build a wide variety of applications—
from customer service chatbots to complex earthquake or crime prediction programs.

WHY DO WE USE AI? WHAT IS THE DIFFERENCE BETWEEN AI

AI offers an exciting extension of many human capabilities AND MACHINE LEARNING?
such as observation, processing, and decision-making. The Machine learning (or ML) enables AI systems. While ML
output and outcomes of AI systems are nearly focuses on learning and prediction, AI applications often
instantaneous, offering humans powerful efficiencies that rely on the predictions from ML to create, plan, or have
did not exist just a few years ago. The computing power an impact in the real world. Automated decisions might
and systems used for AI technologies far exceed human be directly implemented (like in robotics) or suggested to
cognitive capabilities, allow for constant “machine a human decision-maker (in the form of product
learning” without human supervision, and include recommendations in an online shopping session).2
consideration of patterns that are typically impossible for
humans to discern (e.g., the ability to identify an individual
WHY DOES AI NEED DATA?
based on their gait without ever seeing their face). AI can
Large datasets, such as those generated by a state or
also use dynamic nudging to create instant incentives for
national education system, enable AI. While AI systems
compliance (e.g., in the commercial context, a guided
have actually been around since the 1950s, AI has recently
selection of benefits, personalized to the customer,
received increased attention due to its new capabilities,
designed to promote specific economic behavior).
powered by larger quantities of available data and more
powerful computing systems. For instance, data on
HOW DOES AI WORK? population movements, the spread of epidemics, and
In an AI system, predefined algorithms follow a series of climate change can significantly enhance the capabilities of
instructions to transform data into outputs that can be today’s AI systems to monitor emerging trends and
used for making decisions, either by the computer system provide evidence to advance on the UN Sustainable
itself or a human. Many AI algorithms learn directly from Development Goals.
data, by identifying patterns and relationships, without
rules-based instructions from humans—as is the case in
traditional statistical programming.1 The “black box” For reflection
nature of AI systems refers to system inputs and Access to sufficient, unbiased, good-
operations which are not visible to the end-user or other
quality data is a foundational driver of
parties, and sometimes even to the data scientists who
build AI systems. Algorithms that can act independently
trustworthy and ethical AI.
without control or supervision from humans are
considered autonomous.

For reflection
AI is dependent on the data it receives
for decision-making. As with all
computing systems, if the data used
is biased or corrupted, so will be the
outputs or recommended decisions by
the AI system.

1. Prasanna Lal Das (2022). Algorithms in Government: A Magic Formula or a Divisive Force? https://dial.global/research/algorithms-in-
government-a-magic-formula-or-a-divisive-force/
2. USAID Digital Strategy (2020-2024). https://www.usaid.gov/digital-development/usaid-digital-strategy.
 en boga debido a una mayor El acceso a datos suficientes,
capacidad
4
deARTIFICIAL
cómputo. Por
INTELLIGENCE (AI) ETHICS GUIDE
imparciales y de buena calidad es un
tos de datos sobre movimientos
pagación deWHAT
epidemias y el
IS AI ETHICS?
factor fundamental para lograr una
ntre muchosthatotros,
ethics is pueden
IA éticaFor
Ethics is the science of proper behavior. Aristotle argued
the study of human relations in their most
y confiable.
reflection
mente las capacidades
perfect form; he de la IA
claimed that ethics is the basis for
creating an ideal model of human interrelations, ensuring
There is no common definition of AI
ndencias emergentes y proveer
optimal communication between people and a reference ethics.
zar en los Objetivos de Desarrollo
point for creating a structure of moral consciousness. The
practice of AI ethics is the consideration of moral
e las Naciones Unidas.
problems related to the interaction of technology,
humans, and society. Just as AI is evolving rapidly, so too
are AI ethics considerations. Ultimately, the overall goal of
AI ethics is to seek to create an optimal model of
interrelations between humans and technology.

TICA DE LA IA?
WHY DO WE USE AI IN THE
a del comportamiento
EDUCATIONadecuado. SECTOR?
AI systems change the nature of educational systems
ntó que la ética
throughestheelability
estudio
to analyzede
large and complex datasets,
nas en su forma más perfecta.
automate processes, understand changes in student
performance, create new ways of interaction between
se para crearstudents
un modelo
and teachers,óptimo
customize learning methods, and
more. At the same time, the potential for careless
umanas y justas, asegurando
implementation of AI creates serious ethical risks with
ptima entrelong-term
las personas y un generates prejudice, can
negative consequences,
inhibit motivation, or even trigger social unrest.
para crear una estructura de
See more definitions of concepts related to AI ethics in the
a práctica deGlossary
la ética
at the de
end oflathis
IAGuide.
es
problemas morales relacionados
e la tecnología,For losreflection
humanos y la
The ethics of AI today is more about
la IA evoluciona rápidamente,
the right questions than the right
s consideraciones éticas de la IA;
answers.
etivo general de la ética de la IA
modelo óptimo de interrelaciones
cnología.

AMOS IA EN EL SECTOR Para reflexionar

modifican la naturaleza de los No existe una definición común de

a través de la capacidad de la ética de la IA.
e datos grandes y complejos,
os, comprender los cambios
e los estudiantes, crear
teracción entre estudiantes y
5 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

WHY AI ETHICS IS Case study

IMPORTANT Imperfect AI solutions in
While AI can bring many benefits for human beings, it
education
also comes with ethical considerations that cannot be Country: UK
ignored. As these systems are increasingly being used
across sectors, AI can have significant effects on credit, Year: 2020
employment, education, competition, and more. What happened: The students enrollment
However, without ethics embedded in AI algorithms, it is algorithm favored students from private
hardly possible to guarantee that AI will not enable schools and affluent areas, leaving high-
certain actors to cause more harm than good. With the achievers from free, state-schools
wider adoption of AI in recent years, it has been used to disproportionately affected.
sow distrust in public information and and has been held
responsible for perpetuating discrimination in the delivery Why it is important for the purposes
of services and unfavorably profiling segments of the of this Guide? Many policymakers focus
population, raising many other moral concerns. only on positive effects that AI bring to the
education sector, however, the identification
The responsibility of data scientists, teams of software and mitigation of AI’s risks should be
developers, and other participants in the AI lifecycle3 prioritized over implementation.
rarely extends to AI ethics. Software engineers frequently
See details in the Annex.
pay more attention to how well a system or product is
performing its intended function than to its social and
ethical implications and longer-term ramifications. That is
why it is important to engage different stakeholder
groups, including a diverse range of members of civil
society, to ensure ethical design and deployment of AI
systems.

3. “What are AI Blindspots?” AI Blindspot Media. Accessed June 9, 2023. https://aiblindspot.media.mit.edu.

6 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

WHAT ARE AI-RELATED MORAL

PROBLEMS? Case study
AI algorithms can be opaque and complex, and subject to Flawed facial recognition
error, bias, profiling, discrimination, and other unfair
systems
practices. This happens, in part, because algorithms are
created by human AI engineers—and humans are not Country: USA
objective. This is also due to historical biases and
Year: 2018
prejudices in the datasets AI systems learn from. If AI
engineers do not identify address data bias, AI systems What happened: Google facial recognition
may replicate them. Ethical issues appear in many aspects software had a bias against African
of AI adoption. Examples include the choices that Americans; image recognition algorithms
self-driving cars are programmed to make when a crash is in Google Photos were classifying African
inevitable;4 racial, gender, and age biases in facial Americans as “gorillas.”
recognition; biases for people with disabilities; and AI
algorithms making discriminatory decisions in the
Why it is important for the purposes
insurance and banking sectors. There are recent cases in
of this Guide? Currently, numerous
which people have blamed algorithms for injustices that
government entities consider the use of facial
negatively impact their lives, such as being denied bail by
recognition systems, without considering
judges reliant on automated systems or their children these systems’ potential risks and flaws, as
being unfairly deprived of college admissions. 5 illustrated by this use case.
See details in the Annex.

For reflection
How would machines solve the “Trolly
Dilemma”?

WHO IS RESPONSIBLE FOR AI

OUTPUTS?
Humans curate datasets, create and serve as “role
models” for algorithms, and consume the outputs of
algorithms. Humans also provide an essential point of
control for algorithms, either as testers or validators of
algorithmic decisions. The societal question of who,
exactly bears responsibility for errors in AI decision-
making is a profound one: is the fallout from an AI-made
error the fault of the developer of the AI algorithm, the
owner of the smart device, the operator of the algorithm
or AI-tool, the person who provided input data to the
algorithm, or someone else entirely? The lack of
certainty on this front may perpetuate and exacerbate
distrust of AI systems, leading to long-lasting
consequences for innovation and adoption of technology.

4. Orin Kerr. “A ‘Trolley Problem’ Dilemma for the FBI.” The New York Times. January 27, 2016. https://www.nytimes.com/
roomfordebate/2016/01/27/the-ethics-of-a-child-pornography-sting/a-trolley-problem-dilemma-for-the-fbi
5. “Racism In, Racism Out. A Primer on Algorithmic Racism.” 2021. Public Citizen. https://www.citizen.org/article/algorithmic-racism/.
7 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

WHAT ARE SOME OF THE MORAL

ISSUES SURROUNDING AI? For reflection
AI applications such as chatbots, criminal assessment If an AI system generates an incorrect
systems, facial recognition systems have provided sexist or
medical diagnosis that leads to the
racist outputs because of biases inherent in input data.
These biases are often found in open-source data, which death of a patient, who is responsible?
is not necessarily representative of global populations. An
AI algorithm draws conclusions from what it has studied;
when deep learning systems are trained on open data, no
one can control what exactly they learn. Using historical Case study
data may also create modern bias for AI tools as older
AI against people with
datasets may not include data on women, people with
disabilities, indigenous peoples, or other historically disabilities
disenfranchised groups, many of whom are still on the Country: Global
“unfavorable” side of the digital divide. In some cases, an
AI algorithm would likely consider these populations as What happened: Persons with disabilities
outliers and create models that lead to further errors that may be interpreted as outliers by AI
perpetuate their disenfranchisements. applications that may mimic the direct and
indirect discrimination they face in society.
WHAT ARE THE RISKS? Why it is important for the purposes
Ethical problems in AI can lead to a variety of of this Guide? When considering the
consequences with different levels of severity; these development and adoption of an AI system
consequences include everything from increased inequality at the local level, it is important to identify
to extended litigation processes to resist social uprising. if there is representative data that includes
The profiling and biases of algorithms against a particular people with disabilities, for example, to avoid
race, gender, or specific category of people can affect how their exclusion from or discrimination within
education, healthcare, financial, and democratic systems the results or recommendations of such a
work. AI can be also used maliciously, in any of these fields system.
to fake data, steal passwords, and interfere with the work
See details in the Annex.
of other software and machines, thus undermining public
trust in technology even more. These digital crimes put
core human values such as personal privacy, data
protection, fairness, and autonomy at risk.

6. AI Decolonial Manyfesto.” Mayfesto.ai. Accessed June 9, 2023. https://manyfesto.ai.

7. “Principles of Māori Data Sovereignty.” 2018. Te Mana Raraunga. https://cdn.auckland.ac.nz/assets/psych/about/our-research/documents/
TMR+M%C4%81ori+Data+Sovereignty+Principles+Oct+2018.pdf
enas o cualquier otro grupo que históricamente de esta guía? Equipos de Recursos Human
cuentran8 en ARTIFICIAL
el ladoINTELLIGENCE
desfavorable (AI) ETHICS de la brecha
GUIDE tanto en los sectores privado como público
l. En esos casos, los algoritmos pueden están apostando al aprovechamiento de IA pa
derar estas poblaciones
WHO SHOULD como valores atípicos
BE INTERESTED y
IN AI ETHICS?
eficientar los procesos de contratación, sin
modelos que conducen
Everyone. Citizens, a errores
businesses, y problemas
governments, and academia are all userstener en ofcuenta
or adopters ejemplos
AI applications de casos como éste
and technologies.

s de manera They
queallyface, y los efectos negativos de una contratación
or will face, AI-related ethical challenges that will impact their lives or livelihoods in some way. And they all
esas poblaciones seguirán
have a perspective to add to our society’s conversation surrounding ethical AI use.
o invisibles en la era moderna. sesgada.
DO AI SYSTEMS HAVE “GLOBAL” BIASES?
Ver detalles en el anexo.
AI systems development, data collection, and standards creation are happening in the Global North, where most AI
research institutions and big tech companies are located, while most countries of the Global South still experience data
poverty and are still forming reliable and robust digital infrastructures. Thus, there are concerns about whether Western
conceptions of “fairness” and equality should be considered universal,6 and if they should be applied in the same manner
in developing countries as they are in advanced economies. Many parties in developing nations have openly questioned the
presence of Western “values” in AI systems and wonder whether alternative beliefs held in developing nations could
inform ethical AI systems that could be accepted globally. Other concerns, related to maintaining the data privacy of
global indigenous peoples in the face of new AI technologies, have been raised as well. One example is Maori Principles
for Data Sovereignty.7
Los 9gobiernos
ARTIFICIALde todo el(AI)
INTELLIGENCE mundo han comenzado a desarrollar estrategias tanto para aprovechar las
ETHICS GUIDE

oportunidades como para mitigar los riesgos relacionados con el uso de la IA. Un método que han
empleado es el desarrollo de pautas éticas para el uso de IA y otros sistemas automatizados. Si bien
muchas deHOW estas guíasTO APPLYporAIinstituciones
son producidas ETHICS u organismos gubernamentales que carecen de
atribuciones legales para su aplicación, sientan la base de reglas, comienzan a establecer estándares y
Governments around the globe are considering both the opportunities and risks of AI deployment.
crean conciencia entrehave
Certain entities losdrafted
distintos actores
ethical gubernamentales.
guidelines Además,
for use of AI and other las organizaciones
automated systems, and while delmany
sector
privado, oflasthese
instituciones académicas
guides are produced y las organizaciones
by institutions or governmental nooffices without explicittambién
gubernamentales 8
han creado sus
legal or legislative
propios principios éticos o pautas para un público más amplio. A continuación se comparten
powers, such guides create a baseline of rules, begin to establish standards, and build awareness for una serie
de buenasprominent
prácticasgovernment actors.hacia
en el avance Further,
unaprivate sector organizations, academic institutions, and non-
IA ética.
government organizations have also created their own sets of AI ethical principles or guidelines for a
8

wider audience. Below, you will find several key suggestions that appear in one form or another in many
these guides that can help foster the ethical use of AI:
1. CREAR CONCIENCIA E INICIAR EL
DIÁLOGO
1. CREATE AWARENESS AND INITIATE CASO DE ESTUDIO: ALEMANIA,
COMISIÓN
Case study DE ÉTICA
La ética de laDIALOGUE
IA es igualmente importante para el
AI ethics is equally important for government, businesses,
gobierno,andlasindividuals
empresas y las personas; debe AI against
País: people
Alemania Año:with2018
to understand; however, mostdiscutirse
people do
abierta y not
ampliamente. Sin embargo, el tema está disabilities
have a full understanding of it. Thus, a good starting ¿Qué sucedió?: La Comisión de Ética de
fuera del radar para muchos. Es por eso que un
point during AI implementation is to launch awareness Country:
Datos se Germany
estableció para producir puntos de
buen puntoactivities that bringsería
de partida theselanzar
stakeholders to the table.
actividades de It is Year: 2018
referencia éticos, pautas y recomendaciones
important to ensure that these activities catalyze
concientización y traer a la mesa a las distintas partes
wide-ranging engagement and discussion on AI ethics para elhappened:
What desarrolloThe
y uso
DatadeEthics
IA para el gobierno.
interesadas.
issuesEsand
importante
how they impactgarantizar
national, que exista and
sub-national, Commission was established to produce
la participación
municipalyagencies;
el debate más
small- andamplios posibles
medium-sized enterprises ¿Por qué es relevante para efectos
ethical benchmarks, guidelines, and
sobre cuestiones
(SMEs) andde ética
larger de la IA,
businesses; incluidos
professional los
associations; de esta guía? for
recommendations Esta
thebuena prácticaand
development puede
and citizens.
organismos nacionales, subnacionales y municipales, serofcontextualizada
use por gobiernos locales
AI for the government.
las empresas (tanto grandes empresas como las que busquen
Why contar for
it is important conthe
grupos de trabajo
purposes
pequeñas y medianas), las asociaciones profesionales y especializado
of this Guide? en
Thislagood
materia.
practice can be
contextualized by local governments that
la ciudadanía. Vertodetalles
seek en elwork
have specialized anexo.
groups for
responsible use of AI on this matter.
See details in the Annex.

8. “AI Ethics for Latin America.” 2022. C Minds. https://www.cminds.co/booklet.

8 Véase el booklet para emprendimientos y empresas Ética de la IA para América Latina de C Minds y Meta. Disponible en
10 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

2. ADOPT GUIDING PRINCIPLES

Many countries, cities, and organizations have developed Case study
their own principles and guidelines, codes of ethics, and
Canada sets up guiding
self-assessment toolkits on AI for public officials and
business organizations.9 At the heart of each independent
principles
document lies a universal list of guiding principles that do Country: Canada
not bear any legislative power but help to shape ethical
application of AI for all. The general list comprises the Year: 2018
following five principles: What happened: The Treasury Board
developed a set of guiding principles to help
government officials explore AI in a way that
Transparency
is “governed by clear values, ethics, and laws.
AI uses open, understandable, auditable, and
documented use of any data with clear and Why it is important for the purposes
up-to-date security measures. of this Guide? Aspects and lessons learned
from the Canadian experience can be
Non-maleficence and trust considered in the design of guiding principles
AI does no harm or inflicts the least possible aligned to the local reality of state and
harm to reach an outcome. Continuous municipal governments.
operator readiness to interfere in AI systems
See details in the Annex.
is ensured.
Equitability and fairness
AI systems and solutions subordinate to
human- defined rules and laws, thus ensuring
human rights above all.
Autonomy and responsibility
The AI system discloses who is responsible
for AI works and solutions. At the same time,
it recognizes that humans willingly give up
some of decision-making power to AI.
Explainability
To ensure accountability, AI systems and
their decision-making processes need to be
explainable to a human being at a basic and
acceptable level.

There are other AI guiding principles adopted by different

organizations and that are fixed in internationally
recognized documents, such as the Montreal Declaration
for Responsible AI10 or OECD’s principles,11 which are
worth getting acquainted with in detail.

9. See, for instance, the UN Moratorium on use of AI that threatens human rights, or the https://fairlac.iadb.org/en/fair-lac-box (2022).
10. “The Montréal Declaration for Responsible AI Development.” 2018. Université de Montréal. https://monoskop.org/images/d/d2/Montreal_
Declaration_for_a_Responsible_Development_of_Artificial_Intelligence_2018.pdf.
11. “The OECD AI Principles.” 2019. OECD.AI Policy Observatory. https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449.
11 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

3. DEVELOP A LEGAL FRAMEWORK FOR AI ETHICS

Most AI ethics initiatives developed today have inspired AI guidelines and frameworks that are recommended but not
compulsory. Developing ethical standards for data processing using AI can be an important step towards setting up a legal
foundation for ethical AI. The standard can be in the form of methodological recommendations for developing ethical
codes for organizations participating in the development and introduction of AI technologies.

Case study
Open explanations of AI from Singapore
Country: Singapore Year: 2020
What happened: An AI startup disclosed the exact parameters used in developing its AI model to its
clients in the healthcare sector. The startup made a conscious decision to declare the use of AI and in
its analysis and prediction.
Why it is important for the purposes of this Guide? This example sets a precedent for a
successful business case where there is a balance between transparency in the use of AI and privacy of
user data.
See details in the Annex.

4. DRAFT SECTOR-SPECIFIC GUIDELINES FOR THE USE OF AI

AI ethics varies when it comes to AI implementation in specific sectors. Expecting actors in sectors like healthcare,
e-commerce, and education to simply adhere to common principles of fairness and justice when using AI will generally
not be a sufficient approach; instead, it is pivotal for guidelines for AI use in these sectors will need to provide a detailed
framework relevant to each sector to guide their deployment of AI technologies. Sectoral agreements on such
frameworks can be beneficial for the ubiquitous application of AI.

Case study
Resource for UK public authorities
Country: UK Year: 2020
What happened: The government developed an AI guide for public authorities to better understand
how AI works and how it can be applied in the public sector.
Why it is important for the purposes of this Guide? The UK is a leading country in developing
guidelines for the responsible and ethical use of AI. This exercise represents a good practice that
groups guiding principles on the matter.
See details in the Annex.

5. ENCOURAGE TRANSPARENCY AND DISCUSSION OF NOVEL DATA USES WITH

ETHICAL IMPLICATIONS AS THEY EMERGE
The data economy is a highly dynamic field, and the adoption of laws, norms, and standards often does not keep pace
with the introduction and use of technologies—especially AI technologies. Permanent and practical discussions around
emerging ethical implications of AI are important if a sustainable digital and data-driven transformation is to be achieved.
Close collaboration with civil society could help monitor these rapid technological changes and ensure continued
transparency regarding AI.
12 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

6. ENGAGE WITH THE DEVELOPMENT OF INTERNATIONAL PRINCIPLES OF AI

ETHICS
Countries such as the United Kingdom, Canada, Germany, Japan, Argentina, the United Arab Emirates, along with
international organizations such as the European Commission, the Organization for Economic Co-operation and
Development (OECD), the Global System for Mobile Communications Association (GSMA), and many others actively
contribute to the AI Ethics agenda by issuing codes of ethics and guidelines and organizing consortiums such as the Global
Partnership for AI (GPAI) to work together on facilitating ethical adoption of AI.12

Case study
Resource for UK public authorities
Country: Global Year: 2019
What happened: OECD published its AI Principles, which have been accepted by all 37 OECD
countries and seven non-member partners, including Mexico. These principles also formed the basis of
G20 AI Principles.
Why it is important for the purposes of this Guide? It is important that the local government
exploring the potential use of AI takes into consideration whether their country has adhered to the
OECD AI Principles in order to take them as a starting point for any public policy development.
See details in the Annex.

7. UTILIZE PRIVATE SECTOR INITIATIVES FOR PUBLIC GOOD

The private sector is also deploying publicly available AI tools that help to mitigate AI risks. Examples include the AI
Fairness 360 tool by IBM, Google’s People + AI Research (PAIR), the Aequitas Bias and Audit Toolkit from Carnegie
Mellon University, Fairlearn, Datasheets for Datasets from Microsoft, and Tensorflow Model Cards. These tools have
profound potential applications for AI ethics; for instance, the IBM AI Fairness tool is a comprehensive open-source
toolkit of metrics to check for unwanted bias in datasets and machine learning models, and its state-of-the-art algorithms
to mitigate such bias are applied by many practitioners. USAID also supports these endeavors. In 2021, the Agency
introduced Managing Machine Learning Projects in International Development, a practical toolkit for those in
governments or public sectors working with work with data scientists to develop AI.

12. “The Global Partnership on Artificial Intelligence.” GPAI.ai. Accessed June 9, 2023. https://gpai.ai/
13 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

GLOSSARY
Definitions below are derived from multiple USAID studies,131unless specified.

Adoption: Changes that happen when people or Open Data: Refers to data made freely available and
institutions begin to use a new technology and deliberately stored in an easily read data format,
incorporate it into their existing routines or processes. particularly by other computers, and thereby repurposed.
For example, people who use a mobile-money account to
receive remittances and pay bills would be considered Data Privacy: The right of an individual or group to
“adopters,” while those who make a one-time withdrawal maintain control over, and the confidentiality of,
to empty a cash-transfer account would not. information about themselves, especially when that
intrusion results from undue or illegal gathering and use of
Algorithm: A step-by-step procedure to turn any given data about that individual or group.
inputs into useful outputs. A computer algorithm follows a
series of instructions to transform inputs (data) into Data Protection: The practice of ensuring the
outputs that can be used for making decisions, either by protection of data from unauthorized access, use,
the computer system or a human. disclosure, disruption, modification, or destruction, to
provide confidentiality, integrity, and availability.
Internet of Things (IoT): Refers to connected devices
and machines that gather data, connect it with intelligent Digital Economy: The use of digital and Internet
analytics, and adapt their behavior/responses based on the infrastructure by individuals, businesses, and government
information in the communication network. Smartphones to interact with each other, engage in economic activity,
are IoT devices. and access both digital and non-digital goods and services.
A diverse array of technologies and platforms facilitate
Cybersecurity: The prevention of damage to, activity in the digital economy; however, much activity
protection of, and restoration of computers, electronic relies in some measure on the Internet, mobile phones,
communications systems, electronic communications digital data, and digital payments.
services, wire communication, and electronic
communication, including information contained therein, Digital Infrastructure: The foundational components
to ensure its availability, integrity, authentication, that enable digital technologies and services. Examples of
confidentiality, and non-repudiation. digital infrastructure include fiber-optic cables, cell towers,
satellites, data centers, software platforms, and end-user
Cyber Hygiene: The practices and steps that users of devices.
computers and other devices take to maintain system
health and improve online security. These practices are Digital Literacy: The ability to “access, manage,
often part of a routine to ensure the safety of identity and understand, integrate, communicate, evaluate, and create
other details that could be stolen or corrupted.142 information safely and appropriately through digital
devices and networked technologies for participation in
Digital Literacy: The ability to access, manage, economic and social life.”4 This may include competencies
understand, integrate, communicate, evaluate, and create that are variously referred to as computer literacy,
information safely and appropriately through digital information and communications technology (ICT)
technologies for employment, decent jobs, and literacy, information literacy, and media literacy.
entrepreneurship.153
Platform: A group of technologies used as a base upon
Machine Learning: The statistical process of deriving a which other technologies can be built or applications and
rule or pattern from a large body of data to predict future services run. For example, the Internet is a platform that
data. enables web applications and services.

13. “USAID DECA Toolkit.” U.S. Agency for International Development. 2022. https://www.usaid.gov/digital-development/deca-toolkit.
“USAID Digital Strategy.” U.S. Agency for International Development. 2018. https://www.usaid.gov/digital-development/usaid-digital-
strategy.
“USAID Managing Machine Learning Projects in International Development: A Practical Guide.” U.S. Agency for International Development.
n.d. https://www.usaid.gov/digital-development/managing-machine-learning-projects.
14. Chris Brook, “What is Cyber Hygiene? A Definition of Cyber Hygiene, Benefits, Best Practices, and More.” Fortra. May 6, 2023. https://
www.digitalguardian.com/blog/what-cyber-hygiene-definition-cyber-hygiene-benefits-best-practices-and-more.
15. “The Digital Literacy Global Framework (DLGF). 2018. The United Nations Educational, Scientific and Cultural Organization. https://uis.
unesco.org/sites/default/files/documents/ip51-global-framework-reference-digital-literacy-skills-2018-en.pdf.
16. USAID Digital Literacy Primer. U.S. Agency for International Development. 2022. https://www.usaid.gov/digital-development/digital-
literacy-primer.
14 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

ANNEX: ETHICAL AI CASES IN DETAIL

Examples of cases related to the ethics of AI, representing both good practices and negative impacts.

AI DOWNGRADES STUDENTS 50th-percentile male. The most commonly used dummy

In the UK, the students enrollment algorithm favored is five foot nine inches (175cm) tall and weighs 167
students from private schools and affluent areas, leaving pounds (75.74kg) – measurements which are significantly
high-achievers from free, state-schools disproportionately larger than an average woman. This male-skewed data is
affected. Subsequent reviews suggested that the factored into automotive design and is one reason why
algorithms might have been biased (reinforcing prejudices women are 17 percent more likely to be killed and 47
in historical data, plus favoring smaller schools). percent more likely to be injured in car crashes than men.
Additionally, many students had their university places A lack of gender-disaggregated data also affects modern
revoked after the algorithm downgraded students who digital tools. A recent study of common fitness monitors
temporarily could not sit for exams due to the COVID- found that they underestimated steps women take during
19 pandemic. Students took to the streets and the courts housework by up to 74 percent and underestimated
for redress, forcing the government to retract the lower calories burned during housework by as much as 34
grades. Source. percent. Further, women are significantly excluded from
medical research because the findings are not gender
disaggregated. This lack of disaggregated data means that
FACING PRIVACY AND SURVEILLANCE
half of the population is misrepresented or does not have
ISSUES WHILE EMBEDDING AI IN equal access to public or private services and
EDUCATION technologies. Source.
Artificial intelligence can help students get useful feedback
faster and, among other things, reduce the burden on
AMAZON AI AGAINST WOMEN
teachers. Artificial intelligence systems can track how
Amazon abandoned its AI hiring program because of its
users interacts with digital tools; the resulting experience
bias against women. The algorithm began training on the
provides a personalized experience. In education, this may
resumes of the candidates for jobs posted over the
include systems that identify strengths and weaknesses
previous ten years. Because most of the applicants were
and patterns in student performance. While teachers do
men, it developed a bias to prefer men and penalized
this to some extent in their teaching, monitoring and
features associated with women. The program failed to
tracking online conversations and student actions may
remove the gender bias as it was profoundly embedded in
also limit student participation, as students feel
the datasets. Source.
uncomfortable knowing their activities and responses are
being monitored and analyzed by AI. Source.
POLICE AI ACTS AGAINST BLACK
FLAWED FACIAL RECOGNITION PEOPLE
A study of AI tools that U.S. authorities use to determine
SYSTEMS
the likelihood that a criminal reoffends found that
Google is one of the leaders in AI, however, its facial
algorithms produced different results for Black and white
recognition software has shown bias against African
people under the same conditions, with Black people
Americans. In several stated cases, image recognition
being defined as more likely to reoffend. This
algorithms in Google Photos were classifying African
controversial use of AI for “predictive” policing, has led to
Americans as “gorillas.” Google has not yet found a
sharp criticism of law enforcement’s use of AI. Source.
solution for this error, instead implementing a
workaround in which it erased gorillas and several other
primates from the platform’s lexicon to prevent further RACIST AI BOT ON TWITTER
misclassification. Source. In 2016, Microsoft launched an AI bot on Twitter that
could interact and learn from users of the social media
platform. However, it exhibited racist and sexist
LACK OF GENDER-DISAGGREGATED
behaviors a few hours after learning from open data on
DATA LEADS TO FATAL OUTCOMES Twitter. Microsoft shut it down less than a day after its
Crash-test dummies were first introduced in the 1950s,
release. Source.
and for decades they were modeled on data for the
15 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

AI-ENABLED TRANSLATION SERVICES AI “PREDICTED” TEEN PREGNANCY

ARE BIASED TOWARDS MALES In 2018, the Ministry of Early Childhood in the
Google Translate systematically changes the gender of Argentinian province of Salta partnered with Microsoft to
translations in a way that favors males. Stereotypes sneak pilot an algorithmic system to predict teenage pregnancy.
into translations because Google optimizes translations They called it the Technology Platform for Social
for English. Experiments showed that, in many cases, Intervention. The goal of the pilot was to forecast which
Google changes the gender of the word; for instance, the females from low-income areas would become pregnant
German phrase “vier Historikerinnen und Historiker” within the next five years. The implications upon women
(four male and female historians) is rendered as “cuatro and girls of being declared “predestined” for motherhood,
historiadores” (four male historians) in Spanish, meaning or how knowing this information would then help prevent
that female historians were removed from the text. adolescent conception, were not acknowledged or
Google Translate provided similar results in Italian, French publicly discussed by the ministry or Microsoft. The
and Polish. Source. system was based on analyzing data—including age,
ethnicity, country of origin, disability, and whether the
subject’s home had hot water in the bathroom—from
AI AGAINST PEOPLE WITH DISABILITIES
200,000 residents in the city of Salta, included 12,000
There are examples of unintentional discrimination
women and girls between the ages of 10 and 19. The
against persons with disabilities by AI applications.
Technology Platform for Social Intervention was never
Persons with disabilities may be interpreted as outliers by
the subject of a formal review, and its effects on girls and
an AI application, mimicking the direct and indirect
women have not been examined since due to the absence
discrimination these persons face in society. For example,
of national AI legislation. No formal information about the
AI systems programmed on past employee data may
platform’s accuracy or results has been released since.
interpret a disclosed disability as a negative characteristic
However, it has since been discovered that the system’s
if past applicants with disabilities were frequently
database only contained information on racial and
screened out at early stages. Source.
socioeconomic groups and did not have information on
access to sex education or contraception, which are
widely acknowledged by public health organizations as the
most effective methods for lowering the rate of teen
pregnancy. Source.

AI Ethics Solutions

GERMAN DATA ETHICS COMMISSION GUIDING PRINCIPLES FROM CANADA

In Germany, a Data Ethics Commission was established to Canada is exemplary in its deployment of tools for public
produce ethical benchmarks, guidelines, and officials that help them explore AI in ways that are
recommendations for the development and use of AI for “governed by clear values, ethics and laws.” The Canadian
the government. The result of their work was a approach is comprehensive: it provides Guiding Principles
publication of ethical guidelines with a set of Specific to ensure ethical use of AI, a list of businesses looking to
Recommendations for Action, with the intention of sell AI solutions to the government, and an algorithm
“protecting the individual, preserving social cohesion, and impact. The latter helps government bodies assess the
safeguarding and promoting prosperity in the information potential risks of deploying an automated decision-making
age.” Notably, in Germany, many private organizations system. It is presented in the form of an 80-point
establish their own overarching AI guidelines. For questionnaire related to business process, data, system
instance, Deutsche Telekom established Guidelines for design, algorithm, and system design decisions. The results
Artificial Intelligence that describe how AI at Deutsche provided by the assessment inform the body around the
Telekom should be used and how AI-based products potential impact of the proposed AI and provide
should be developed. information about applicable requirements.
16 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

AI LAWS IN THE USA

The National Artificial Intelligence Initiative was adopted in 2020.175However, this initiative has not defined bias or
mentioned gender bias as of this review. A new Blueprint for an AI Bill of Rights from the Biden-Harris Administration has
introduced five protections necessary for individuals in the AI age. Specifically, the Algorithmic Discrimination Protection
declares, “You should not face discrimination by algorithms, and systems should be used and designed equitably.”186In
October 2022, the Law on Artificial Intelligence Training for the Acquisition Workforce Act (also known as the “AI
Training Act”)197was adopted to provide an AI training program for the acquired workforces of executive agencies. The
purpose of the program is to ensure the public workforce has knowledge of the capabilities and risks associated with AI.
Meanwhile, several federal and state government agencies and private sector actors have launched initiatives to prevent
algorithmic bias. An industry initiative led by the Data & Trust Alliance has developed Algorithmic Bias Safeguards for the
Workforce, a structured questionnaire that businesses can use for procuring software to evaluate workers.

OPEN EXPLANATIONS OF AI FROM SINGAPORE

UCARE.AI, a Singapore-based startup, offers an AI-powered “cost predictor” product on its platform that works with
hospitals to deliver accurate estimations of hospital bills to patients. To build greater confidence and trust in the use of AI,
the company was mindful to be transparent in its use of AI with various stakeholders. UCARE.AI not only disclosed the
exact parameters used in developing the AI model to its clients, but also provided detailed explanations on all algorithms
that had any impact on operations, revenue, or customer base. Realizing that the accuracy of bill projection is highly
regarded by hospitals and patients, UCARE.AI made a conscious decision to declare the use of AI in its analysis and
prediction of bill amounts to client data managers and its patients. Source.

AI GUIDE FOR PUBLIC AUTHORITIES IN THE UK

The UK Government developed a guide to using AI in the public sector to better understand how AI works and how it
can be applied in the public sector. A final illustrative document from the government is the Guidelines for AI
Procurement that provide a set of principles on how to vet and buy AI technology, as well as insights on tackling
challenges that may arise during procurement.

AI ETHICS—A JOINT EFFORT

Recognizing that issues relevant to AI transcend borders, countries are increasingly adopting regional approaches to AI,
including coordinated efforts in the European Union and the African Union, among Nordic-Baltic states and Arab nations,
and within the G7 and the G20. The OECD has also strengthened its AI-related efforts in recent years, spearheaded by
the OECD.AI Policy Observatory. Indeed, the OECD AI Principles adopted in 2019 are the first intergovernmental
standards on AI. The OECD created its guidelines that provide a list of overarching principles and policy
recommendations with an aim to guide governments, organizations, and individuals in designing and running AI systems in
a way that puts people’s best interests first and ensuring that designers and operators are held accountable for their
proper functioning. Source.

17. “National Artificial Intelligence Initiative Act of 2020.” 2020. Congress.gov. https://www.congress.gov/bill/116th-congress/house-bill/6216.
18. “Algorithmic Discrimination Protections.” 2020. The White House https://www.whitehouse.gov/ostp/ai-bill-of-rights/algorithmic-
discrimination-protections-2/.
19. “Artificial Intelligence Training for the Acquisition Workforce Act or the AI Training Act.” 2022. Congress.gov. https://www.congress.gov/
bill/117th-congress/senate-bill/2551.
17 ARTIFICIAL INTELLIGENCE (AI) ETHICS GUIDE

U.S. Agency for International Development

1300 Pennsylvania Avenue, NW
Washington, DC 20523
Tel: (202) 712-0000
www.usaid.gov