Notes From The Ai Frontier: Applying Ai For Social Good

NOTES FROM
THE AI FRONTIER
APPLYING AI FOR
SOCIAL GOOD
DISCUSSION PAPER
DECEMBER 2018
Michael Chui | San Francisco

Martin Harryson | Silicon Valley
James Manyika | San Francisco
Roger Roberts | Silicon Valley
Rita Chung | Silicon Valley
Ashley van Heteren | Amsterdam
Pieter Nel | New York
Since its founding in 1990, the McKinsey Global Institute (MGI) has sought
to develop a deeper understanding of the evolving global economy. As the
business and economics research arm of McKinsey & Company, MGI aims
to provide leaders in the commercial, public, and social sectors with the facts
and insights on which to base management and policy decisions.
MGI research combines the disciplines of economics and management,

employing the analytical tools of economics with the insights of business
leaders. Our “micro-to-macro” methodology examines microeconomic
industry trends to better understand the broad macroeconomic forces
affecting business strategy and public policy. MGI’s in-depth reports have
covered more than 20 countries and 30 industries. Current research focuses
on six themes: productivity and growth, natural resources, labor markets,
the evolution of global financial markets, the economic impact of technology
and innovation, and urbanization. Recent reports have assessed the
digital economy, the impact of AI and automation on employment, income
inequality, the productivity puzzle, the economic benefits of tackling gender
inequality, a new era of global competition, Chinese innovation, and digital and
financial globalization.
MGI is led by three McKinsey & Company senior partners: Jacques Bughin,

Jonathan Woetzel, and James Manyika, who also serves as the chairman
of MGI. Michael Chui, Susan Lund, Anu Madgavkar, Jan Mischke,
Sree Ramaswamy, and Jaana Remes are MGI partners, and Mekala Krishnan
and Jeongmin Seong are MGI senior fellows.
Project teams are led by the MGI partners and a group of senior fellows,
and include consultants from McKinsey offices around the world. These
teams draw on McKinsey’s global network of partners and industry and
management experts. Advice and input to MGI research are provided by the
MGI Council, members of which are also involved in MGI’s research. MGI
Council members are drawn from around the world and from various sectors
and include Andrés Cadena, Sandrine Devillard, Tarek Elmasry, Katy George,
Rajat Gupta, Eric Hazan, Acha Leke, Scott Nyquist, Gary Pinkus, Sven Smit,
Oliver Tonby, and Eckart Windhagen. In addition, leading economists,
including Nobel laureates, act as research advisers to MGI research.
The partners of McKinsey fund MGI’s research; it is not commissioned by any

business, government, or other institution. For further information about MGI
and to download reports, please visit www.mckinsey.com/mgi.
Copyright © McKinsey & Company 2018
2 McKinsey Global Institute

SUMMARY
WHAT’S INSIDE
NOTES FROM THE AI 1. Mapping AI use cases to
FRONTIER: APPLYING AI domains of social good
FOR SOCIAL GOOD Page 1
2. How AI capabilities can be

Artificial intelligence, while not a silver bullet,
used for societal benefit
could contribute to the multi-pronged efforts
Page 10
to tackle some of the world’s most challenging
social problems. AI is already being leveraged in 3. Six illustrative use cases
research to tackle societal “moon shot” challenges Page 18
such as curing cancer and climate science. The
focus of this paper is on other social benefit uses 4. Bottlenecks to overcome
of AI that do not require scientific breakthroughs Page 30
but that add to existing efforts to help individuals
or groups in both advanced and developing 5. Risks to be managed
economies who are experiencing challenges or Page 35
crises and who often live beyond the reach of
6. Scaling up the use of AI for
traditional or commercial solutions. We assess the
social good
AI capabilities that are currently most applicable
Page 42
for such challenges and identify domains where
their deployment would be most powerful. We also
identify limiting factors and risks to be addressed
and mitigated if the social impact potential is to
be realized.
Through an analysis of about 160 AI social impact use cases, we have identified and
characterized ten domains where adding AI to the solution mix could have large-scale
social impact. These range across all 17 of the United Nations Sustainable Development
Goals and could potentially help hundreds of millions of people worldwide. Real-life
examples show AI already being applied to some degree in about one-third of these use
cases, ranging from helping blind people navigate their surroundings to aiding disaster
relief efforts.
Several AI capabilities, primarily in the categories of computer vision and natural

language processing, are especially applicable to a wide range of societal challenges.
As in the commercial sector, these capabilities are good at recognizing patterns from
the types of data they use, particularly unstructured data rich in information, such as
images, video, and text, and they are particularly effective at completing classification
and prediction tasks. Structured deep learning, which applies deep learning techniques
to traditional tabular data, is a third AI capability that has broad potential uses for social
good. Deep learning applied to structured data can provide advantages over other
analytical techniques because it can automate basic feature engineering and can be
applied despite lower levels of domain expertise.
These AI capabilities are especially pertinent in four large domains—health and hunger,
education, security and justice, and equality and inclusion—where the potential usage
frequency is high and where typically a large target population would be affected. In
health, for example, AI-enabled wearable devices, which can already detect potential
early signs of diabetes through heart rate sensor data with 85 percent accuracy, could
potentially contribute to helping more than 400 million people afflicted by the disease
worldwide if made sufficiently affordable. In education, more than 1.5 billion students
could benefit from application of adaptive learning technology, which tailors content to
students based on their abilities.
Scaling up AI usage for social good will require overcoming some significant bottlenecks,
especially around data accessibility and talent. In many cases, sensitive or monetizable
data that could have societal applications are privately owned, or only available in
commercial contexts where they have business value and must be purchased, and
are not readily accessible to social or nongovernmental organizations. In other cases,
bureaucratic inertia keeps data that could be used to enable solutions locked up, for
example in government agencies. In most cases, the needed data have not been
collected. Talent with high-level AI expertise able to improve upon AI capabilities and
develop models is in short supply, at a time when competition for it from the for-profit
sector is fierce. Deployment also often faces “last mile” implementation challenges
even where data and technology maturity challenges are solved. While some of these
challenges are nontechnical and common to most social good endeavors, others are
tech-related: NGOs may lack the data scientists and translators needed to address the
problem and interpret results and output from AI models accurately.
Large-scale use of AI for social good entails risks that will need to be mitigated, and
some tradeoffs to be made, to avoid hurting the very individuals the AI application was
intended to help. AI’s tools and techniques can be misused by authorities and others
with access to them, and principles for their use will need to be established. Bias may be
embedded in AI models or data sets that could amplify existing inequalities. Data privacy
will need to be protected to prevent sensitive personal information from being made
public and to comply with the law, and AI applications will need to be safe for human
use. The continuing difficulty of making some AI-produced decisions transparent and
explainable could also hamper its acceptance and use, especially for sensitive topics
such as criminal justice. Solutions being developed to improve accuracy, including
model validation techniques and “human in the loop” quality checks, could address
some of these risks and concerns.
Stakeholders from both the private and public sector have essential roles to play in
ensuring that AI can achieve its potential for social good. Collectors and generators of
data, whether governments or companies, could grant greater access to NGOs and
others seeking to use the data for public service and could potentially be mandated to
do so in certain cases. To resolve implementation issues will require many more data
scientists or those with AI experience to help deploy AI solutions at scale. Capability
building, including that funded through philanthropy, can help: talent shortages at this
level can be overcome with a focus on accessible education opportunities such as online
courses and freely available guides, as well as contributions of time by organizations
such as technology companies that employ highly skilled AI talent. Indeed, finding
solutions that apply AI to specific societal goals could be accelerated if technology
players dedicated some of their resources and encouraged their AI experts to take on
projects that benefit the common good.
The application of AI for societal benefit is an emerging topic and many research questions
and issues remain unanswered. Our library of use cases is evolving and not comprehensive;
while we expect to build on it, data about technological innovations and their potential
applications are incomplete. Our hope is that this paper sparks further discussion about
where AI capabilities can be applied for social good, and scaled up, so that their full societal
potential can be realized.
1. MAPPING AI USE CASES TO DOMAINS OF
SOCIAL GOOD
Artificial intelligence (AI), which for the purposes of this paper we use as shorthand to refer
specifically to deep learning techniques, is increasingly moving out of the research lab and
into the world of business.1 Beyond its commercial uses, now increasingly widespread in
mobile and other consumer applications, AI has noncommercial potential to do good. While
AI is not a silver bullet or cure-all, the technology’s powerful capabilities could be harnessed
and added to the mix of approaches to address some of the biggest challenges of our age,
from hunger and disease to climate change and disaster relief.2
Examples of where some of these capabilities are already being deployed illustrate how
broad AI’s impact could be. To cite just three: Planet Labs, an Earth-imaging Silicon Valley
startup, partnered with Paul G. Allen Philanthropies and leading research scientists to
create a global map of shallow-water coral reefs by applying object detection to satellite
imagery in correlation with geospatial data. This map is used to monitor change over time
and inform conservation interventions for the reef ecosystems that are under threat.3 At
Thorn, an international anti–human trafficking nonprofit organization, a combination of face
detection and person identification, social network analysis, natural language processing,
and analytics is being used to identify victims of sexual exploitation on the internet and dark
web. Thorn works in collaboration with a group of technology companies, including Google,
Microsoft, and Facebook, and has found a total of 5,791 child victims since 2016.4 AI is also
being used in the battle against cancer: researchers at the MIT Media Lab, for example,
have applied reinforcement learning, a capability in which systems essentially learn by trial
and error, in clinical trials with patients diagnosed with glioblastoma (the most aggressive
form of brain cancer) to successfully reduce toxic chemotherapy and radiotherapy dosing.
This example is particularly exciting as it shows capabilities still in development being
applied to social good use cases; reducing chemotherapy doses helps improve quality of
life of cancer patients and reduce the cost of their treatment. As further research continues
to improve reinforcement learning, the practical applications of the solutions will extend
beyond clinical trials to customization of patient treatment.5
In all, we have collected about 160 such social good use cases to date. They touch on
some aspect of all 17 of the United Nations Sustainable Development Goals and potentially
could help hundreds of millions of people worldwide. This use case library, which continues
to grow and evolve, provides the basis for an in-depth examination of the domains where
AI could be used and the applications that are likely to be the most impactful, as well as
bottlenecks to impact and risks that will need to be addressed.
1
We recognize that the line of demarcation between artificial intelligence capabilities and other analytical
capabilities is not universally shared, with different people holding different definitions, over time. In our use
case library, we did estimate the potential for other analytical capabilities, including for the use of machine
learning, as described in the Flint, Michigan, case in Chapter 3. Our use of “deep learning” refers to machine
learning techniques on very large artificial (simulated) neural networks.
2
AI capabilities can be used for bad or malicious purposes as well as for social good. For a discussion of the
ethics of AI, see Box 2, on page 36.
3
Andrew Zolli, Planet, Paul G Allen Philanthropies, & leading scientists team up to map & monitor world’s corals
in unprecedented detail, Planet, June 4, 2018, planet.com/pulse/planet-paul-g-allen-coral-map/.
4
Thorn’s user surveys indicate that in the past two years, its “Spotlight” tool was used in 21,044 cases and
identified 6,553 traffickers. wearethorn.org/spotlight/.
5
Rob Matheson, “Artificial intelligence model ‘learns’ from patient data to make cancer treatment less toxic,”
MIT News, August 9, 2018, news.mit.edu/2018/artificial-intelligence-model-learns-patient-data-cancer-
treatment-less-toxic-0810.
McKinsey Global Institute Notes from the AI frontier: Applying AI for social good 1
AI’S POTENTIAL SOCIETAL IMPACT IS BROAD, BASED ON OUR MAPPING OF
USE CASES TO DOMAINS
To build our library of use cases, which forms the basis of our analysis, we adopted a dual
approach, both societal and technological (Exhibit 1). Each use case highlights a type
of meaningful problem that can be solved by an AI capability or some combination of AI
capabilities. To measure the relative potential of AI we used usage frequency as a proxy
(see Box 1, “Building a library of AI use cases for social good to understand comparative
relevance of AI across domains”). For about one-third of the use cases in our library to
date, we identified an actual AI deployment in some form (Exhibit 2). Since many of these
solutions are small test cases to determine feasibility, their functionality and scope of
deployment often suggest that additional potential could be captured. For three-quarters of
our use cases, we have seen deployment of solutions that employ some level of advanced
analytics; most of these use cases, although not all, would further benefit from the use of
AI applications.
Exhibit 1
Mapping domains to issue types and use cases in our library.
Number of use cases per domain
Crisis response
▪ Disease outbreak
▪ Migration crises
▪ Natural and man-made
disasters
Security and justice ▪ Search and rescue
▪ Harm prevention Economic empowerment
▪ Fair prosecution ▪ Agricultural quality and yield
▪ Policing ▪ Financial inclusion
17 ▪ Initiatives for economic growth
Public and social sector ▪ Labor supply and demand matching
▪ Effective management 16 15
of public sector Education
▪ Effective management ▪ Access and completion of
of social sector education
▪ Fundraising ▪ Maximizing student
16 13 achievement
▪ Public finance
management ▪ Teacher and administration
▪ Services to citizens productivity
Infrastructure Environment
▪ Energy ▪ Animal and plant
▪ Real estate 15 21 conservation
▪ Transportation ▪ Climate change
▪ Urban planning and adaptation
▪ Water and waste ▪ Energy efficiency
management 4 11 and sustainability
Information ▪ Land, air, and
verification 28 water conservation
and validation ▪ Equality and inclusion
▪ False news ▪ Accessibility and disabilities
▪ Polarization ▪ Exploitation
Health and hunger ▪ Marginalized communities
▪ Treatment delivery
▪ Prediction and prevention
▪ Treatment and long-term
care
▪ Mental wellness
▪ Hunger
SOURCE: McKinsey Global Institute analysis
2 McKinsey Global Institute Notes from the AI frontier: Applying AI for social good
Exhibit 2
About one-third use cases in our library have been deployed in some form, USE CASE LIBRARY
leveraging AI capabilities. NOT EXHAUSTIVE
Number of use cases Number of use cases where Number of use cases with
where some form of AI only analytics has been deployed AI potential and no known
has been deployed but AI potential exists AI or analytics deployment
Social
impact Use case profile breakdown
domain per domain Remarks
Crisis 7 10 This domain has high potential for AI use, although problems can be
17
response complex and developing the right AI solutions may take time.
Existing use cases deploying AI typically related to agriculture;

Economic commercial market has supported AI solutions that could be adapted for
5 8 2 15
empowerment societal good use cases. Many existing use cases that use analytics could
benefit from structured deep learning, which is greatly underused today.
Most existing AI use cases employ natural language processing (NLP).

Education 5 3 5 13
For now, adaptive learning only leverages analytics.
Research institutions and organizations working on AI use for social

Environment 12 4 5 21
causes have supported AI deployment.
Equality and Many solutions in this domain rely on AI. Most use NLP and computer
8 2 1 11
inclusion vision. Much room remains to raise the quality of solutions.
Health and Existing cases that use AI are mainly focused on medical diagnoses, and
10 8 10 28
hunger deployment is not yet at scale.
Information
High potential use of AI, but problem is complex and development of
verification 31 4
appropriate solutions may take time.
and validation
Most use cases in our library either do not have known case studies or
Infrastructure 1 9 5 15 use only analytics; type of problem to solve and data used mainly revolve
around optimization using structured data.
All use cases in this domain in our library have existing case studies and
Public and
16 use only analytics, though NLP and structured deep learning would likely
social sector
add value.
Many potential AI solutions in this domain have not been implemented

Security and
3 10 3 16 because of fear of negative repercussions. Existing use cases largely
justice
leverage analytics.
NOTE: Our library of about 160 use cases with societal impact is evolving and this chart should not be read as a comprehensive gauge of the potential
application of AI or analytics capabilities.
Box 1. Building a library of use cases for social good to understand the comparative relevance of AI
across domains
Our library of use cases, which forms the basis of our issues can theoretically affect all seven billion people
analysis, currently has about 160 use cases in ten social on the planet. However imperfect, the proxy of potential
impact domains. To build the library, we adopted a usage frequency of AI allows for comparisons between
two-pronged approach, both societal and technological use cases individually or at an aggregate level across
(Exhibit 3). all domains, in terms of comparative magnitude of AI
usefulness and impact.
Each use case highlights a type of meaningful problem
that can be solved by an AI capability or some To calculate AI usage frequency, we estimated the
combination of AI capabilities. The problem to solve was number of times that models trained using AI would be
given a broad enough definition so that similar solutions used in a year to predict an outcome. This quantitative
would be grouped together. Most domains have around approach provided a directional means to identify AI
15 use cases, with two outlier domains of health and capabilities with higher potential to bring about social
hunger (28 use cases) and information verification and impact, and others where AI deployment would be useful
validation (four use cases). but not as impactful. (A criterion for our use cases is that
they reach a threshold of having “meaningful” societal
From a societal point of view, we sought to identify key
value potential, as agreed by domain experts.)
problems that are known to the social-sector community
and determine where AI could aid efforts to resolve them. AI usage frequency takes into account the number of
From a technological point of view, we took a curated list individual cases for which a model would need to be run
of 18 AI capabilities and sought to identify which types of and how often the model would be run. This could be the
social problems they could best contribute to solving. number of lives affected in a use case and how often per
year a model would be run on each individual.
For each use case, we tried to identify at least one
existing case study. Where none were identified, we For example, in a use case on predicting students at risk
worked iteratively with experts to identify gaps and of dropping out of school, the base is the number of K-12
added corresponding use cases to our library. To guide students worldwide; the model in this case has to be run
our thinking, we conducted interviews with some 80 separately for each individual student approximately once
AI practitioners, social entrepreneurs, domain experts, per month to predict the likelihood that they will drop out.
academics, and technology company executives. For an AI solution that uses a combination of capabilities
including image classification, object detection, OCR,
The library is not comprehensive, but it nonetheless
and emotion recognition to narrate the environment for
showcases a wide range of problems where AI can be
the visually impaired, the base is the number of visually
applied for social good. As AI capabilities evolve and
impaired people globally, and we estimate that it would
as technical and social impact practitioners continue to
be run nearly continuously, that is, once per minute for 16
identify more ways in which these capabilities can be
active hours a day. The base number of individuals is not
applied, we expect the library to grow.
always the number of people. One example is from a use
Measuring usage frequency case where drones are used to detect poacher activity.
To provide a rough (and admittedly imperfect) measure Here, the metric we use is the 307 wildlife sanctuaries in
of the relative potential of AI, we employed usage the world, and we estimate that the system would be run
frequency as a proxy for societal value. Unlike AI usage once a minute for 12 hours a day (at night) when poachers
for commercial purposes, where the value is typically are active.
measured in dollars, social value is harder to measure
Use cases can take in various data types and are often
across all domains and use cases using one metric.
associated with more than one AI capability. We found
The cost of human suffering, whatever the cause, and
that these vary significantly across social impact domains,
the benefits of alleviating it, are impossible to precisely
based on our library. The heat map in Chapter 2 of
gauge and compare. Even comparisons using number
this discussion paper focusing on AI usage frequency
of lives affected can quickly become meaningless
demonstrates this variety, highlighting the intersection of
both across domains and within them; for example,
the domain or issue type and the specific AI capability.
use cases that contribute to solving climate change
Box 1. Building a library of use cases for social good to understand the comparative relevance of AI
across domains (continued)
Exhibit 3
We built a library of use cases of AI for societal good using both social-first and tech-first approaches.
Social-first approach
Social problems AI capabilities
Mapping
Evaluating
capabilities to Perspective
Impact domains Use cases potential barriers
domains and in this paper
and risks
use cases
AI capabilities Social problems
Tech-first approach
We grouped use cases into ten social impact domains based on examining and integrating
taxonomies used by social-sector organizations, such as the AI for Good Foundation and
the World Bank. Use cases within each domain are further grouped into two to five issue
types.6 The following is the list of the social impact domains we examined.
Crisis response. Specific crisis-related challenges, such as responding to natural and

man-made disasters in search and rescue missions and at times of disease outbreak.
Examples of use cases with high potential usage frequency include using AI on satellite
data to map and predict wildfire progression to optimize firefighter response. Drones with
AI capabilities can also be used to find missing persons in wilderness areas.
Economic empowerment. Opening access to economic resources and opportunities,

Box 1
including jobs, skills development, and market information, with an emphasis on
currently vulnerable populations. For example, AI can be used for early detection of plant
damage through low-altitude sensors, including smartphones and drones, to improve
yield in small farms if farmers have access to technology; one project called FarmBeats
is building edge-computing technology that could one day make data-driven farming
accessible for even the poorest farmers.7
6
AI for Good Foundation, ai4good.org/active-projects/. In July 2016, the World Bank introduced a new
taxonomy of theme codes, projects.worldbank.org/theme.
7
GatesNotes, “Can the Wi-Fi chip in your phone help feed the world?”, blog entry by Bill Gates, October 9,
2018, gatesnotes.com/Development/FarmBeats.
Educational challenges. These include maximizing student achievement and
improving teacher productivity. For example, adaptive learning technology could be
used to recommend content to students based on past success and engagement with
the material. AI could also be used to detect student distress early, before a teacher
has noticed.
Environmental challenges. These include sustaining biodiversity and combating

natural resource depletion, pollution, and climate change. For example, robots with
AI capabilities can be used to sort recyclable material from waste. The Rainforest
Connection, a Bay Area nonprofit, uses AI tools such as Google’s TensorFlow in
conservation efforts across the world. Its platform can detect illegal logging in vulnerable
forest areas through analysis of audio sensor data.8 Other applications include using
satellite imagery to predict routes and behavior of illegal fishing vessels.9
Equality and inclusion. Addressing equality, inclusion, and self-determination

challenges, such as reducing or eliminating bias based on race, sexual orientation,
religion, citizenship, and disabilities. One use case, based on work by Affectiva, which
was spun out of the MIT Media Lab, and Autism Glass, a Stanford research project,
involves use of AI to automate emotion recognition and provide social cues to help
individuals along the autism spectrum interact in social environments.10 Another example
is the creation of an alternative identification verification system for individuals without
traditional forms of ID, such as driver’s licenses.
Health and hunger. Addressing health and hunger challenges, including early-
stage diagnosis and optimized food distribution. Researchers at the University of
Heidelberg and Stanford University have created a disease detection AI system, using
visual diagnosis of natural images such as images of skin lesions to determine if they
are cancerous; the system outperformed professional dermatologists.11 AI-enabled
wearable devices, which can already detect potential early signs of diabetes through
heart rate sensor data with 85 percent accuracy, could potentially help more than
400 million people worldwide afflicted by the disease if the devices could be made
sufficiently affordable.12 Other use cases include combining various types of alternative
data sources such as geospatial data, social media data, telecommunications data,
online search data, and vaccination data to help predict virus and disease transmission
patterns, or using an AI solution to optimize food distribution networks in areas facing
shortages and famine.
Information verification and validation. The challenge of facilitating provision,

validation, and recommendation of helpful, valuable, and reliable information to all. This
domain differs from the others in that it focuses on filtering or counteracting content
that could mislead and distort, including false and polarizing information disseminated
through the relatively new channels of the internet and social media. Such content
can have severely negative consequences, including the manipulation of election
results and the mob killings in India and Mexico that have been triggered by false news
8
“What have we done so far?” Rainforest Connection, rfcx.org/home.
9
“Using satellite imagery to combat illegal fishing,” The Maritime Executive, July 17, 2017.
10
affectiva.com and autismglass.stanford.edu. See also, David Talbot, “Digital summit: First emotion-
reading apps for kids with autism,” MIT Technology Review, June 9, 2014, https://www.technologyreview.
com/s/528191/digital-summit-first-emotion-reading-apps-for-kids-with-autism/.
11
“Computer learns to detect skin cancer more accurately than doctors,” Guardian, May 29, 2018.
12
Brandon Ballinger et al., DeepHeart: Semi-supervised sequence learning for cardiovascular risk prediction,
32nd AAAI Conference on Artificial Intelligence, New Orleans, LA, February 2–7, 2018, aaai.org/ocs/index.
php/AAAI/AAAI18/paper/view/16967/15916.
dissemination via messaging applications.13 Use cases in this domain include actively
presenting opposing views to ideologically isolated pockets in social media. Problems in
this domain are both in the public interest and technically very complex. The main actors
able to address them are most likely to be tech companies that own these platforms
and technologies.
Infrastructure management. Infrastructure challenges that could provide public

good in the categories of energy, water and waste management, transportation,
real estate, and urban planning. For example, traffic light networks can be optimized
using real-time traffic camera data and Internet of Things sensors to maximize vehicle
throughput. AI can also be used for predictive maintenance of public transportation
systems such as trains and public infrastructure, including bridges, to identify potentially
malfunctioning components.
Public and social sector management. Initiatives that are related to the efficiency and
effective management of public- and social-sector entities, including strong institutions,
transparency, and financial management. For example, AI can be used to identify tax
fraud using alternative data such as browsing data, retail data, and payments history.
Another instance where AI can prove effective is in providing automated question
answering via email to improve government interaction with citizens.
Security and justice. Challenges in society that include harm prevention—both from
crime and other physical dangers—as well as tracking criminals and mitigating bias of
police forces. This domain focuses on security, policing, and criminal justice issues as
a unique category adjacent to public-sector management. An example is using AI to
create solutions that help firefighters determine safe paths through burning buildings
using data from IoT devices.
THE SOCIAL DOMAINS COVERED BY OUR USE CASES RANGE ACROSS ALL
17 UNITED NATIONS SUSTAINABLE DEVELOPMENT GOALS
The United Nations Sustainable Development Goals (SDGs) are among the best-known
and most frequently cited societal challenges, and our use cases map to all 17 of the goals,
supporting some aspect of each one (Exhibit 4). The SDGs are contained in a 2030 Agenda
for Sustainable Development adopted by all UN member states in 2015 which lays out
strategies for ending poverty and other deprivations at the same time as improving health
and education, reducing inequality, preserving the environment, and boosting economic
growth, among other priorities.
Our use case library does not use the same taxonomy as the SDGs because their goals are
not directly related to AI usage, unlike ours; about 20 of the cases in our library do not map
to the SDGs at all.14
13
Vindu Goel, Suhasini Raj, and Priyadarshini Ravichandran, “How WhatsApp leads mobs to murder in India,”
New York Times, July 18, 2018; Patrick J. McDonnell and Cecilia Sanchez, “When fake news kills: Lynchings
in Mexico are linked to viral child-kidnap rumors,” Los Angeles Times, September 21, 2018.
14
For details of the 17 UN Sustainable Development Goals, see sustainabledevelopment.un.org/.
Exhibit 4
Use cases in our library support the UN Sustainable Development Goals.
UN Sustainable Development Goals1

Number of use cases
UN goals AI use case breakdown
Life below water 1

Security and justice Crisis response
Affordable and clean energy Public and Economic 2

social sector empowerment
Clean water and sanitation AI 2
Infrastructure use case Education
key
Responsible consumption and production Info verification 3
Environment
and validation
Sustainable cities and communities 3
Health and hunger Equality and inclusion
Gender equality 3
Partnerships for the goals 4
Zero hunger 4
Decent work and economic growth 5
Climate action 6
Reduced inequalities 7
Industry, innovation, and infrastructure 9
No poverty 10
Life on land 10
Quality education 13
Peace, justice, and strong institutions 24
Good health and well-being 29
1 Chart is a partial list of use cases, as 21 of the 156 identified use cases do not target any of the UN’s Sustainable Development Goals.
NOTE: This chart reflects the number and distribution of use cases and should not be read as a comprehensive evaluation of AI potential for each SDG; if an
SDG has a low number of cases, that is a reflection of our library rather than of AI applicability to that SDG. The chart also does not reflect all use cases in
the library, more than 20 of which do not map to any SDGs. These mainly focus on effective management in the public and social sectors, or belong to the
issue types of disaster response and search and rescue in the crisis response domain.
SOURCE: United Nations; McKinsey Global Institute analysis
STAKEHOLDERS INCLUDING PUBLIC AND SOCIAL SECTOR ORGANIZATIONS
ARE BUILDING THE MOMENTUM FOR SOCIAL BENEFIT APPLICATIONS
While AI solutions are not a first nor an obvious choice for many social-sector organizations,
a variety of stakeholders in the social and public sectors, in universities, and in a number
of commercial firms are starting to focus on the technologies in their work for social good.
Two US-based NGOs already cited are examples: Thorn, co-founded by Hollywood stars
Ashton Kutcher and Demi Moore, uses a range of computer vision and natural language
processing (NLP) capabilities alongside analytics to identify and rescue victims of sexual
trafficking, while Rainforest Connection uses sound detection from audio sensors to detect
illegal logging in rainforests and then sends real-time alerts to partners on the ground.
In Chapter 3, we describe six case studies in more detail. They show how a variety of
organizations and academic institutions are taking the first steps to use AI solutions for
social good. The organizations are not necessarily noncommercial; some are commercial
companies either working on social good causes or using solutions that can have
social value.
Beyond the examples we describe, our research found several organizations with AI
expertise that are frequently cited in the sector for their work on applying AI to social good
issues. They include AI4All, the AI for Good Foundation, DataKind, and Data Science for
Social Good.15 We expect that, with the right support from the broader AI for Social Good
ecosystem, more organizations of all types will start to include AI as an option in their toolkit
for solving social problems.
The geography of AI for social good is, for now, also quite narrow. While the use cases in our
library provide coverage of AI being applied to social good around the world, most of the
organizations behind these initiatives are based in the United States.. This likely is a result of
the concentration of high-level AI expertise, and it raises questions of how organizations with
that expertise can better connect to communities around the world. As our library evolves,
we anticipate seeing more use cases where not just deployment but also organization base
and solution development location are in many different regions around the world.
McKinsey & Company supports AI for social good causes through an initiative that sponsors projects and
15
seeks to foster an ecosystem of joint efforts across NGOs, government organizations, technology partners,
and others. Projects supported include ones to combat the spread of measles, help welfare case workers
identify individuals who may need additional support in finding work, and help victims of human trafficking,
including by training algorithms to search for signals in police and investigative databases.
2. HOW AI CAPABILITIES CAN BE USED FOR
SOCIETAL BENEFIT
We identified 18 AI capabilities that could potentially be used for social benefit, of which 14
fall into three major categories: computer vision, natural language processing, and speech
and audio processing. The remaining four, which we treated as stand-alone capabilities,
include reinforcement learning, content generation, and structured deep learning. We also
included a category for other analytics techniques (Exhibit 5). For the purposes of this paper,
we use AI as shorthand specifically to refer to deep learning techniques that use artificial
neural networks. While some machine learning techniques are sometimes referred to as AI,
if they do not use deep learning we group them in this paper in the analytics category.
When we subsequently mapped these capabilities to domains (aggregating use cases)

in a heat map, we found some interesting patterns (Exhibit 6). Making comparisons
horizontally across each row highlights which capabilities could be particularly suitable for
social good use cases. We see that a few are relevant across domains, a few are relevant
to only specific domains, and some are barely used for now. We also made comparisons
vertically to assess domains with regard to the potential use of AI. While many other insights
can be drawn by mapping use cases, domains, and capabilities, we will describe some of
our findings from this cross-domain view. The heat map is not a key from which to derive
absolute impact, but a comparative index of potential and relevance of AI capabilities to
certain domains and their use cases.
COMPUTER VISION, NATURAL LANGUAGE PROCESSING, AND STRUCTURED

DEEP LEARNING HAVE BROAD POTENTIAL APPLICATION ACROSS DOMAINS
Based on AI usage frequency, several AI capabilities within the categories of computer
vision and natural language processing and the capability of structured deep learning had
the most widespread and greatest potential for social good application.
Image classification and object detection are powerful capabilities with

multiple applications for social good
Within computer vision, the specific capabilities of image classification and object detection
stand out for their potential applications for social good. Image classification is an AI
capability that classifies an image or video clip into one of a set of predetermined classes;
for example, image classification can tell you whether the image it reviews contains a cat
or a dog. It can be used to solve problems such as mapping and predicting poverty levels
of different neighborhoods based on nighttime luminosity and daytime imagery. Object
detection is a capability that finds all instances of all trained classes of objects and reports
their locations within an image; for example, object detection can find and highlight all
locations where dogs appear in an image. This can be used to solve problems such as the
detection of fires in satellite imagery.
These capabilities are often used in conjunction with one another. An example is when
drones need computer vision capabilities to navigate a complex forest environment for
search-and-rescue purposes. In this case, image classification may be used to distinguish
normal ground cover from a footpath, guiding the drone’s directional navigation, while object
detection is used to circumvent obstacles such as trees. These capabilities are relevant in
use cases where imagery is available and from which useful information can be extracted to
solve a problem of social good.
Exhibit 5
Some AI capabilities could be used for societal benefit. NOT EXHAUSTIVE

Capability Relative
category maturity Capability Example of problems the AI capability can solve
Computer vision More Person identification (image  Identifying a known missing child through publicly
developed and video) posted pictures and video (commonly referred to as
facial recognition)
Face detection (image and  Detecting the presence of people in surveillance
video) camera footage
Image and video  Identifying endangered animals in image and video
classification for enhanced protection
 Detecting explicit content
Near-duplicate or similar  Detecting hate-speech content for removal of image
detection (images and video) or video
Object detection and  Detecting fires in satellite imagery
localization (images and
video)
Optical character and  Digitizing hard-copy records for quicker patient
handwriting recognition health history search
(OCR, images)
Tracking  Tracking illegal fishing vessels via satellite imagery
Developing Emotion recognition (image  Measuring level of student engagement in
and video) classrooms
Speech and audio More Person identification (speech)  Verifying individuals through mobile phone for
processing developed inclusive banking access based on sound and
pattern of voice
Speech-to-text (audio)  Real-time captioning for the deaf or people hard of
hearing to facilitate live conversation
Sound detection and  Identifying chain-saw sounds in rainforests for alerts
recognition (audio) on illegal logging activities
Developing Emotion recognition (speech)  Assisting individuals on the autism spectrum who
have difficulty in social interactions
Natural language More Person identification (text)  Detecting a paper’s author through handwriting
processing developed analysis and identification of syntax patterns
Language translation (text)  Enabling larger distribution of online education
services to underserved populations
Other natural language  Identifying plagiarism in student assignments to
processing (text) enhance instructor productivity
Developing Sentiment analysis (text)  Using automated review of public sentiment about
specific topics to inform policy
Language understanding  Enabling chatbots that understand abstract
concepts and ambiguous language, eg, ones that
can do second-level, nuanced health screens
Content Developing Content generation  Generating text and media (video, audio) content for
generation educational purposes with quick production
turnaround for wide distribution
Reinforcement Developing Reinforcement learning  Large-scale and high-speed simulation modeling,
learning for example in drug trials, doing millions of
simulations to determine best treatment for breast
cancer in population with a specific genetic makeup
Deep learning on More Structured deep learning  Identifying tax fraud and underreporting of income
structured data developed based on tax return data
Analytics More Analytics  Any analytics technique not involving deep learning,
developed eg, for optimization, journey mapping, network
analysis
Exhibit 6
Mapping usage frequency of AI capabilities to ten social impact domains identifies

patterns of the relevance and applicability of AI for social good. Lower Higher
Usage frequency of AI capability for each domain

Domain  Equality and Health and Info verification Economic
inclusion hunger and validation empowerment Environment
Security and Public and
Capability  Education justice Crisis response social sector Infrastructure
Deep learning on
structured data
Natural language
processing2
Image and video
classification
Object detection and
localization
Language understanding
Sound detection and

recognition
Sentiment analysis
Language translation
Face detection
Tracking
Emotion recognition
Person identification
Optical character and

handwriting recognition
Speech-to-text
Content generation
Reinforcement learning
Near-duplicate or similar
detection
Usage frequency of analytics capability for each domain
Analytics, eg, optimization,
network analysis
1 Log base 10 scale. Deployment frequency capped at once per hour per year to prevent skewing; capping affected only a small number of use cases.
2 Excluding sentiment analysis, speech-to-text, language understanding, and translation.
NOTE: Our library of about 160 use cases with societal impact is evolving and this heatmap should not be read as a comprehensive gauge of the potential
application of AI or analytics capabilities. Usage frequency estimates the number of times that models trained using AI would be used in a year to predict an
outcome.
Some of these use cases consist of tasks individual humans could potentially accomplish,
but where the required number of instances is so large that it exceeds human capacity,
such as finding flooded and unusable roads across a large area after a hurricane. In other
cases, an AI system can perform with greater accuracy than a human (often by processing
more information), for example early identification of plant disease to prevent infection of
the entire crop, which can be devastating to small-scale farmers. In some cases, a machine
may be able to pick up key features of an image where a human would not or could not. One
example is the use of person identification to identify victims of exploitation.
Computer vision capabilities such as person identification, face detection, and emotion
recognition are relevant only in select domains and use cases, including for crisis response,
security, equality, and education—but where they are relevant, their impact is great. In these
use cases, the common theme is the need to identify individuals, which is most easily done
through analysis of images. An example is the use of face detection on surveillance footage
to detect the presence of escaped criminals in a specific area.16 In education, emotion
recognition on video or image data can be helpful in determining which specific students
need extra attention and help. As we discuss elsewhere, these capabilities also introduce
significant risks, for example around privacy, that must be understood and managed.
Natural language processing can be applied for societal impact where language
and communication barriers are a roadblock
Some aspects of natural language processing, including sentiment analysis, language
translation, and language understanding, also stand out as applicable to a wide range of
domains and use cases. Natural language processing is most useful in domains where
information is commonly stored in unstructured textual form, such as incident reports,
health records, newspaper articles, and social media posts such as tweets.
As with computer vision–based methods, in some cases a human may be able to perform
the task with greater accuracy than a trained machine learning model. However, the speed
of “good enough” automated systems can enable meaningful scale efficiencies—for
example, providing automated answers to questions asked by citizens through email. There
are also cases where AI models could outperform humans in effectiveness, especially
in situations that require processing and analyzing vast amounts of information quickly.
Examples include monitoring disease outbreaks by analyzing tweets sent in multiple
local languages.
As with computer vision capabilities, some language-related capabilities are not used
in many domains, but they add substantial incremental value where they are used.
For instance, language translation provides value in domains where language and
communication barriers are a major roadblock, for example when migrant populations
have difficulty communicating in the local language. In health, we often see challenges in
treatment due to language barriers between doctors and patients. Language translation can
also be used in education, to help students learn new languages.
Some capabilities, or combinations of capabilities, can open possibilities for the

target population that would not otherwise exist—especially in use cases that involve
understanding the natural environment through interpretation of vision, sound, and speech.
For example, real-time description of one’s environment could be game-changing in helping
blind people navigate their surroundings. Another example is the use of AI to help educate
children who are on the autism spectrum. Although professional therapists have proven
effective in creating behavioral learning plans for children with autism spectrum disorder,
Face detection applications detect the presence of people in an image or video frame. This is not the same as
16
facial recognition, which is used to identify individuals by their features.
waiting lists for therapy can be long.17 AI tools, primarily using emotion recognition and
face detection capabilities, can increase access to this education by providing cues to
help children identify and ultimately learn facial expressions among their family members
and friends.
Structured deep learning analyzes traditional tabular data sets, which are often
accessible for societal impact uses
The potential of structured deep learning across domains that we find in our use case library
is perhaps more surprising because it is not commonly used today. This capability involves
analysis of traditional tabular data sets using deep learning. It can contribute to solving
problems ranging from identifying fraud based on tax return data to finding patterns of
insights in electronic health records that would be very hard for humans to discover.
Structured deep learning (SDL) has been gaining momentum in the commercial sector,
and we expect to see that trend spill over into solutions built for social good use cases,
particularly given the abundance of tabular data in the public and social sectors. The
advantage of SDL solutions is that they reduce the need for domain expertise or for an
innate understanding of the data by automating aspects of basic feature engineering, the
process of selecting and encoding the features in the data that will be most relevant. This
could make SDL easier to use compared with other analytics models, especially for groups
like NGOs that have limited resources and talent.
For now, the value of using structured deep learning is only starting to emerge. For example,
in commercial applications, Instacart has successfully used SDL to enable its professional
shoppers to efficiently navigate stores, reducing shopping time by minutes per delivery—
an important efficiency gain.18 In another example, Pinterest developed an SDL system
to surface relevant recommendations that led to a 5 percent increase in its “related pin”
engagement, a funnel to money-making features such as direct advertising and monetizable
pins.19 A third example is the use of trip data to train structured deep learning systems that
predict taxi trajectories. In a Kaggle competition, this solution outperformed all non–deep
learning solutions to win first place, increasing the efficiency of electronic taxi dispatching
systems and helping optimize public transport.20
While our research has so far not identified actual deployment of SDL for social good in the
world today, our use case library suggests that it has considerable potential to be used for
noncommercial purposes across a range of domains because tabular data exists in relative
abundance in almost every domain. It is by far the most common data type in the public-
and social-sector management and infrastructure domains. For example, it is easy to see
potential for public-sector agencies to deploy types of SDL architecture similar to Instacart’s
shopper navigation solution to optimize deployment of emergency vehicles, police, and
other agents for greater workforce efficiency than can be gained through optimization
methods using traditional analytics. Other potential examples include using structured deep
learning in education to help recommend educational and vocational opportunities using
academic history and financial information. In the equality and inclusion domain, it could be
used to automatically enroll individuals on welfare in programs of which they were not aware,
but for which they were qualified. In the health and hunger domain, too, broad applications,
such as increasing patient support through identifying drivers of hospital performance,
have potential.
17
“Autism therapy wait list changes ‘a difficult process,’ minister acknowledges,” CBC News, April 2, 2016,
cbc.ca/news/canada/ottawa/autism-wait-list-ontario-minister-1.3517508.
18
Jeremy Stanley, “Deep learning with emojis (not math),” Medium, March 29, 2017, tech.instacart.com/
deep-learning-with-emojis-not-math-660ba1ad6cdc.
19
Kevin Ma, “Applying deep learning to Related Pins,” Medium, January 12, 2017,
medium.com/the-graph/applying-deep-learning-to-related-pins-a6fee3c92f5e.
20
No free hunch, “Taxi trajectory winners’ interview: 1st place, team?,” blog entry by Kaggle Team, July 27,
2015, blog.kaggle.com/2015/07/27/taxi-trajectory-winners-interview-1st-place-team-%F0%9F%9A%95/.
OTHER CAPABILITIES HAVE SOCIAL POTENTIAL, INCLUDING SOUND
RECOGNITION, REINFORCEMENT LEARNING, AND ADVANCED ANALYTICS
Beyond these three capabilities, our use case library suggests that other capabilities,
including both developing AI ones and more established advanced analytics, have potential
applications for social benefit.
Sound detection and recognition could become increasingly relevant as

auditory sensors are deployed for use cases across domains
Sound detection and recognition is also a fairly mature capability that has widespread reach
given the prevalence of audio in the natural world. It is a common AI capability that can
be used across many domains. As more auditory sensors are deployed, the scope of this
capability will increase correspondingly. Today, sound detection is an important capability
for use cases such as the exposure of illegal logging, diagnosis of medical or neurological
conditions such as Parkinson’s disease, predictive maintenance of public transportation
systems, and providing adaptive learning to students.
Reinforcement learning and content generation, while nascent capabilities,

have potential for social use
Reinforcement learning and content generation are relatively young capabilities and
are not particularly prevalent in our use case library, but they do feature in a few cases.
Reinforcement learning is a capability in which systems are trained by receiving virtual
“rewards” or “punishments,” essentially learning by trial and error. It was the primary
technique used in the breakthrough AlphaGo program, which became the first program
to beat a human professional Go player. It can be used, for example, to train models that
recommend precision medicine–based treatments for individual patients. Researchers from
the MIT Media Lab created a model that explores previous oncology drug regimens and at
each monthly checkpoint determines updates that need to be made to patient doses, with
the end goal of shrinking tumor size.21
Recent progress in visual content generation has been made using generative adversarial
networks (GANs), a neural network architecture. Given their ability to mimic data sets,
GANs are particularly effective at augmenting existing data sets.22 GANs can also carry
out a number of image-related tasks, including transforming visuals into a different style,
improving resolution of low-resolution images, and repairing holes in images. The ability to
generate language content is still developing and will need more improvement before the
potential to generate meaningful text, such as in chatbot responses for mental wellness and
medical consultations, can be fully captured.
Advanced analytics can be a more time- and cost-effective solution than AI for
some use cases
Some use cases in our library are better suited to analytics techniques other than those that
involve deep learning, since analytics are often less complex to deploy than those involving
AI techniques. Moreover, for certain tasks, other analytical techniques can be better suited
than deep learning. For example, in cases where there is a premium on explainability,
decision tree–based models can often be more easily understood by humans. A financial
institution might deploy into production tree-based models for underwriting, despite having
parameters that were set based on experiments with deep learning.
21
Rob Matheson, “Artificial intelligence model ‘learns’ from patient data to make cancer treatment less toxic,”
MIT News, August 9, 2018, news.mit.edu/2018/artificial-intelligence-model-learns-patient-data-cancer-
treatment-less-toxic-0810.
22
GANs operate by use of two neural networks: the “generator” creates new instances of data, while the
“discriminator” evaluates their authenticity.
The full range of analytics techniques is likely to continue playing a key role because
structured deep learning is only just gaining momentum and has not shown marked
improvement over traditional methods for optimization problems, except in select
commercial use cases. One example is the use case of optimizing public-sector agent
deployment (for example, firefighters and EMTs). While SDL may surpass the use of analytics
for this use case in the next few years, it is important to note that analytics techniques have a
lower barrier to deployment and may thus be a preferable solution for an organization, given
the available resources and technical capabilities.
THE VERTICAL VIEW OF DOMAIN COMPARISONS BRINGS OUT NUANCES IN

THE PREVALENCE OF DIFFERENT DATA MODALITIES
In our use, modality refers to the type of data an AI system uses—for example, data from
light sensors (that is, image and video) or from audio sensors. A multimodal AI solution
is one that takes in different types of data inputs. Adaptive learning, assisting those with
disabilities, assisting firefighters in navigating paths in a fire, and using images, audio,
and even geospatial data to identify extremist online content for removal are all examples
of multimodal applications for social good. Modality is a potentially important issue,
especially for social good, because multimodal AI solutions are highly complex and may
take more time and more resources to build (but might also have higher performance) than
unimodal solutions.
Across domains, a subset of use cases (only 28 percent of the total) require multimodality
to unlock maximum societal value. Three-quarters of them are concentrated in the same
five domains where many AI capabilities are applicable: crisis response, education, equality
and inclusion, health and hunger, and security and justice (Exhibit 7). Overall, among these
five domains, only 38 percent of use cases are multimodal; the remaining 62 percent either
use only a single source of unstructured data and perhaps structured data or use only
structured data.
These domains split into two groups. For crisis response, education, and health and hunger,
we find many AI capabilities that can be deployed because their use cases require a wide
variety of types of data input, and most use only one type of unstructured data. This results
in many AI capabilities being applicable across these domains.
For the equality and inclusion and the security and justice domains, however, we find a high
percentage of multimodal use cases (73 percent and 63 percent, respectively), due to an
emphasis on natural environment and human behavior understanding in these domains.
Use cases can include, for example, narrating the environment for the blind and decreasing
search time for criminal suspects using audio and video surveillance footage, which typically
require multimodality. Such solutions will be more complex to build, and the impact in these
two domains may thus not be realized as quickly as in other domains.
Exhibit 7
More than 70 percent of use cases require only one modality. Five domains for which many AI capabilities are
relevant have the biggest share of multimodal use cases.
Modality is the way something is experienced, eg, seen (image and video), heard (audio)
Number of use cases in library Distribution of data type modality for use cases in each domain
using each type of data1
Multimodal unstructured Single unstructured Structured
Text data source2 data source2 data only
55 The 5 domains where many AI capabilities are relevant
Audio Crisis response 5 17
29
Education 5 13
Video
26 Equality and inclusion 8 11
Image
Health and hunger 4 28
64
Structured Security and justice 10 16
119
Economic empowerment 2 15
Environment 1 21
Information verification
4 4
and validation
Infrastructure 2 15
Public and social sector 2 16
1 Aggregate sum of use cases in this chart exceeds 156 because use cases can employ more than one type of data.
2 The category “Single unstructured data source” represents use cases that employ only one type of unstructured data (eg, text, image, video, or audio) but
may or may not also make use of structured data; for “Multimodal unstructured data source,” two or more types of unstructured data are required by the use
cases, and structured data may be used as well.
3. SIX ILLUSTRATIVE USE CASES
AI will be deployed in different ways depending on the crises such as Hurricane Harvey in Houston (number 6).
domain, capability, barriers, and risk profiles of specific The fifth case, focused on the prediction of houses in
use cases. To illustrate the breadth of areas where AI Flint, Michigan, that may have lead pipes, uses machine
could be applied for social good, this section goes into learning techniques. While these are often considered to
more specific detail about six cases. While these are just be AI, they are classified under the umbrella of analytics
a specific set of examples, they highlight the range of for this paper, since we use a narrower definition of AI
cases, the different capabilities that could be used, and as deep learning. This use case example underscores
the potential impact. Several of the cases use capabilities how AI is one solution among others; machine
in the computer vision category, including helping the learning and analytics solutions are also applicable and
visually impaired globally (number 1), tracking wildlife sometimes better suited for some social problems than
poachers in South Africa (number 4), and responding to deep learning.
1. Using computer vision to help While advances in computer vision have improved
the visually impaired navigate scene description technology, this remains a developing
capability. As it improves, it will provide visually impaired
their environment
people with a richer and deeper understanding of their
An estimated 250 million people worldwide have environment, including recognizing other people and
moderate to severe visual impairment, and about describing colors, for example.
90 percent of them live in developing nations (Exhibit 8).
Not being able to see is a major impediment to work; Limited access to technology in emerging
about 70 percent of working age adults with visual economies is a potential barrier that will need to
impairment in the United States are unemployed. The be overcome
World Health Organization estimates that the global One of the most significant barriers to AI in this area
economic impact of unaccommodated blindness and low is “last mile” implementation in the form of access to
vision worldwide exceeds $42 billion.23 technology. Most visually impaired people live in emerging
economies where internet access and bandwidth
Better medical care, nutrition, and other measures can may be lacking, limiting the ability of the user to send a
help prevent blindness or ease the suffering of those picture to the cloud for processing. Many people do not
afflicted by it. Alongside these solutions, AI has the own the smartphone needed for existing applications;
potential to make life easier for the visually impaired smartphone penetration globally is below 40 percent.25
by providing navigational assistance via smartphone. Building partnerships with NGOs and governments to
AI’s computer vision capabilities can identify objects provide basic technology access to individuals in poor
and read text, converting handwriting or printed text communities could help address this limitation.
to digital text that can then be spoken aloud. Existing
mobile applications include Microsoft’s Seeing AI, which Talent is also a potential bottleneck. Developing AI
is accessible to users in 70 countries around the world applications to tailor them more closely to the needs of
and free of charge. Reviews of the application indicate the visually impaired in various regions, including with
that it has very large daily applicability.24 Similar solutions more detailed description of the surrounding environment,
include the OrCam MyEye camera, which is mounted on may require the efforts of high-level AI researchers over a
standard glasses and converts what is seen into spoken period of years.
output. The device is easily portable and operates without
the need for a smartphone but carries a $3,000 price tag.
23
The World Health Organization and more than 20 international NGOs in 1999 launched a global initiative, Vision 2020, that aims to eliminate
avoidable blindness by the year 2020. The initiative provides guidance and technical and resource support to countries that have formally
adopted its agenda. See iapb.org/vision-2020/.
24
See for example, Steven Kelley, “Seeing AI: Artificial Intelligence for blind and visually impaired use,” VisionAware, American Foundation for the
Blind, visionaware.org/info/everyday-living/helpful-products/using-apps/seeing-ai-app/1235.
25
The mobile economy 2018, GSMA, 2018, gsma.com/mobileeconomy/wp-content/uploads/2018/05/The-Mobile-Economy-2018.pdf.
Exhibit 8
Most visually impaired people in the world live in emerging economies.
2015
Moderate and severe 0–0.5 0.5–1.0 1.0–1.5 1.5–2.0 2.0–2.5

vision impairment
% 2.5–3.0 3.0–3.5 3.5–4.0 4.0–4.5 >4.5
SOURCE: The Vision Loss Expert Group; McKinsey Global Institute analysis
Among the risks, data privacy looms large. Users of 2. Using computer vision to diagnose
such a mobile application would capture images of the skin cancer from mobile phone photos
environment that often include other individuals or private
data, for example pictures of credit cards, and may also One in three cancer diagnoses is for skin cancer. When
rely on training data that could include private information. detected at an early stage, skin cancer survival rates
Since these images could be automatically uploaded to can be as high as 97 percent, but they drop as low as
the cloud, application developers will need to take steps 14 percent with late-stage detection.26 Detection today
toward protecting data privacy. Anonymized storage of is largely done by dermatologists looking at moles with
data provided through the app may ensure security and the naked eye or a dermoscope. Residents of rural
privacy in the event of a data breach, although there are communities without dermatologists in the surrounding
some doubts in the industry about the effectiveness of area are consequently at particular risk of late detection.
anonymization. Another strategy is to rely more heavily
Some experiments suggest that AI can diagnose skin
on edge computing and process the data on the device
cancer with greater accuracy than human dermatologists;
where it is being generated (the user’s smartphone)
in a pilot, AI classification of skin lesions as melanomas
instead of in a centralized data-processing facility (the
versus benign beat classification by 58 dermatologists. AI
cloud). This minimizes the chances that a central data
diagnosed cancerous moles with 95 percent accuracy,
storage breach will affect a user whose information is
while the dermatologists’ accuracy was 86 percent.27 This
mainly stored in a distributed fashion.
26
Taylor Kubota, “Deep learning algorithm does as well as dermatologists in identifying skin cancer, Stanford News, January 25, 2017,
https://news.stanford.edu/2017/01/25/artificial-intelligence-used-identify-skin-cancer/.
27
Thirty of the dermatologists were deemed experts, defined as having had more than five years of experience; 11 were deemed “skilled,” with
between two and five years’ experience; and 17 were “beginners,” with less than two years’ experience. H. A. Haenssle et al., “Man against
machine: Diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58
dermatologists, Annals of Oncology, August 2018, Volume 29, Issue 8, pp. 1836–1842, https://www.ncbi.nlm.nih.gov/pubmed/29846502.
creates the potential for a mobile application using image the possibility of sharing healthcare data, the FinnGen
recognition to make testing accessible to all, including genomic research partnership was able to create a
rural communities around the world that do not have easy system to gather and share blood sample data through
access to dermatologists.28 biobanks. The University of Helsinki led the project,
working with the Helsinki University Central Hospital
This AI solution would use object detection and image and seven pharmaceutical companies. All Finns can
classification of skin cancer images. A convolutional participate in the research by giving consent to the
neural network system could be leveraged to build biobanks to use their sample. Digital healthcare data
detection of cancerous versus regular skin lesions. from national registries were also provided to support the
The model would be trained using image data at a raw FinnGen research consortium.29
pixel level with associated cancer versus non-cancer
tags. Training would require the aggregation of data “Last mile” implementation issues could also be a
sets with thousands of such images and diagnostic bottleneck. The most prominent roadblock to deployment
information. During deployment, images could potentially is the question of which party would be liable for
be submitted through an internet-based mobile app or misdiagnoses: the solution provider, the healthcare
through MMS to maximize the user base. provider, or the insurer? This is not dissimilar to the
unresolved accountability questions about self-driving
Further testing and development will be needed cars. Once diagnosis through a mobile application
to ensure that the benefits are realized on a is conducted, healthcare access, explanations, and
broad scale coverage are needed. Specialized care is required to act
While the technology for deployment already exists and on the diagnosis from a mobile application. This can be a
has been piloted, a solution for mobile usage would need challenge in many cases, including for rural communities.
to be packaged and further tested before it could be While smartphone penetration is rising globally, it remains
used and scaled. This can be accomplished fairly easily below 50 percent in many developing regions (Exhibit 9).
by high-level engineering talent. However, to improve
diagnosis of additional skin cancers not included in the Risks involved with mobile cancer diagnosis will need to
initial model and continue to increase accuracy, further be mitigated. Privacy violation is one such risk. Patient
technical work by AI talent with high-level expertise will consent is required to leverage photos to be used for
likely be needed. research purposes and diagnosis. Some countries do
not allow healthcare data to leave the country (the data
Even if such a solution were to be deployed in a localized must be stored on servers within the countries’ borders).
setting, data availability bottlenecks would need to be A solution may be to build a global diagnostic model
overcome to ensure that the benefits could be realized on with anonymized databases, including country-specific
a broader scale. Alternatively, emerging techniques such databases when regulations apply, and require patient
as transfer learning and new model architectures could consent for doctors and others deploying the AI model.
enable solutions to be derived from small data sets. For example, 23andme users opt in for a service to
allow for data collection. This requires explanation of the
Data sets containing images of skin cancer and other
implications of opting in and adequate administrators to
conditions are not centralized and will need to be
evaluate adherence to rules.30
collected from many different healthcare providers.
Potential solutions include consolidating provider data Beyond privacy issues, there are also risks associated
where possible on a platform and augmenting the with ensuring safety. Legal repercussions could also
database using physician-provided skin photos and result in case of misdiagnosis, though the question of
diagnostic information. Another option would be to who would be liable remains unresolved, as discussed
collaborate with large health systems including single above. A solution could be to require “human in the
country payors. In Finland, in an example illustrating loop” intervention, for example a physician to validate
28
Some commercial applications for this technology are also underway. See for example, Jonah Comstock, “SkinVision gets $7.6M to continue
expanding skin cancer app,” Mobihealthnews, July 30, 2018, www.mobihealthnews.com/content/skinvision-gets-76m-continue-expanding-
skin-cancer-app.
29
Mari Kaunisto, FinnGen taps into a unique gene pool to find the next breakthroughs in disease prevention, diagnosis and treatment, University
of Helsinki FinnGen press release, December 19, 2017, helsinki.fi/en/news/health/finngen-taps-into-a-unique-gene-pool-to-find-the-next-
breakthroughs-in-disease-prevention-diagnosis-and-treatment.
30
Research participation and consent, 23andme, customercare.23andme.com/hc/en-us/articles/212195708-Research-Participation-and-
Consent.
a diagnosis made by an AI solution. 31 Rigorous testing body. Training data from people with different skin
would also be required to ensure that diagnoses are pigmentation will be required to enable effective and
accurate even if the context changes, for example, inclusive diagnosis.
different skin tones or lesions in atypical parts of the
One team of researchers using machine learning to diagnose skin cancer discovered that their system reported a higher probability of skin
31
cancer when the image included a ruler, because the training data included images of skin cancer that were typically accompanied by a ruler
to illustrate the size of the lesion. Jan Bowers, “Should dermatologists fear machine learning?” Dermatology World, February 2018, digital.
ipcprintservices.com/publication/?i=469994&article_id=2989257.
Exhibit 9
Smartphone penetration is rising but still low in emerging regions, where most visually impaired and unbanked or
financially vulnerable people live.
Phone type penetration per region Developing No Feature Smartphone

% regions phone phone
2017 actual 2025 estimate
79 60
Sub-Saharan Africa 13 2
8 38
64 49
Middle East and
17 3
North Africa
19 48
59 37
Asia-Pacific 17 3
24 60
50 34
Latin America 17 3
33 63
48 28
Commonwealth of
26 0
Independent States
26 72
28 18
Europe 16 0
56 82
27 20
North America 6 2
67 78
NOTE: Feature phone figures assume equivalent to 2G penetration, while smartphone figures assume penetration of phones that use 3G or beyond. Figures
may not sum to 100% because of rounding.
SOURCE: The mobile economy 2018, GSMA, 2018; McKinsey Global Institute analysis
3. Using AI models to improve financial be used as an additional layer of verification to confirm
inclusion in emerging economies an individual’s identity. Structured and unstructured data
from sources including social media, browsing history,
About 1.7 billion people worldwide are “unbanked,” that telecom, and know-your-customer data can be used to
is, they do not have an account at a financial institution train AI models. Solutions are likely to start with external
or through a mobile money provider.32 About half of them data such as longevity as a telecom customer, and the
are from the poorest households in the global economy. model is then augmented against a client’s actual product
Companies in some countries can also have difficulties borrowing performance.
in accessing finance; according to the World Bank,
20 percent of small and medium-sized enterprises in Successful implementation will require
Africa cite finance as the biggest obstacle to maintaining integration of multiple data sources
their business.33 Key challenges to overcome include integrating multiple
data sources, given different methods of storing
AI can help address core difficulties holding back financial information. The outcomes must be tested rigorously
inclusion: difficulties in verifying identities and in a lack and explainable where necessary, given that an AI model
of traditional data for underwriting services to vulnerable is analyzing personal data to sort people, assess the
populations. AI-based financial inclusion products, credit risk of customers, and potentially reject some.
which are already being deployed, bypass the need for a Representative positive and negative data need to be
traditional credit score through analyzing digital footprints. collected to help reduce unwanted biases. Identifying
Many fintech startups are entering the alternative credit- the levers considered most strongly by the model when
scoring space with AI-enhanced solutions, particularly determining a credit score could help, as could providing
in countries such as Bangladesh and Pakistan where information on the model decision-making process,
the populations are large and significant portions are particularly for rejected users.
unbanked (Exhibit 10).
“Last mile” implementation is also a potential bottleneck,
Companies such as CreditVidya, ZestFinance, and because many adults in emerging economies who are
Lenddo capture alternative data by device, browser, and unbanked also do not have mobile phones or internet
social media fingerprinting to generate a predictive model access. This prevents the creation of a digital footprint,
of creditworthiness. M-Shwari banking, which leverages including telecommunications and online social history,
the M-Pesa mobile money system in Kenya, incorporates which are the data needed for this form of alternative
telecommunications history in its assessment of credit credit scoring. Overcoming such bottlenecks may require
risk. One in five adults in Kenya is currently an active building partnerships with NGOs to provide funding for
user of the service, and M-Shwari is regarded as one basic technology access.
of the most successful solutions for financial inclusion.
Behind its SMS and internet-based interface, predictive Other implementation challenges may depend on the
algorithms leverage several AI capabilities to analyze willingness of financial institutions to provide a mobile
social and telecom data and assess creditworthiness. money infrastructure and support the use of alternate
The information is then processed in minutes and credit ratings, and on providing financial education and
produces a credit score, which determines the size of the transparency to customers. For example, they will need
loan allowed. to understand the implications of various actions on
assessments of their creditworthiness.
A range of capabilities can be leveraged for these
products, including natural language processing, In terms of risk—as with other solutions that draw inputs,
structured deep learning, and person identification including social media data and purchase histories, from
of social and telecom data. Long short-term memory highly personal information—data privacy and security
recurrent neural networks can be trained to recognize an are essential.
individual’s credit risk. Image-processing capabilities can
32
For a detailed discussion of financial inclusion, see Digital finance for all: Powering inclusive growth in emerging economies, McKinsey Global
Institute, September 2016, and the Global Financial Inclusion database, World Bank. globalfindex.worldbank.org.
33
Access to finance for small and medium enterprises in Africa, World Bank, 2016, acetforafrica.org/acet/wp-content/uploads/
publications/2016/03/Access-to-Finance-for-SMEs-Paper.pdf.
Exhibit 10
Financial inclusion is high in advanced economies but millions of people in emerging economies are unbanked.
Estimated adult population with an account at a bank or other financial institution in 20171
%
South Asia 68
Sub-Saharan
33
Africa
Middle East and
43
North Africa
Excluding
high- Latin America
53
income and Caribbean
countries
Europe and
65
Central Asia
East Asia and
70
Pacific
High-income:
High- 94
World
income
countries High-income:
95
OECD
World 67
In 21 countries, fewer than 20% of adults had an account at a bank or other financial institution in 2017
Countries include
 Afghanistan
 Burkina Faso
 Cambodia
 Central African Republic
 Chad
 Democratic Republic of Congo
 Republic of the Congo
 Ivory Coast
 Guinea
 Iraq
 Liberia
 Madagascar
 Malawi
 Mali
 Mauritania
 Niger
 Pakistan
 Senegal
 Sierra Leone
 South Sudan
 Tanzania
1 Figures are based on percentage of respondents to a World Bank survey who indicated they have an account at a bank or other financial institution.
SOURCE: The World Bank Group Global Findex Database, globalfindex.worldbank.org; McKinsey Global Institute analysis
4. Using AI-powered drones to spot and Scaling the AI solution for use in multiple
stop poachers parks will be needed to improve efficiency and
reduce costs
The United Nations estimates that illegal wildlife trade Since each park has different flora, fauna, weather,
worldwide could be worth $8 billion to $10 billion and other conditions, data acquisition may need to be
annually.34 The value of the ivory trade alone contributes customized to optimize performance in any new park. To
about $1 billion.35 avoid having to generate a large amount of training data at
the outset, the program could potentially be scaled to first
For now, foot patrols and drone-based surveillance have
target parks like those in the pilot programs, in order to
not been effective at preventing poaching, and these
minimize the labor required for additional training.
efforts are labor-intensive and under resourced. One
AI-based solution has already been built and tested, and Real-time processing of data requires access to GPU-
it had some initial success in combating poachers. The powered or other systems either in the cloud or on local
SPOT system, built by researchers from the University computers. A local machine requires a significantly
of Southern California’s Center for Artificial Intelligence larger initial investment, while the cloud needs a reliable
in Society and piloted by the organization Air Shepherd, internet connection. Sharing a public cloud instance
automates the process of detecting poachers in infrared or a GPU-powered local machine between multiple
video feeds, freeing park rangers for other tasks and parks would minimize the upfront investment as well as
increasing the reliability of surveillance (see illustration, operating costs.
“How AI can be deployed to catch wildlife poachers”).36
In general, hiring AI talent who can build and train this
The solution uses image classification and object model could be challenging. Not all organizations looking
detection to find animals and poachers on infrared video to do something similar would have access to the high-
captured by a drone at night. A convolutional neural level AI expertise that could develop the model, and even
network model is trained to recognize both poachers if they do, they may not have long-term support to refresh
and animals despite their small size in the video feed. (or customize), troubleshoot, and improve on the model
The SPOT model does not need to be customized by over time.
the drone user and can be used in the field immediately,
although a trained drone operator is still required. Implementation costs for state-of-the-art drones and
infrared cameras could be significant for resource-
Air Shepherd has reported some success in South Africa constrained parks, although building partnerships with
and plans for wider rollout in Botswana. In one area NGOs and governments to provide access to the required
where as many as 19 rhinos were killed each month, there funds to purchase and maintain the equipment could be a
were no deaths for at least six months after the program cost-sharing solution.
was deployed.
Implementation talent (non-AI) is also a challenge: while
With more development, SPOT and similar solutions SPOT will eventually pilot the drone, the aircraft will still
could guide drones autonomously, adjusting flight routes require trained personnel for launching, maintaining,
to track poachers and removing the need for highly and troubleshooting of equipment, including the infrared
trained pilots and systems operators in the more than 300 camera. The limited supply and cost of employing these
wildlife parks in the world. professionals can limit scalability. Scaling the program
could involve parks sharing drone operators, driving down
costs. Drone operators can also train park personnel to
use the drones. Once they are proficient, park personnel
would need to contact the professional operators only for
specialized repairs.
34
Wildlife crime worth USD 8-10 billion annually, ranking it alongside human trafficking, arms and drug dealing in terms of profits, United Nations
Office on Drugs and Crime, May 13, 2014, unodc.org/unodc/en/frontpage/2014/May/wildlife-crime-worth-8-10-billion-annually.html.
35
Jason Bellini, “Price of ivory? 100 elephants killed per day,” Wall Street Journal, March 19, 2015, blogs.wsj.com/briefly/2015/03/19/price-of-
ivory-100-elephants-killed-per-day-the-short-answer/.
36
Elizabeth Bondi et al., SPOT poachers in action: Augmenting conservation drones with automatic detection in near real time, 32nd AAAI
Conference on Artificial Intelligence, April 27, 2018, teamcore.usc.edu/papers/2018/spot-camera-ready.pdf.
How AI can be deployed to catch wildlife poachers
Six steps from offline training of AI model to online detection
1. Offline training 2. Drone deployment 3. User interface

A neural network is trained Drones are flown over Video and still images are
on 70 videos, containing wildlife sanctuaries, transmitted via radio
animals and poachers, capturing thermal infrared waves to a computer.
that have been labeled. images.
The model is tested with
other videos.
4. Pre-processing 5. Detection 6. Output

The infrared images may The video is processed in The neural network
need to be converted to batches and sent to the outputs annotations that
“white-hot” format, where cloud for analysis. Each are overlaid on top of the
warm objects are lighter image is treated as an original image. This
against a dark input into the neural enables identification of
background. network. the poachers’
whereabouts.
SOURCE: Elizabeth Bondi et al., SPOT poachers in action: Augmenting conservation drones with automatic detection in near real time,
32nd AAAI Conference on Artificial Intelligence, April 27, 2018; McKinsey Global Institute analysis
5. In Flint, Michigan, detecting water learning—often considered an AI technique—and
service lines containing lead Bayesian analysis to make pipe material predictions.
It uses a combination of city parcel data, census
In the water crisis in Flint, Michigan, lead contamination data, and incomplete records of water service lines
of water from pipes resulted in a lead concentration in the to predict the homes most likely to have lead pipes
water supply that was 300 times the U.S. Environmental (see illustration, “How the lead pipe detection model in
Protection Agency limit.37 About 9,000 children aged six Flint, Michigan, works”). In addition to the application
or younger were exposed to the risk of permanent harm in Flint, this approach could be relevant in other water
to brain development, impaired learning abilities, and pipe replacement efforts, although local data would be
behavioral disorders.38 required. As the case illustrates, a broad view of analytical
and AI techniques that could be applicable to any given
The primary contributor to lead in the water system
problem can be more appropriate in some cases than
is thought to be water service lines, which connect
using only deep learning.
the city water supply to homes. However, finding and
replacing water service lines in Flint requires excavation, Data availability is a key to this solution and will
because of a lack of records about pipe materials. This need to be scaled
is where AI and advanced analytics can come in. The A stringent and systematic data-gathering process by the
University of Michigan has developed a model called city is key. Digitization of such data is also a prerequisite
ActiveRemediation that can predict with 98 percent for AI deployment. One of the cost factors in Flint was the
accuracy whether a water service line is lead. Deployment lack of digital records.
of the predictive model in Flint reduced unnecessary
replacement excavations from 18.8 percent to Implementation was relatively straightforward, but the
2 percent.39 challenge will be to scale this type of solution. The pipe
material model was developed specifically for Flint with
While the other use cases in this section deploy deep the data available in Flint. That means other data sources
learning, this example combines an instance of machine will be required for application to other areas.
37
Christopher Ingraham, “This is how toxic Flint’s water really is,” Washington Post, January 15, 2016, washingtonpost.com/news/wonk/
wp/2016/01/15/this-is-how-toxic-flints-water-really-is.
38
The University of Michigan’s ActiveRemediation model for detecting lead pipes is described in detail in Jacob Abernethy et al.,
ActiveRemediation: The search for lead pipes in Flint, Michigan, Cornell University Library, August 17, 2018. https://arxiv.org/abs/1806.10692.
See also Lead poisoning and health, World Health Organization, August 23, 2018, www.who.int/news-room/fact-sheets/detail/lead-poisoning-
and-health.
39
Ibid. Jacob Abernethy et al., Cornell University Library, August 17, 2018.
How the lead pipe detection model in Flint, Michigan, works
Six steps from offline training of AI model to online detection
Data collection
1. City of Flint provided team with a dataset 2. Digitization of city records of
describing each of the 55,893 parcels in service lines using OCR on data,
the city. It includes attributes of each home, such as hand-annotated maps
such as property owner, address, value, containing markings for each parcel
and building characteristics. and incomplete records of materials
used for each home’s service line.
Model building
3. Machine learning model

predicts probability that
the portion of a service line
attached to a particular
10.3% 12.8% 21.0% 94.4% 91.2% 17.5%
home is hazardous based
on known features, such
as home value.
4. A Bayesian model is
used to fine-tune
predictions to account
for relationships between 91.8% 14.8% 13.3%
parcels and precincts.
5. Algorithm determines
which homes to observe 10.3%
next, by efficiently
allocating resources.
12.2% 19.7%
12.2% 10.3%
6. Determination of final
list of homes to be visited
by the replacement crew, 74.1%
based on the probability 12.5%
that they have lead pipes.
Blocks of houses in the Grand Traverse district in Flint, Michigan. Percentages on each house denote
the machine learning–derived probability of lead water pipes.
SOURCE: Jacob Abernethy et al., ActiveRemediation: The search for lead pipes in Flint, Michigan, Cornell University Library, August 17, 2018;
McKinsey Global Institute analysis
6. Analyzing satellite data to help cope granular portions of Earth, such as small agricultural fields
with the aftermath of natural disasters in developing nations or clusters of buildings in urban
areas. Satellite imagery can also provide perspective into
Natural disasters kill more than 50,000 people and infrastructure projects in remote, inaccessible locations
displace tens of millions more each year. The total where on-foot monitoring is often infeasible due to the
economic damage amounts to more than $100 billion arduous, time-consuming nature of collecting information
annually. Populations in less developed countries are on the ground. The Indian government has employed
hardest hit: in countries with a medium or low human satellite-based monitoring systems to understand the
development index, up to six times as many people status of projects on a timely basis and detect where fund
can be affected by natural disasters compared to more allocation is not effective.42
affluent countries, and the economic damage can be four
times greater.40 Use cases that pair satellite data with other data include
augmenting remote sensing data from satellites with
To coordinate and prioritize emergency response, prices to create a forecast of agricultural production.
governments and first responders must have an accurate Satellite data could also be used to measure economic
and complete view of disaster zones. Frequent and broad activity by detecting car density, construction rates, and
area satellite imagery enables new AI-based systems electricity consumption through nighttime illumination.
to quickly and accurately detect infrastructure changes The following illustration is a photograph showing how
that may affect evacuation and response. AI can assist object detection software was applied to satellite imagery
in improving relief efforts and emergency preparedness to detect flooded roads after Hurricane Harvey in 2017.
with greater accuracy and on a much larger scale than The object detection software is trained to detect and
human workers. For example, following the passage of map roads in imagery. During a flood, the software can
Hurricane Harvey in 2017, a collaboration between Planet rapidly identify which roads are no longer distinguishable,
Labs, provider of satellite imagery, and CrowdAI provided and therefore impassable, due to flood waters. In the
an immediate view of the greater Houston area and was photograph, flooded roads are marked in red, while
able to detect road outages due to flooding and quantify nonflooded ones are in orange.
infrastructure damage.
Access to satellite data and rigorous testing are
The solution leverages computer vision capabilities— essential for AI use for disasters
specifically, object detection—to determine which To use satellite data in disaster scenarios, access is the
portions of satellite imagery belong to the target feature. In key challenge to overcome. Most satellite data are held
the case of detecting road outages, the model was able to by institutions that often do not make their data freely
identify all roads in a satellite image. Other assessments available, and access to the data may be prohibitively
of critical infrastructure damage required identification of expensive for NGOs. However, most space agencies
objects such as building outlines. The solution was also and space system operators—in addition to some
able to identify the presence or absence of water in broad nongovernmental satellite data owners—belong to the
areas affected by floods on false color satellite imagery International Charter on Space and Major Disasters.
using various analytic methods. Combined, the resulting Through that worldwide collaboration, satellite
mapping provides an accurate view of usable roads data are made available at no cost for purposes of
and undamaged buildings, updated daily as the satellite disaster management.
imagery is refreshed.41
In terms of risk, as with other solutions that provide
Satellite data can be used for many other solutions critical information, the need for rigorous testing to
ranging from short-term weather forecasts to tracking ensure accuracy is essential. If the solution incorrectly
deforestation. Indeed, satellite data can power AI indicates that a road is clear of flooding and directs
applications across all ten of our domains. This is thousands of people toward it, there could be significant
especially the case for infrastructure and economic consequences, such as increased risk of harm and
empowerment, where advances in satellite imagery reduced evacuation speed.
resolution enable visualization and monitoring of more
40
More affluent countries are defined as those with either a very high or high human development index. Hannah Ritchie and Max Roser, “Natural
catastrophes,” Ourworldindata.org, 2018, ourworldindata.org/natural-catastrophes.
41
“Anatomy of a catastrophe: Using imagery to assess Harvey’s impact on Houston,” Planet.org, 2018, planet.com/insights/anatomy-of-a-
catastrophe/.
42
Santosh Subramanian, “Infrastructure monitoring: A case study,” SkyMap Global, June 19, 2018, skymapglobal.com/infrastructure-monitoring-
with-satellite-imagery-a-case-study/.
Object detection using satellite imagery
Showing flooded roads (in red) and non-flooded roads (in orange) in Houston,
following the passage of Hurricane Harvey in 2017
SOURCE: PlanetLabs; McKinsey Global Institute analysis
4. BOTTLENECKS TO OVERCOME
While the social impact of adding AI to the mix of solutions targeting the world’s most
pressing challenges is potentially very large, some AI-specific bottlenecks will need to be
overcome if even some of that potential is to be realized. In all, based on interviews with
social domain experts and AI researchers, we identified 18 potential bottlenecks that could
stand in the way of successful AI deployments for social good. Having identified the 18, we
then tested them on use cases in our library. Exhibit 11 shows these bottlenecks grouped in
four categories of criticality.
Exhibit 11
Bottlenecks limiting the use of AI for societal good can be grouped into four categories.
Critical for most domains
 Data accessibility Critical for select cases1
 Data quality  Data volume Contextual challenges
 High-level AI expertise  Data labeling  Data availability

availability
 AI practitioner talent  Data integration Potential bottleneck
 High-level AI expertise availability
accessibility  Access to technology  Availability of
 AI practitioner talent organizations that can
 Regulatory limitations accessibility  Organization scale AI deployment
receptiveness
 Organization  Access to computing  Access to software
deployment efficacy capacity  Public resistance libraries and other tools
1 Bottlenecks that are critical for some domains as a whole or for individual use cases within those domains.
NOTE: This list of bottlenecks was derived from interviews with social domain experts and AI researchers and tested against our use case library.
The most significant bottlenecks we identified, and which we describe in detail in this
chapter are data accessibility, a shortage of talent to develop AI solutions, and “last mile”
implementation challenges.43
The availability of computing infrastructure on which to train AI models is no longer as

significant a barrier as it once was. Competition among vendors has greatly reduced the
cost of cloud-based computing capacity, including graphics processing units (GPUs),
and increased access to affordable computing capacity.44 Cloud-based computation as a
service is now widely accessible and requires relatively small investment on a pay-as-you-
go basis. For example, fast.ai has shown that, using its framework and a cloud computing
instance, a model can be trained on the ImageNet corpus of images to create a model with a
93 percent accuracy rate to identify objects in images for only $25.45
43
Other AI research we have conducted suggests that AI’s application could be unequal, with a divide
opening up between countries and companies. See Jacques Bughin and Nicolas van Zeebroeck, “The
promise and pitfalls of AI,” Project Syndicate, September 6, 2018, https://www.project-syndicate.org/
commentary/artificial-intelligence-digital-divide-widens-inequality-by-jacques-bughin-and-nicolas-van-
zeebroeck-2018-09.
44
Access to GPU capacity can nonetheless be tight in the short term.
45
Jeremy Howard, Training Imagenet in 3 hours for $25; and CIFAR10 for $0.26, fast.ai, May 2, 2018, fast.
ai/2018/04/30/dawnbench-fastai/.
Similarly, advances in open-source AI libraries have substantially simplified many of the
tasks previously requiring significant programming skill and knowledge of AI algorithms.
Software libraries such as PyTorch, TensorFlow, and fast.ai are available to a global public on
an open-source basis, and the ease with which they can be deployed and used continues to
improve over time.
DATA NEEDED FOR SOCIAL IMPACT USES MAY NOT BE EASILY ACCESSIBLE
FOR NGOS AND OTHERS IN THE SOCIAL SECTOR
Data barriers we identified are primarily tied to issues of the accessibility, quality, and
quantity of data. Data volume issues—essentially, the difficulty of obtaining sufficiently
large amounts of rich data, including input from video, image, and text with which to train
algorithms—remain an obstacle. However, as the AI field continues to advance, and more
models are pretrained with large amounts of data in various domains, the incremental
amount of data required to solve individual problems can often be reduced. For example,
transfer learning, in which an AI model is trained to accomplish a certain task and then
applies that learning to a similar but distinct activity, will reduce the requirement for massive
training data volume for each individual activity.
Data accessibility remains a significant challenge. To resolve it will require a willingness

by both private- and public-sector organizations to share their data. Much of the data
that are essential or useful for social good applications are in private hands or in public
institutions that might not be willing to share their data. Organizations with large data sets
include telecommunications and satellite companies, social media platforms, financial
institutions (for details such as credit history), hospitals, doctors and other health providers
(medical information), and governments (including private individuals’ tax information).
Obtaining access to these types of data sets by social entrepreneurs and NGOs can be
difficult because of regulations on data use, privacy and other concerns around risks, and
bureaucratic inertia. The data may have business value and be commercially available
for purchase. Given the challenges of distinguishing between social use and commercial
use, the price may be too high for NGOs and others wanting to deploy the data for societal
benefit. Owners of the data may also be unwilling to dedicate the time and resources
needed to clean and share data.
Indeed, data quality can be its own challenge. If the data used to build and run accurate and
fair AI models are not representative or of sufficiently high quality—for example, if some data
are partially missing, outdated, or contain errors—this can be a serious risk, as we discuss
in more detail in Chapter 5. Exhibit 12 shows how our use cases map to data accessibility
and AI capability maturity across domains. For example, the crisis response and public and
social sector management domains require mostly mature capabilities, while solutions for
equality and inclusion are still developing.
Exhibit 12
A mapping of use cases to data accessibility and AI capability maturity shows that the impact is highly varied.
Sum of usage frequency score of use cases mapped to level of data accessibility and capability maturity barriers1
Crisis response
High
Data
accessibility
Low
Deve- More
loping developed
AI capability maturity
Economic empowerment Education Environment
Equality and inclusion Health and hunger Information verification and

validation
Infrastructure Public and social sector Security and justice
1 Capability maturity refers to the combination of the level of talent needed to build a solution and the level of maturity of the capabilities required.
THE EXPERT AI TALENT NEEDED TO DEVELOP AND TRAIN AI MODELS IS IN
SHORT SUPPLY IN THE SOCIAL SECTOR
While just over half the use cases in our library can leverage solutions that can be created
by talent with relatively lower levels of AI experience, the remaining use cases have added
complexity due to a combination of factors, depending on the specific case. These need
high-level AI expertise, that is, people who may have PhDs or considerable experience with
the technologies—and who are in short supply.
For the use cases requiring less AI expertise, the solution builders needed are data
scientists or software developers with AI experience but not necessarily high-level expertise.
This is because some AI capabilities are less technically complex to deploy. Most of these
use cases rely on single modes of data input.
Problem complexity increases significantly where use cases rely on several AI capabilities to
work together cohesively and require multiple different data-type inputs. Making progress
in developing solutions to these cases will thus require high-level talent, for which demand
far outstrips supply and competition is fierce.46 Reflecting this supply-demand imbalance,
compensation for AI practitioners with high-level expertise is very high in the commercial
sector, straining the ability of social-sector organizations to recruit. Some areas of research
tend to attract talent more than others; for example, ambitious “moonshot” challenges may
be more attractive to some experts than practical AI applications.
“LAST MILE” IMPLEMENTATION CHALLENGES ARE ALSO A SIGNIFICANT

BOTTLENECK FOR AI DEPLOYMENT FOR SOCIAL GOOD
Even when high-level AI expertise is not required, NGOs and other social-sector
organizations can still face technical problems in deploying and sustaining AI models for
which they will need continued access to some level of AI-related skills. The talent required
might include engineers who can maintain or improve on the models, data scientists who
can extract meaningful output from AI models, and so on. Failed handoffs will occur when
solutions providers only set up the solution and then disappear without ensuring that a
sustainable plan is in place.
One example cited by experts we interviewed concerned an AI-powered research tool

built for a federal agency. The agency did not understand the technical documentation
describing how to install and run the tool, and the tool became unused “shelfware” once the
agency’s contract with the private research group that devised the solution expired. To avoid
such problems, the organization deploying the AI solution could either ensure that it has the
capability to maintain and operate a tool in the long term or contract for technical support,
including updating and maintaining the model.
Organizations may also have difficulty interpreting AI model results. Even if a model achieves
a desired level of accuracy on test data, new or unanticipated failure cases can often
emerge in real-life scenarios. Without an understanding of how the solution works, which
may require data scientists or “translators”—that is, people who can bridge the technical
expertise of data engineers and data scientists with the operational expertise of frontline
managers—the NGO or other implementing organization may be overly trusting of the model
results, even though most AI models cannot perform accurately all the time. The models are
often described as “brittle,” that is, failing when inputs stray in specific ways from the data
sets on which the models were trained.47
Cade Metz, “Tech giants are paying huge salaries for scarce A.I. talent,” New York Times, October 22, 2017.
46
See Nicolaus Henke, Jordan Levine, and Paul McInerney, “You don’t have to be a data scientist to fill this
47
must-have analytics role,” Harvard Business Review, February 5, 2018, hbr.org/2018/02/you-dont-have-to-

be-a-data-scientist-to-fill-this-must-have-analytics-role.
Other “last mile” implementation challenges that are not technical in nature may seem more
mundane but could be equally important and difficult to resolve. Change management of
organizations, including adapting processes to be able to integrate AI-powered solutions, is
one of the largest obstacles for both commercial and noncommercial deployment. Another
potential hurdle is a lack of critical infrastructure. At the individual level, for example, finding
an AI-powered solution that uses smartphones will be of little use to people who do not
have them. Likewise, in areas without electricity, even simple devices that require recharging
will not achieve the desired effect. Funding could help overcome such limitations but may
need to be on a very large scale. For example, 90 percent of the 215 million visually impaired
persons worldwide who could benefit from environment understanding software on
smartphones live in developing countries, where smartphone penetration is low. Obtaining
financial commitments for these investments, reductions in the costs of technology, or both
will be needed to solve this challenge.
5. RISKS TO BE MANAGED
Risks associated with AI are becoming an increasingly important area of research,
especially (but not exclusively) in the field of ethics applied to AI. AI’s tools and techniques
can be misused by authorities and others with access to them, and principles for their use
will need to be established (see Box 2, “A growing body of research on the ethics of AI”). In
the worst case, AI solutions can unintentionally harm the very people they are supposed
to help.
Some AI-related risks spring from the way AI models are trained. For example, if data sets
used to train algorithms are based on historical data that incorporate racial or gender bias
(even unintentionally, resulting solely from sampling bias), the applications derived from the
algorithms will perpetuate and may aggravate that bias.
In general, risks relating to AI for social good are quite similar to those for more commercial
uses. One of the biggest risks is that AI’s tools and techniques can be misused by authorities
and others with access to them; malicious uses can harm individuals, organizations, and
society at large.48 AI can be used maliciously to threaten the physical and emotional safety of
individuals, as well as their digital safety, financial security, and equity and fair treatment. For
organizations, malicious use often implies reputation and legal compliance risks, although
there may be fewer commercial risks than those that could potentially harm for-profit
companies. For society, malicious uses of AI could threaten national security, economic
stability, political stability, labor market stability, and infrastructure.
One general difference between the risks associated with commercial and noncommercial
purposes concerns the effects of labor displacement on workers and the workforce. Much
of the public debate around AI in the for-profit world focuses on the potential for labor
displacement. An examination of our use case library suggests that this may be less of a risk
in many of the applications of AI for social good. Indeed, as outlined in the earlier discussion
of bottlenecks, AI’s application for social good uses tends to be hampered by a shortage of
personnel with the skills and technical know-how required. If AI were to increase access to
health or education, for example, this could serve as a potential net boost to employment of
doctors, nurses, or teachers.
An analysis of our use case library found four main categories of risk that are particularly
relevant when leveraging AI solutions for social good, as we describe below. They are bias
and fairness, privacy, safe use and security, and “explainability”—being able to identify the
feature or data set that leads to a particular decision or prediction. The types of risk and their
magnitude differ considerably from case to case. All domains carry some level of risk, but
in general our analysis suggests that domains in which data are sensitive and predictions
identify individuals—for example economic empowerment, education, equality, health, and
security—face the highest magnitude of risk (Exhibit 13).
Inaccurate AI poses high risks in other domains, such as crisis response. For example,
erroneous predictions of the location of missing persons could prove fatal. The scoring in the
exhibit is based on perspectives from our interviews with domain experts and AI specialists.
To come up with the risk profiles, we tagged individual use cases from low to high for each
of the five categories: the risk of bias; the risk of privacy violation; the risk of unsafe use of
the AI solution; the level of explainability that is required to reduce or mitigate risks; and
considerations of the risk of negative impact to the workforce and workers.
See Miles Brundage et al., The malicious use of artificial intelligence: Forecasting, prevention, and mitigation,
48
Future of Humanity Institute, February 2018.
Box 2. A growing body of research on the ethics of AI
As AI capabilities develop and are increasingly deployed in both the
commercial and noncommercial worlds, the ethics surrounding use of artificial
intelligence have spurred a growing body of research.
Donors including technology company executives have stepped up funding

for major scholarship programs in the past months, including at MIT, Harvard’s
Berkman Klein Center for Internet and Society, Stanford University, and the
University of Toronto, to study the implications of AI, including how it will affect
people’s lives and serve humanity.1
Technical professional bodies and international bodies including the World

Economic Forum and the UN’s International Telecommunication Union have
also focused attention on societal uses and potential abuses of AI.2 The
ITU has initiated an AI for Good conference that ties AI outcomes to the UN
Sustainable Development Goals.
Several nonprofit organizations also do important work on ethical uses

of AI, including the Partnership on AI, which brings together academics,
researchers, civil society organizations, and companies building and utilizing
AI technology to better understand AI’s impacts.3 And major technology
companies including Microsoft and Google have articulated their philosophy
and practices on AI; Google’s first principle is “be socially beneficial.”4
One recurring concern is the impact of AI on work, and whether it could create
a gap even wider than the existing digital divide, not just between high- and
low-income workers but also between countries.5 While proposed solutions to
these potential gaps differ, there is broad agreement that a substantial change
in the way we educate our children is needed to prepare them for the future of
work, and that large-scale retraining will be required for midcareer workers as
skill requirements shift.6
Safety is a concern in social impact uses, just as it is in commercial use, for

example with self-driving cars, since accidents can create ethical dilemmas
in life-and-death situations. In healthcare-related uses, errors and decisions
can systematically burden groups of people if solution design is flawed. An
example of this is a diagnostic tool that may overall be better than a human
doctor but that makes significantly more diagnosis errors when the patient is a
member of an ethnic minority.
1
Steve Lohr, “M.I.T. plans college for artificial intelligence, backed by $1 billion,” New
York Times, October 15, 2018; Maria Di Mento, “Donors pour $583 million into artificial-
intelligence programs and research,” Chronicle of Philanthropy, October 15, 2018.
2
See, for example, How to prevent discriminatory outcomes in machine learning, World
Economic Forum, March 12, 2018. The Institute for Electrical and Electronics Engineers
has announced the approval of three standards projects inspired by the work of its Global
Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems. The
new standards projects directly reflect the institute publication. See Kyarash Shahriari and
Mana Shahriari, IEEE standard review—Ethically aligned design: A vision for prioritizing
human well-being with artificial intelligence and autonomous systems, IEEE Canada
International Humanitarian Technology Conference, Toronto, Canada, July 21–22, 2017.
3
McKinsey is a member of the Partnership on AI, partnershiponai.org/.
4
See Microsoft AI principles, microsoft.com/en-us/ai/our-approach-to-ai, and The Keyword,
“AI at Google: Our principles,” blog entry by Sundar Pichai, June 7, 2018, blog.google/
technology/ai/ai-principles/.
5
Erik Brynjolfsson and Andrew McAfee, The Second Machine Age, New York: W. W. Norton,
2016.
6
See Jobs lost, jobs gained: Workforce transitions in a time of automation, McKinsey Global
Institute, December 2017; Skill shift: Automation and the future of the workforce, McKinsey
Global Institute, May 2018.
Box 2. A growing body of research on the ethics of AI (continued)
One of the most intense areas of research concerns bias. While algorithms
and AI can improve on human decision making to reduce bias in some cases,
the models can also end up reflecting intrinsic bias contained in the data
sets used to train them.7 For example, a recent report by MIT and Stanford
researchers showed how some commercially available facial recognition
systems perform poorly when applied to faces of women and people of
color.8 A study by Harvard’s Berkman Klein Center shows how some popular
AI use cases could both positively and negatively impact specific aspects
of the Universal Declaration of Human Rights, depending on how they
are implemented.9
Given the potential for negative impacts, some call for AI to be regulated
by government. The benefits of regulation will need to be balanced against
potential unintended consequences, including a dampening of innovation.10
Many governments have produced formal AI strategies to promote the

use and development of AI, and most include specific sections on the
development of ethics regulations.11
History shows that many new technologies have created misgivings, dating
back at least to the 16th century with the invention of the stocking frame.12 As
has happened in the past, all stakeholders, from civil society to AI researchers
to government, will need to collaborate to define what is—and is not—
acceptable, if the positive benefits that the technologies offer are to become
a reality.
7
One example of AI being less biased than humans is in bail decisions; one paper has found
that algorithmic decision making in bail decisions could reduce jail populations by 42 percent
with no increase in crime rates. Jon Kleinberg et al., Human decisions and machine
predictions, NBER working paper number 23180, February 2017.
8
Joy Buolamwini and Timnit Gebru, “Gender shades: Intersectional accuracy disparities
in commercial gender classification,” Proceedings of Machine Learning Research, 2018,
Volume 81, http://proceedings.mlr.press/v81/buolamwini18a.html.
9
Filippo Raso et al., Artificial intelligence and human rights: Opportunities and risks, Berkman
Klein Center for Internet and Society at Harvard University, September 25, 2018, https://
cyber.harvard.edu/sites/default/files/2018-09/2018-09_AIHumanRightsSmall.pdf.
10
Christopher Fonzone and Kate Heinzelman, “Should the government regulate artificial
intelligence? It already is,” The Hill, February 26, 2018, https://thehill.com/opinion/
technology/375606-should-the-government-regulate-artificial-intelligence-it-already-is.
11
See for example, Cédric Villani, For a meaningful artificial intelligence: Towards a French and
European strategy, March 2018, aiforhumanity.fr/pdfs/MissionVillani_Report_ENG-VF.pdf;
Will Hurd and Robin Kelly, Rise of the machines: Artificial intelligence and its growing impact
on U.S. policy, US House of Representatives, Subcommittee on Information Technology,
September 2018, oversight.house.gov/wp-content/uploads/2018/09/AI-White-Paper.pdf;
Pan-Canadian artificial intelligence strategy, CIFAR, https://www.cifar.ca/ai/pan-canadian-
artificial-intelligence-strategy; Tim Dutton, “An overview of national AI strategies,” Medium,
June 28, 2018, medium.com/politics-ai/an-overview-of-national-ai-strategies-2a70ec6edfd.
12
In 1589, England’s Queen Elizabeth I refused to grant a patent to a stocking frame invented
by William Lee because she was supposedly concerned about the effect on hand knitters.
R. L Hills, “William Lee and his knitting machine,” Journal of the Textile Institute, July 1989,
Volume 80, Number 2.
Exhibit 13
Risk profiles of social impact domains differ, with some of the biggest Risk profiles of different social
potential risks found in the health and hunger domain. impact domains, level of risk
A Bias B Privacy violation C Unsafe use
Low Medium High
D Level of explainability required E Potential for workforce displacement
A B C D E A B C D E
Crisis response Health and hunger
Disease outbreak Treatment delivery
Migration crises Prediction and prevention
Natural and man-made Treatment and long-term care
disasters
Mental wellness
Search and rescue
Hunger and malnutrition
Economic empowerment
Information verification and validation
Agricultural quality and yield
False news
Financial inclusion
Polarization
Initiatives for economic
growth
Infrastructure
Labor supply and demand
matching Energy
Real estate
Education Transportation
Access and completion of Urban planning
education
Water and waste
Maximizing student management
achievement
Teacher and administration Public and social sector management

productivity
Effective management of
public sector
Environment
Effective management of
Animal and plant social sector
conservation
Fundraising
Climate change and
adaptation Public finance management
Energy efficiency and Services to citizens
sustainability
Land, air, and water Security and justice

conservation
Harm prevention
Equality and inclusion Fair prosecution
Accessibility and disabilities Policing
Exploitation
A B C D E
Marginalized communities
A B C D E
NOTE: The risk scoring in this exhibit is based on interviews with domain experts and AI specialists. We tagged individual use cases in our library from low to
high in each domain and then aggregated the score to the issue type level. Our use case library continues to evolve and these findings should not be read as
exhaustive.
BIAS THAT LEADS TO UNFAIR OUTCOMES IS A RISK, INCLUDING WHEN
ALGORITHMS ARE TRAINED ON BIASED HISTORICAL DATA
Bias in AI may perpetuate and aggravate existing prejudices and social inequalities, affecting
already vulnerable populations and potentially amplifying existing cultural prejudices. This
bias may come about through problematic historical data, including sample sizes that are
not representative or are inaccurate. For example, AI-based risk scoring for criminal justice
purposes may be trained on historic criminal data that include biases (for example, where
African Americans are unfairly labeled as high risk). As a result, risk scores from AI would
continue to perpetuate this bias. The use of an AI-based sentencing solution containing
biased data could have a life-changing impact on individuals. Algorithm and data choices
can also introduce bias, for example by producing unequal outcomes for certain groups
because of weaker statistical correlations for such groups. Some AI applications already
highlight large disparities in accuracy depending on the data used for training algorithms; for
example, examination of facial-analysis software shows error rates of 0.8 percent for light-
skinned men and 34.7 percent for dark-skinned women.49
One key source of bias can be poor data quality. An example of this type of data quality
issue arises when using data on past employment records to identify future candidates.
Since many employers and industries have historically had disparate distributions of gender
and race in their corporate hierarchies, using purely historical data to train AI models could
perpetuate discrimination. For example, an AI-powered recruiting tool used by one tech
company was abandoned recently after several years of trials because it appeared to show
systematic bias against women, resulting from patterns in training data from years of hiring
history.50 To counteract such bias, skilled and diverse data science teams need to take into
account potential issues in the training data or otherwise sample intelligently from them.
BREACHING PRIVACY OVER PERSONAL INFORMATION COULD CAUSE HARM

Privacy concerns about sensitive personal data are already rife, and the ability to assuage
these worries could help public acceptance of widespread AI use, for for-profit as well as
for nonprofit ends. The risk here is that information about individuals, such as financial, tax,
or health records, could be made accessible through porous AI systems to those without a
legitimate need to access them, causing embarrassment and potentially harm. For example,
data on movie recommendations and viewing habits combined with movie database data
could potentially identify sexual and political information, which in some countries could put
individuals in physical or psychological danger.51 Another example is when a mental health
screening solution has access to private information such as emails and health records.
The underlying data are already extremely sensitive, and if mental health status becomes
public, the exposure could significantly hurt individuals and set back their recovery, given
the continuing stigma surrounding mental health issues.
49
Joy Buolamwini and Timnit Gebru. “Gender shades: Intersectional accuracy disparities in commercial gender
classification,” Proceedings of Machine Learning Research, 2018, volume 81.
50
Jeffrey Dastin, “Amazon scraps secret AI recruiting tool that showed bias against women,” Reuters, October
10, 2018, reuters.com/article/amazoncom-jobs-automation/rpt-insight-amazon-scraps-secret-ai-recruiting-
tool-that-showed-bias-against-women-idUSL2N1WP1RO.
51
Ryan Singel, “Netflix spilled your Brokeback Mountain secret, lawsuit claims,” Wired, December 17, 2009,
wired.com/2009/12/netflix-privacy-lawsuit/.
SAFE USE AND SECURITY ARE ESSENTIAL FOR SOCIAL GOOD USES OF AI
Ensuring that AI applications are safe and responsible for human use is an essential
prerequisite for widespread deployment for social aims. Seeking to further social good with
technologies that are dangerous for humans would not only run counter to their core mission
but could also spark a backlash, given the potentially large numbers of people involved. For
technologies that could affect life and well-being, it will be important to have in place safety
mechanisms that comply with existing laws and regulations. For example, if AI misdiagnoses
patients in hospitals without a safety mechanism in place, particularly if the systems are
directly connected to treatment processes, the outcomes could be catastrophic. Another
example is in predictive maintenance: if an AI model fails to recognize that a component of
a bus or train needs to be replaced, that could result in a major accident. The framework for
accountability and liability for harm done by AI is still evolving.
DECISIONS MADE BY COMPLEX AI MODELS WILL NEED TO BECOME MORE

READILY EXPLAINABLE
Explaining in human terms the results from large and complex AI models remains one of
the key challenges to achieving user and regulatory acceptance.52 Opening the AI “black
box” to show how decisions are made, which factors are decisive and which are not, will
be important for social use of AI. This is especially true when working with stakeholders,
including NGOs, who require a basic level of transparency of use and will likely want to
provide individuals with clear explanations. Not all AI models are complex, and work is
under way to design neural network models that are more readily explainable. Ensuring
explainability is especially important in use cases relating to any decision making about
individuals, and in particular for cases related to justice and criminal identification, when an
accused person needs to be able to appeal a decision in a meaningful way. While a black
box solution that predicts recidivism may provide accurate results, it could be difficult to
interact with and to understand exactly why the solution identified a person as a threat or
a potential reoffender. Use of AI models in cases such as identifying tax fraud could be
subject to the similar expectations and/or requirements and therefore need a high level of
explainability—or run the risk of being considered unusable.
Michael Chui, James Manyika, and Mehdi Miremadi, “What AI can and can’t do (yet) for your business,”
52
McKinsey Quarterly, January 2018.
MITIGATING RISKS
Effective mitigation strategies typically involve “human in the loop” interventions, in which
humans are involved in the decision or analysis loop to validate models and double-check
results from AI solutions. They may require cross-functional teams, including domain
experts, engineers, product managers, user experience researchers, legal professionals,
and others, to test, assess, and flag possible unintended consequences. This must be done
on an ongoing basis.
Human analysis of data used to train models may be able to identify issues such as bias
and lack of representation. Fairness and security “red teams” could carry out solution tests,
and in some cases third parties could be brought in to test solutions using an adversarial
approach. Some methods of mitigating this bias demonstrated by university researchers
include sampling the data with recognition of the inherent bias and creating synthetic data
sets based on known statistics. The appropriate solution depends on the specific use case.
For example, if the objective of the AI solution is to detect the incidence of stroke in a patient
and identify the subtype of stroke, then a promising solution may be to oversample minority
class data to overcome initial imbalances.53 In the case of predictions of sexual assault
prevalence based on geography, a synthetic data set may be the better option because
available sets of real data may be too small for effective sampling.
Guardrails to prevent users from blindly trusting AI can be put into place. In medicine, for
example, misdiagnoses can be devastating to patients, whether through false positive
results that cause distress, wrong or unnecessary treatment or surgeries, or, even worse,
false negatives, which could lead to patients missing diagnoses until a disease has reached
terminal stage. Ensuring education for patients and mandating that a disclaimer be read
every time an AI solution gives a result to a patient could be helpful.
Technology may also be able to find some solutions to these challenges, including
explainability. For example, nascent approaches to model transparency include local
interpretable model-agnostic (LIME) explanations, which attempt to identify those parts
of input data on which a trained model relies most heavily to make predictions. Similarly, a
system built by DeepMind that recommends treatments for eye diseases provides doctors
with an indication of which aspects of a medical scan prompted the diagnosis.
Yizhao Ni et al., “Towards phenotyping stroke: Leveraging data from a large-scale epidemiological
53
study to detect stroke diagnosis,” PLoS ONE, 2018, Volume 13, Number 2, journals.plos.org/plosone/
article?id=10.1371/journal.pone.0192586.
6. SCALING UP THE USE OF AI FOR SOCIAL GOOD
The potential for AI to be used for social good is compelling, given both the large numbers of
people who could be helped and the otherwise intractable problems that may be solved. As
with any technology deployment for social good, the scaling up and successful application
of AI for societal benefit will depend on the willingness of relevant groups of stakeholders,
including collectors and generators of data as well as governments and NGOs, to engage.
These are still early days for AI deployment for social use, and considerable progress will be
needed before the vast potential becomes a reality.
For now, AI capabilities are being tested and deployed, and they already show promise
across a range of domains. The technology itself is advancing rapidly. As we have sought
to describe in this paper, the next steps will need to focus on scaling up AI solutions and
overcoming the bottlenecks and market failures that are holding it back for now.
Public- and private-sector players all have a role to play. Industries that control or generate
data, including satellite and telecommunications companies and major tech firms with
social media or other platforms, can accelerate their involvement. That in turn will require
agreements on how to handle the cross-collaboration with data in a way that is protective of
users and takes into account ethics concerns. NGOs and other social-sector organizations
will also need to raise their ability to work with AI solutions. All this leaves ample room for
philanthropists and others to help build the capabilities needed to increase AI deployment
for social good.
In this final chapter, we suggest some areas in which stakeholders could make a meaningful
contribution to further the use of AI for the benefit of society, especially in overcoming the
key impediments of data accessibility, talent, and implementation.
IMPROVING DATA ACCESSIBILITY FOR SOCIAL IMPACT CASES WILL REQUIRE

ACTIVE PARTICIPATION OF DATA COLLECTORS AND GENERATORS
A wide range of stakeholders owns, controls, collects, or generates the data that could
be deployed for AI solutions. Governments are among the most significant collectors of
information, which can include tax, health, and education data, but private companies—
including satellite operators, telecommunications firms, utilities, and technology companies
that run digital platforms, as well as social media sites and search operations—also collect
massive volumes of data. These data sets may contain highly confidential personal data
that cannot be shared without being anonymized. But, in the case of private operators, data
sets may also be monetized by their collectors, and thus not available for pro-bono social
good cases.
Overcoming this accessibility challenge will likely require a call to make specific data sets
more readily available for well-defined societal initiatives.
Data collectors and generators will need to be encouraged, and possibly mandated, to open
access to a subset of their data when it could be in the clear public interest. Such dual use
is already starting to happen in some areas, and new business models are being tested that
may facilitate data sharing. For example, many satellite data companies participate in the
International Charter on Space and Major Disasters, which commits them to open access
to satellite data during emergencies such as the September 2018 tsunami in Indonesia
and Hurricane Michael, which hit the US East Coast in October 2018.54 Other data sharing
initiatives with private companies are also being worked on, including OPAL, an open
54
See disasterscharter.org. For other examples of data collaboration, see Stefaan G. Verhulst and Andrew
Young, “How the data that Internet companies collect can be used for the public good,” Harvard Business
Review, January 23, 2018.
algorithm collaboration between the World Economic Forum, MIT Media Lab, Orange,
and others to derive aggregated insights from a company’s data without data leaving the
company’s server. If proven successful, this could be a powerful tool in unlocking private
data for social causes. It started in 2017 with pilots in Colombia and Senegal in partnership
with their governments and two telecommunications operators.55
Close collaboration between NGOs and data collectors and generators could also
help to facilitate this push to make data more accessible. Funding will be required from
governments and foundations for initiatives to record and store data that could be used for
social ends.
Even if the data are to become accessible, using them presents challenges. Continued
investment will be needed to support high-quality data labeling. And multiple stakeholders
will need to commit to storing data that can be accessed in a coordinated way and to use
the same data standards where possible to ensure seamless interoperability.
Where data sets are freely available, moreover, the data may not have sufficiently large
volume for deep learning, which will restrict application of these capabilities—although with
advances in transfer learning and pretrained models, some capabilities may not need data
volumes as large as would previously have been the case.
Issues of data quality as well as potential bias and fairness will also need to be addressed
if the data are to be deployed usefully. Transparency will be a key for the latter issues. A
deep understanding of the data, their provenance, and their characteristics will need to be
captured so that others using the data set are aware of potential flaws.
This is a long list, and it is likely to require collaboration among companies, governments,
and NGOs to set up regular data forums in each industry to work on data availability,
accessibility, and connectivity issues, ideally setting global industry standards and
collaborating closely on use cases, to ensure that implementation becomes feasible.
OVERCOMING AI TALENT SHORTAGES IS ESSENTIAL FOR SOLVING

TECHNICAL CHALLENGES AND IMPLEMENTING AI SOLUTIONS
While solving data challenges will require goodwill and careful coordination, the talent-
related challenges we have identified with respect to applying AI for societal good are
potentially long-term issues that start with changes in education systems.
The talent challenge is twofold: a shortage of workers with high-level AI expertise—including

(but not limited to) PhD-level experience—who are able to develop and train more complex
models and a lack of data scientists, translators, and other AI practitioners who can
become involved in the implementation phase. The social sector has an especially difficult
challenge in hiring and retaining both types of talent, given the high salaries that experienced
practitioners can earn at commercial companies.
The long-term solution to these challenges will be to recruit more students to be

trained in AI. That could be spurred by significant increases in funding, both grants and
scholarships, for tertiary education and PhDs in AI-related fields. With the high salaries that
AI expertise commands today, the employment market has signaled a surge in demand for
such education.
Sustaining and increasing current training opportunities would be helpful. They include “AI
residencies”—one-year training programs at corporate research labs (such as OpenAI,
Google, Facebook, and Microsoft)—as well as shorter-term AI boot camps and academies
for midcareer professionals. An advanced degree typically is not a prerequisite for these
See opalproject.org/about-opal.
55
programs, which can train participants in the practice of AI research without having to spend
years in a PhD program.
In the absence of experienced professionals in the social sector, companies with AI

talent could play a major role in focusing more efforts on AI solutions that have a social
impact. For example, they could encourage employees to volunteer and support or coach
noncommercial organizations to adopt, deploy, and sustain high-impact AI solutions.
Companies and universities with AI talent could also allocate some of their research
capacity to new AI capabilities or solutions that focus on societal benefit but are unable to
attract people with the requisite skills.
Overcoming the shortage of talent able to manage AI implementations will likely require
government and educational providers to work with companies and social-sector
organizations to develop more free or low-cost online training courses. The courses would
help those with a basic understanding of computer science acquire the skills needed to
pull the data together and be on the frontline of AI implementation (see Box 3, “Ten steps to
AI deployment for social good, and some barriers to overcome: A checklist”). Foundations
could provide funding for such initiatives.
One idea is to create prizes and other programs to spur innovation for creative solutions.
Incentives applied to social good may accelerate the learning process for NGOs, and
monetary prizes can encourage new entrants to the field.
Task forces of tech and business translators from government, corporations, the freelance
ranks, and social organizations could be established to help teach NGOs about AI with
relatable case studies. Beyond coaching, these task forces could potentially help NGOs
scope potential projects, support deployment, and plan sustainable road maps. As with
all technology deployment, success or failure can depend on the existence of necessary
infrastructure as well as the organization’s ability to adapt and raise sufficient funding. While
these and other implementation challenges are not directly AI-related, they should not be
overlooked when attempting to deploy AI for social good.
•••
AI could be the moon landing of our generation, an inspiring scientific leap forward that
brings huge benefits to mankind. From the modest library of use cases that we have begun
to compile, we can already see tremendous potential to address the world’s most important
challenges, from hunger to disease to education. Hundreds of millions of people could
benefit, from autistic children to earthquake survivors. While the potential to do good with
AI is impressive, turning that potential into reality on the scale it deserves will require focus,
collaboration, goodwill, funding, and a determination among many stakeholders to work
for the benefit of society. Many gaps remain. Some are technological and can be overcome
if recent rapid scientific breakthroughs continue. Others relate to talent and the shortage
of humans who can develop and train these systems and make them work on the ground.
We are only just setting out on this journey. Reaching the destination will be a step-by-step
process of confronting barriers and obstacles. We can see the moon—but getting there
will require more work and a solid conviction that the goal is worth all the effort, for the sake
of everyone.
Box 3. Ten steps to AI deployment for social good, and some barriers to overcome:
A checklist
This is a checklist for social-sector organizations thinking ;;
6. Required data set can be accessed: Data released
about deploying AI solutions, including for the first time. It for public use or to the solution builder.
highlights both the actions that may smooth the path to a Related potential barrier to overcome:
successful deployment and the barriers that will need to data accessibility.
be overcome. While there can be entry points anywhere
Related potential risks to consider and mitigate:
along this list, the social-sector organization will need to
data and model bias, data privacy violations.
address steps 1, 2, and 3 first, and then the remaining
items on the checklist, to ensure that AI can be deployed.
;;
7. Required data quality ensured and can be used:
Data preprocessed into standard format that can be
readily fed to a training algorithm.
;;
1. Societal problem clearly defined: Specific problem
in a social impact domain identified, with measurable Related potential barriers to overcome: data
objective and requirements for success. quality, data volume.
;; 2. Translation into technical problem achieved: data and model bias, data privacy violations.
Formulation of technical problem structure with
defined data types and evaluation of technical ;;
8. AI model built and trained on available data set.
feasibility completed by technical experts. Model proves significant utility on test set.
Related potential barriers to overcome: availability
;;
3. AI technology and/or data confirmed to be a
and accessibility of talent with high-level AI
bottleneck in the societal problem and decision to
expertise as well as AI practitioners, access to
deploy AI (or not) made: Comparison of value of AI
computing capacity.
to other solutions (for example, policy or changing
stakeholder incentives) and evaluation of potential
unsafe use of solution, inability to meet
risks and mitigations completed to determine whether
explainability level required.
AI deployment is the right solution.
Related potential barriers to overcome:
;;
9. Trained AI model proved value on the ground:
organization receptiveness, access to technology Model used in the target environment and
for users, organization deployment efficiency, demonstrated sufficient value to drive large-scale
regulatory limitations. adoption by actor (for example, NGO).
Related potential risks to consider and mitigate: Related potential barriers to overcome: availability
data and model bias, data privacy violations, and accessibility of talent with high-level AI
unsafe use of solution, inability to meet expertise as well as AI practitioners, access to
explainability level required. technology for users, organization receptiveness,
organization deployment efficiency,
;;
4. Organization to deploy solution committed: Social-
regulatory limitations.
sector organization with resources to deploy AI
solution at scale has committed to solving the defined
unsafe use of solution, inability to meet
societal problem.
explainability level required.
Related potential barriers to overcome:
organization receptiveness, access to technology
;;
10. Organization has built required technical
for users, organization deployment efficiency, capabilities: Committed organization or organizations
regulatory limitations. have hired or trained required technical capabilities
to run and maintain the AI model independently
;;
5. Required data set available or can be generated:
and sustainably.
The required data exists in some form, and the data
Related potential barriers to overcome: AI
set can feasibly be put together.
practitioner availability and accessibility,
Related potential barriers to overcome: data
organization deployment efficiency.
availability, data integration, data volume.
data and model bias, data privacy violations.
ACKNOWLEDGMENTS
This “Notes from the AI frontier” discussion paper is and Daphne Luong of Google AI, and Micah Berman,
part of an ongoing series of publications that explores Charina Chou, Andrew Dunckelman, Jacquelline Fuller,
aspects of artificial intelligence and its potential impact Brigitte Hoyer Gosselink, Mollie Javerbaum, and
on business, the economy, and society. Previous Justin Steele at Google.org.
papers have looked at AI use cases across sectors and
business functions and AI’s potential impact on the Many others informed our research. We wish to thank:
global economy.56 Aaron Horowitz and Lucia Tian at ACLU; Tess Posner
at AI4All; James Hodson at AI for Good Foundation;
The research was led by Michael Chui, a McKinsey Global Tom Fairburn and Ben Wylie at The Baobab Network;
Institute partner in San Francisco; Martin Harrysson, Paul Duan at Bayes Impact; Amy Guggenheim Shenkan
a McKinsey partner in Silicon Valley; James Manyika, at Common Sense Media; Dean Ramayya Krishnan at
chairman and director of the McKinsey Global Institute Carnegie Mellon University; Jake Porway at Datakind;
and McKinsey senior partner based in San Francisco; Paul van der Boor and Rayid Ghani at Data Science
and Roger Roberts, a McKinsey partner based in Silicon for Social Good; Jim Bildner at DRK Foundation;
Valley. Ashley van Heteren provided guidance and Jeremy Howard at fast.ai; Amy Luers at Future
support. Rita Chung headed the working team, which Earth; Dave Levin at Hala Systems; Anju Khetan at
comprised Tara Balakrishnan, Alexandre Kleis, Pieter Nel, Khan Academy; Alan Donald at Mercy Corps; Professor
and Sigberto Alarcon Viesca. Sandy Pentland at MIT; Will Marshall, Trevor Hammond,
and Tara O’Shea at Planet Labs; Jeremy Pierotti at
We are grateful to colleagues within McKinsey who Sansoro Health; Prem Ramaswami at Sidewalk Labs;
provided valuable advice and analytical support: Professor Marshall Burke at Stanford University;
Shannon Bouton, Stephanie Carlton, Alberto Chaia, Professor Toby Walsh at University of New South
Bertil Chappuis, Ben Cheatham, Kevin Chao, Wales; Professor Milind Tambe at University of Southern
Michael Conway, Matt Craven, Tyler Duvall, California; and Oliver Kharraz at Zocdoc.
Niklas Garemo, Taras Gorishnyy, Lars Hartenstein,
Kimberley Henderson, Solveigh Hieronimus, This discussion paper was edited and produced by
Tarek Mansour, Mekala Krishnan, Marc Krawitz, Peter Gumbel, MGI’s editorial director, editorial production
Pankal Kumar, Dave Levin, Sacha Litman, manager Julie Philpot, senior designers Marisa Carder,
Anu Madgavkar, Dany Matar, Mona Mourshed, Margo Shimasaki, and Patrick White, and Rich Johnson,
Jan Mischke, Keith Otis, Sangeeth Ram, Chloe Rivera, data visualization editor. Rebeca Robboy, MGI director
Matt Rogers, Ben Safran, Hamid Samandari, of external communications, managed dissemination
Dhaval Shah, Jake Silberg, Vivien Singer, Sarah Tucker, and publicity, while digital editor Lauren Meling provided
Nathan Uhlenbrock, Elke Uhrmann-Klingen, support for online and social media treatments.
Jonathan Usuka, Helga Vanthournout, Judy Wade,
Kate Whittington, and Jonathan Woetzel. This report contributes to MGI’s mission to help business
and policy leaders understand the forces transforming the
This independent MGI initiative is based on our own global economy, identify strategic locations, and prepare
research, the experience of our McKinsey colleagues for the next wave of growth. As with all MGI research,
more broadly, and the McKinsey High Tech Practice’s this work is independent, reflects our own views, and has
research collaboration with Google.org and AI not been commissioned by any business, government,
researchers including at Google AI. In particular, we or other institution. We welcome your comments on the
are grateful to Greg Corrado, Kat Chou, Tomas Izo, research at [email protected].
Notes from the AI frontier: Insights from hundreds of use cases, McKinsey Global Institute, April 2018; Notes from the AI frontier: Modeling the
56
impact of AI on the world economy, McKinsey Global Institute, September 2018.
RELATED MGI
AND MCKINSEY RESEARCH
Notes from the AI frontier: Insights from hundreds of use cases
(April 2018)
This paper provides an analysis of more than 400 use cases across 19
industries and nine business functions and highlights the broad use and
significant economic potential of advanced AI techniques.
Notes from the AI frontier: Modeling the impact of AI on the world

economy (September 2018)
Artificial intelligence has large potential to contribute to global economic
activity. But widening gaps among countries, companies, and workers will
need to be managed to maximize the benefits.
Artificial intelligence: The next digital frontier? (June 2017)

This paper discusses AI and how companies new to the space can learn a
great deal from early adopters who have invested billions into AI and are now
beginning to reap a range of benefits.
Jobs lost, jobs gained: Workforce transitions in a time of automation

(December 2017)
This report explains that as many as 375 million workers around the world may
need to switch occupational categories and learn new skills.
A future that works: Automation, employment, and productivity

(January 2017)
This report explores the potential uses of automation in multiple industries and
its effects on employment and productivity.
The age of analytics: Competing in a data-driven world

(December 2016)
Big data’s potential just keeps growing. Taking full advantage means
companies must incorporate analytics into their strategic vision and use it to
make better, faster decisions.
www.mckinsey.com/mgi
Download and listen to MGI podcasts on iTunes or at
www.mckinsey.com/mgi/publications/multimedia/
McKinsey Global Institute
December 2018
Copyright © McKinsey & Company
www.mckinsey.com/mgi
@McKinsey_MGI
McKinseyGlobalInstitute

Notes From The Ai Frontier: Applying Ai For Social Good

Uploaded by

Notes From The Ai Frontier: Applying Ai For Social Good

Uploaded by

NOTES FROM

Michael Chui | San Francisco

MGI research combines the disciplines of economics and management,

MGI is led by three McKinsey & Company senior partners: Jacques Bughin,

The partners of McKinsey fund MGI’s research; it is not commissioned by any

Copyright © McKinsey & Company 2018

2 McKinsey Global Institute ﻿

FOR SOCIAL GOOD Page 1

2. How AI capabilities can be

 Several AI capabilities, primarily in the categories of computer vision and natural

Mapping domains to issue types and use cases in our library.

Number of use cases per domain

SOURCE: McKinsey Global Institute analysis

Existing use cases deploying AI typically related to agriculture;

Most existing AI use cases employ natural language processing (NLP).

Research institutions and organizations working on AI use for social

Many potential AI solutions in this domain have not been implemented

SOURCE: McKinsey Global Institute analysis

Social problems AI capabilities

AI capabilities Social problems

SOURCE: McKinsey Global Institute analysis

 Crisis response. Specific crisis-related challenges, such as responding to natural and

 Economic empowerment. Opening access to economic resources and opportunities,

 Environmental challenges. These include sustaining biodiversity and combating

 Equality and inclusion. Addressing equality, inclusion, and self-determination

 Information verification and validation. The challenge of facilitating provision,

 Infrastructure management. Infrastructure challenges that could provide public

Use cases in our library support the UN Sustainable Development Goals.

UN Sustainable Development Goals1

UN goals AI use case breakdown

Life below water 1

Affordable and clean energy Public and Economic 2

Partnerships for the goals 4

Decent work and economic growth 5

Industry, innovation, and infrastructure 9

Peace, justice, and strong institutions 24

Good health and well-being 29

SOURCE: United Nations; McKinsey Global Institute analysis

When we subsequently mapped these capabilities to domains (aggregating use cases)

COMPUTER VISION, NATURAL LANGUAGE PROCESSING, AND STRUCTURED

Image classification and object detection are powerful capabilities with

Some AI capabilities could be used for societal benefit. NOT EXHAUSTIVE

SOURCE: McKinsey Global Institute analysis

Mapping usage frequency of AI capabilities to ten social impact domains identifies

Usage frequency of AI capability for each domain

Sound detection and

Optical character and

SOURCE: McKinsey Global Institute analysis

Some capabilities, or combinations of capabilities, can open possibilities for the

facial recognition, which is used to identify individuals by their features.

Sound detection and recognition could become increasingly relevant as

Reinforcement learning and content generation, while nascent capabilities,

THE VERTICAL VIEW OF DOMAIN COMPARISONS BRINGS OUT NUANCES IN

55 The 5 domains where many AI capabilities are relevant

Audio Crisis response 5 17

26 Equality and inclusion 8 11

Structured Security and justice 10 16

Public and social sector 2 16

SOURCE: McKinsey Global Institute analysis

Most visually impaired people in the world live in emerging economies.

Moderate and severe 0–0.5 0.5–1.0 1.0–1.5 1.5–2.0 2.0–2.5

Phone type penetration per region Developing No Feature Smartphone

2017 actual 2025 estimate

1. Offline training 2. Drone deployment 3. User interface

4. Pre-processing 5. Detection 6. Output

3. Machine learning model

SOURCE: PlanetLabs; McKinsey Global Institute analysis

Critical for most domains

 Data accessibility Critical for select cases1

 Data quality  Data volume Contextual challenges

 High-level AI expertise  Data labeling  Data availability

SOURCE: McKinsey Global Institute analysis

The availability of computing infrastructure on which to train AI models is no longer as

Data accessibility remains a significant challenge. To resolve it will require a willingness

Economic empowerment Education Environment

Equality and inclusion Health and hunger Information verification and

2 McKinsey Global Institute

Several AI capabilities, primarily in the categories of computer vision and natural

Crisis response. Specific crisis-related challenges, such as responding to natural and

Economic empowerment. Opening access to economic resources and opportunities,

Environmental challenges. These include sustaining biodiversity and combating

Equality and inclusion. Addressing equality, inclusion, and self-determination

Information verification and validation. The challenge of facilitating provision,

Infrastructure management. Infrastructure challenges that could provide public