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. ƒ 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. ƒ 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. ƒ 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. ƒ 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. ƒ 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. ƒ 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 ▪ Labor supply and demand matching Public and social sector 16 15 ▪ Effective management Education of public sector ▪ Effective management ▪ Access and completion of education of social sector ▪ Fundraising 13 ▪ Maximizing student ▪ Public finance 16 achievement ▪ Teacher and administration management ▪ 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 and sustainability 11 ▪ Land, air, and Information 28 water conservation verification 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 AI societal good Discussion paper mc 1129 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 This domain has high potential for AI use, although problems can be 7 10 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 21 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 3 1 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. SOURCE: McKinsey Global Institute analysis McKinsey Global Institute Notes from the AI frontier: Applying AI for social good 3 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 4 McKinsey Global Institute Notes from the AI frontier: Applying AI for social good 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 SOURCE: McKinsey Global Institute analysis 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. ƒ 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. ƒ 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. McKinsey Global Institute Notes from the AI frontier: Applying AI for social good 5 ƒ 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. ƒ 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 ƒ 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. ƒ 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. ƒ 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. 6 McKinsey Global Institute Notes from the AI frontier: Applying AI for social good