Over the past decade, machine learning-based algorithms have been deployed across a wide range of use cases and industries. From the algorithms that assess an individual's creditworthiness, to algorithms that serve up suggested movies and shows to watch on Netflix, the impact of Big Data, analytics, and automation are felt daily by nearly everyone.
One area of life where algorithms have not yet been perfected is with payments made by government or relief organizations to people in the aftermath of a crisis, emergency, or natural disaster, where getting financial relief to the people who need it most is critical. Though there have been pilot programs and limited use of artificial intelligence (AI) to provide targeted aid, the practice is far from widespread.
Key drivers behind the desire to incorporate more automation and data analysis into aid dispersion is the time-consuming nature of assessing who is eligible to receive aid, and then ensuring that aid is only delivered to those legitimate recipients. Due to the scale and compressed timeframes of delivering aid in the wake of a disaster or emergency, many organizations have struggled with this process.
For example, U.S. federal agencies made approximately $281 billion in payment errors during fiscal year 2021, up $75 billion from the previous fiscal year, and nearly double the amount reported in 2017, according to data from the U.S. Government Accountability Office. Errors in the distribution of funds under unemployment insurance and small business loan programs contributed to that total, driven by the COVID-19 pandemic response.
However, it is not only government agencies that struggle. Non-government charitable organizations, many of which are actively involved in relief efforts when unexpected natural disasters occur, also face challenges, usually related to fraud, where scams are used to access funds that are earmarked for actual victims. A 2021 report from Western Union found the cost of financial fraud impacting non-government organizations reached an estimated $5.1 trillion in 2019, with just 9% of non-governmental organizations (NGOs) indicating they have some type of fraud awareness program in place, and 54% of NGOs saying they do not report fraud to law enforcement because of the negative reputational impact it can have on future potential funding.
A pilot program of nonprofit GiveDirectly used an algorithm to direct aid to victims of Hurricane Ian in Florida.
One of the ways in which technology can be used to mitigate fraud is by correlating the level of impact of a disaster or emergency condition with the people who have been affected. In September 2022, Hurricane Ian struck the state of Florida, resulting in significant damage from high winds and flooding across several counties. Instead of asking residents to fill out lengthy forms to receive aid, which requires a significant amount of time and effort to reach out to all possible recipients to determine eligibility, a pilot program conducted by the nonprofit GiveDirectly utilized an algorithm to help direct aid to nearly 3,500 residents of Collier, Charlotte, and Lee counties.
To make the initial assessments of which areas needed assistance, a mapping tool called Delphi was developed by four Google machine learning experts who worked with GiveDirectly over a six-month period. Delphi is used to overlay live maps of storm damage onto data on poverty from sources including the U.S. Centers for Disease Control and Prevention (CDC) to pinpoint communities that are likely to be in need after disasters. The storm damage data is provided by another Google tool called Skai that uses machine learning (ML) to analyze satellite imagery from before and after a disaster, then estimates the level and scope of damage to buildings.
To train the ML algorithms powering Skai's damage assessments, satellite images of hundreds of buildings in the disaster area are manually labeled, so the algorithm can then identify similarly damaged buildings across the entire affected area using unlabeled images. According to the company, the tool was 80% more accurate than manual assessments when it was used on the Beirut, Lebanon, port explosion in 2020, and in the aftermath of Cyclone Yasa in Fiji in 2021.
To provide aid to the people who needed it after Hurricane Ian, GiveDirectly sent a push notification to users of a benefits app called Providers, which manages food stamp payments. By taking the correlated damage assessments from Skai and identifying the people who used Providers who lived in affected areas, the program was able to offer $700 in aid directly to people who needed it. As of October 2022, 900 people had accepted the offer; if all 3,500 eligible recipients were to accept the offer, $2.4 million would be paid out in direct financial aid.
One of the challenges with using algorithms and automation to deliver cash aid to victims is that, in an era where spam and phishing scams are ubiquitous, people who genuinely need assistance may think an unsolicited offer of aid is too good to be true and simply a scam. Indeed, GiveDirectly conducted a test in September 2022 after Hurricane Fiona in Puerto Rico, sending out push notifications informing recipients about the availability of an immediate cash benefit to 700 people, but less than 200 people took up the offer. Sarah Moran, GiveDirectly's U.S. director, told Wired, "That was a lower response than we would have expected," blaming the low uptake on people suspecting the messages were a phishing campaign.
People who genuinely need assistance may think an unsolicited offer of aid is too good to be true.
Still, the use of AI and machine learning to deliver aid is not commonplace; GiveDirectly's Hurricane Ian efforts were the first observed use of this type of technology in the U.S. However, other organizations are realizing the value of using technology to provide disaster or other emergency assessments and disbursing financial and other forms of relief.
Currently, the Red Cross, a large NGO that provides humanitarian relief services and aid, does not yet use AI or ML for any of its relief analyses. However, the organization says it has innovation projects under way and plans to integrate both AI and machine learning into its operational decision making in the future.
That said, the Red Cross utilizes myriad data from a variety of sources to conduct its assessments. The organization's current practice is to collect and integrate CDC Social Vulnerability Index (SVI) data, U.S. Census data, data from the organization's call center on client needs, crowdsourced damage data, Red Cross damage assessments, government-provided storm and fire-track data, and service delivery data such as shelter population trends into RC View, the Red Cross' ArcGIS disaster management system.
The use of AI and ML is not restricted to U.S. organizations. GiveDirectly and the government of Togo in West Africa teamed up with the Center for Effective Global Action (CEGA), a multidisciplinary research center at the University of California, Berkeley, addressing international development challenges, and Innovations for Poverty Action (IPA), a research and policy non-profit focused on promoting solutions for addressing global poverty, to assemble a team to use ML and mobile phone data to provide aid to citizens in poverty due to the COVID-19 lockdowns. The program used ML and mobile phone data to identify, enroll, and pay more than 138,500 Togolese citizens that needed aid. This approach was found to be more accurate in identifying eligible recipients, and reduced the number of eligible recipients who might have been excluded from receiving benefits by 8%-14% compared with alternative geographic targeting measures being considered by the government, including simply selecting residents of the poorest prefectures or cantons, or by using asset tests to determine eligibility.
The challenge with providing disaster relief payments via an ML model is the organization that deploys an algorithm must clearly define the target recipient group and tune the model appropriately. "In Togo, the government's objective was to deliver benefits to the poorest people in the country, so our efforts focused on training a machine-learning model to target the poor," according to Aikens et al. "In other settings, such as following natural disasters, the people most impacted by adverse events may not be the poorest. With high-frequency phone data available in near-real time, related techniques might be used to more dynamically prioritize the people with the greatest need. For example, it may be possible to train a machine learning algorithm to identify people whose consumption fell by the greatest amount, based on changes in patterns of phone use following a crisis."
Still, the use of AI and ML should be approached with a measure of caution, particularly when the inputs and drivers of the algorithm are not disclosed. A prime example is the use of AI by the U.K.'s Department for Work & Pensions, which has been using algorithms to track possible cases of benefits fraud. An algorithm was developed to judge the likelihood recipients' claims about their childcare and housing costs are true when they apply for benefits, using data from law enforcement, benefits offices, and personal credit data, to assess whether people are qualified to receive benefits.
However, a group in Manchester launched a legal challenge in November 2021 following testimonials from disabled people in the area claiming they were being disproportionately targeted for benefit fraud investigations, often without any explanation as to why they were being flagged. The group, the Greater Manchester Coalition of Disabled People, said that once flagged, those targeted in this way can face an invasive, humiliating, and time-consuming investigation lasting up to a year.
The organization that deploys an algorithm must clearly define the target recipient group and tune the model appropriately.
Since that lawsuit was filed, the U.K. government has begun to implement steps to improve algorithmic transparency. The U.K.'s National Data Strategy added in October 2021 a Algorithmic Transparency Data Standard, along with a transparency template and guidance for agencies that use algorithms to support decisionmaking.
In situations where aid-based decisions are being made, algorithms may be best used as a complementary tool to traditional analyses. Indeed, while the Red Cross' innovation team is working with technology partners and developing specific AI and ML projects for this year and beyond, the organization is not planning to abandon traditional analytical or hands-on approaches for identifying need and delivering aid.
"We believe that operational art and science are both necessary to determine who needs help after a disaster," says Nicole Maul, a spokesperson for the Red Cross. "We have been responding to disasters for more than 140 years and what we have found is that in order to best serve disaster survivors, we need to blend the tried-and-true, boots-on-the-ground efforts with the benefits of tools such as geospatial technology and partnerships with other organizations."
A key example of using technology—but not AI—to deliver targeted benefits is the Indian government's transfer of subsidies and benefits to the poor, which included the use of biometrics, electronic banking information, and a secure payments bridge. The initiative, called Aadhaar, was implemented starting in 2013 by requiring citizens to provide biometric data to link them with their individual bank account.
The National Payments Corporation of India (NPCI), an organization created by the Reserve Bank of India and the Indian Banks Association that operates the retail payments and settlement systems in India, created the Aadhaar Payment Bridge (APB) System, a payment system that allows direct payments to be made between the government and individuals. A unique Aadhaar number is used as a central key for electronically channeling government benefits and subsidies in the Aadhaar Enabled Bank Accounts (AEBA) of the intended individual.
By capturing and storing unique, immutable biometric information with a specific ID number, and then linking that ID with a specific bank account, more than a billion Indians have been able to receive benefits payments, subsidies, and other financial relief more efficiently than traditional paper-based benefits disbursement, while also reducing corruption or leakage due to manual processes. The Indian government's National Informatics Centre noted Direct Benefit Transfer (DBT) "played a major role in sustaining life especially of the underprivileged segments of the society impacted by the COVID-19 crisis, helping millions in providing immediate relief in tiding over the turbulent period."
Coronavirus Oversight, U.S. Government Accounting Office: https://www.gao.gov/coronavirus
What NGOs need to know to prevent fraudulent activities?, Western Union, May 21, 2021, https://bit.ly/3FTLsfv
Aiken, E., Bellue, S., Karlan, D., Udry, C., and Blumenstock, J.E.
Machine learning and phone data can improve targeting of humanitarian aid, Nature, March 31, 2022, https://doi.org/10.1038/s41586-022-04484-9
©2023 ACM 0001-0782/23/06
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and full citation on the first page. Copyright for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or fee. Request permission to publish from [email protected] or fax (212) 869-0481.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2023 ACM, Inc.