2024-05-26-世界银行-在数据匮乏的地方扩大社会援助_新数据和人工智能的机遇和局限（英）_66页_6mb

Summary of "Scaling up social assistance where data is scarce: Opportunities and limits of novel data and AI"

This paper examines novel approaches to social protection using geospatial and mobile phone data, artificial intelligence (AI), particularly machine learning, as a complement or substitute to traditional systems in low-income countries during the COVID-19 crisis. It reviews three case studies—DRC's STEP-KIN, Togo's Novissi Model 2, and Nigeria's NASSP SU—to assess the applicability of technology-based cash transfers in both emergency and regular contexts.

Key Findings:

• Coverage and Data Use: DRC used mobile data (SIM count) and satellite imagery. Togo relied on mobile data (CDR) and geospatial data. Nigeria utilized satellite maps for geographic targeting combined with household surveys. Novel data sources facilitated targeting efficiency in data-poor environments but faced challenges from digital exclusion. Many excluded non-phone-users, limiting full inclusivity.
• Implementation Speed and Efficiency: The programs deployed significantly faster (8–12 months) than traditional PMT-based systems, reducing initial registration delays and administrative costs, though long-term MNO access agreements added friction.
• Targeting Accuracy: Results from case studies remain experimental due to a lack of rigorous ex post evaluations and ground-truthing. Novissi's CDR-based targeting performed roughly as well as PMT and asset indexes, with low resolution for household-level targeting. Geographic targeting improved speed but not precision for dynamic needs assessments.

Limitations and Challenges:

• Digital Disparity: Low mobile phone penetration and varying SIM usage hindered full program reach. Digital exclusion created a regress effect, harming the poorest populations.
• Technical and Institutional Bottlenecks: Data quality, computational overhead, and institutional coordination hinder widespread use. Legal and privacy safeguards are underdeveloped, with data-protection frameworks lagging behind technological adoption.
• Geographic and Contextual Constraints: Urban focus in most programs limited rural benefits. Administrative data gaps further restricted targeting refinement, blurring the line between operational gains and accuracy.

Policy Recommendations:

• Governments must balance digital-first approaches with traditional safety nets to address inclusion gaps. Investment in foundational infrastructure, such as public ID systems and interoperability platforms, enhances target accuracy and delivery efficiency.
• Mitigate digital exclusion by blending technology solutions with manual alternatives, such as community-based outreach, along with foundational ID systems. Dedicate funds to ongoing household surveys for ground truth validation.

In conclusion, while novel data and AI help address data scarcity and improve scaling, the marginal gains in targeting are inconsistent. Technological leapfrogging is promising, but comprehensive safeguards and context-driven application are necessary to ensure equitable and effective social assistance.