Gender fundamentally alters who benefits from social protection in low- and middle-income countries, so policy tools — including AI systems used for targeting, monitoring and evaluation — must be designed with gender-disaggregated data, causal methods for subgroup effects, and governance safeguards to prevent reinforcing existing inequalities.

Main Finding

Gender fundamentally shapes both the need for and the outcomes of social protection in low- and middle-income countries (LMICs). Effective poverty- and vulnerability-reduction requires gender-sensitive program design, delivery, monitoring, and research because (a) poverty risks and vulnerability differ by gender; (b) gender norms and intra-household dynamics mediate program impacts; and (c) women and girls are often systematically disadvantaged in opportunities and access. Recent policy and research trends increasingly recognize and attempt to address these points.

Key Points

• Differential prevalence and risk: Poverty drivers and vulnerability profiles vary by gender, so one-size-fits-all programs can miss or underserve important needs.
• Mediating mechanisms: Gender norms, intra-household bargaining, time use, mobility, and access to markets/services shape how individuals participate in and benefit from programs.
• Equity and efficiency case: Targeting gender inequalities via social protection can promote longer-term poverty reduction, productivity gains, and broader social benefits.
• Implicit prioritization: Because social protection aims at equity, there is an argument (and growing practice) to prioritize women and girls in program design and targeting.
• Evolving practice and evidence: Program design and research questions have become explicitly gender-aware over recent decades (e.g., gendered targeting, conditionalities, complementary services).
• Scope of review: The chapter focuses on adulthood in LMICs and synthesizes evidence across three categories: social assistance (noncontributory transfers, public works, fee waivers, subsidies), social care, and social insurance.
• Evidence gaps & divergences: While consensus exists on the importance of gender, findings diverge on mechanisms, heterogeneous effects, and best practices for implementation across contexts.

Data & Methods

• Approach: The chapter uses a "review of reviews" — synthesizing existing reviews and meta-analyses to summarize the consensus and contested findings on social protection and gender in LMICs.
• Scope and focus: Emphasis on adult gender dimensions (complements separate chapters on early childhood and youth). Examines design, delivery, and measured impacts of social assistance, social care, and social insurance.
• Types of evidence considered: Randomized controlled trials, quasi-experimental evaluations, observational studies, qualitative research, and program implementation reports aggregated by prior reviews.
• Common methodological strengths and limitations in the literature:
  • Strengths: growing use of causal evaluation methods; richer program variation to study design features.
  • Limitations: underreporting of gender-disaggregated and intra-household outcomes, limited long-term follow-up, heterogeneous measures of gendered outcomes, and few studies integrating qualitative insights on norms with quantitative causal estimates.
• Identified research priorities: capture intra-household allocation, longer-term and intergenerational effects, mechanisms linking norms to outcomes, and context-specific heterogeneity.

Implications for AI Economics

• Data design and collection
  • Prioritize gender-disaggregated, individual-level, and intra-household datasets (consumption, labor, time-use, mobility, agency, access to services).
  • Collect longitudinal data and qualitative/contextual indicators (norms, decision-making, conditionalities) to support causal and dynamic modeling.
  • Ensure metadata documents measurement choices that may bias gendered inference (who responds, who is observed).
• Modeling and causal inference
  • Use causal ML methods to estimate heterogeneous treatment effects by gender and household role while guarding against overfitting and spurious subgroup claims.
  • Combine structural models and agent-based simulations to represent intra-household bargaining, norm dynamics, and long-run accumulation that ML alone may miss.
  • Develop methods for robust inference in small subgroups and for transportability across contexts with different gender norms.
• Algorithmic targeting and fairness
  • When using predictive models for targeting or enrollment, incorporate fairness constraints and distributional objectives (not just efficiency) to avoid systematically excluding women or reinforcing existing inequities.
  • Beware proxy variables that encode gendered access (e.g., mobile-phone ownership) and can introduce bias into automated targeting.
  • Prioritize transparency, interpretability, and participatory validation with affected communities to detect unintended harms shaped by norms.
• Program design optimization and delivery
  • Use reinforcement learning / policy search methods constrained by normative and ethical considerations to explore trade-offs (e.g., cost vs. gender-equitable outcomes).
  • Apply ML for operational tasks (fraud detection, leakage identification, payment routing) while ensuring these systems do not differentially disadvantage women (e.g., due to lower formal ID or mobile access).
  • Leverage natural-language processing and low-cost sensors for monitoring and beneficiary feedback, but validate across gender and literacy differences.
• Evaluation metrics and objectives
  • Move beyond income-only metrics: include measures of agency, time use, health, care burdens, safety, and intra-household resource allocation to evaluate gendered impacts.
  • Use mixed-methods evaluation (quantitative + qualitative) to uncover norm-mediated channels that pure predictive models miss.
• Ethics, privacy, and governance
  • Protect sensitive gender-related data; ensure informed consent and consider risks from digitalization (e.g., exposing beneficiaries to domestic conflict).
  • Design governance frameworks that include gender-aware stakeholders in model development, deployment, and monitoring.
• Research agenda for AI economists
  • Build and share gender-disaggregated benchmark datasets that respect privacy.
  • Develop causal ML tools tailored to policy evaluation with subgroup inference and transportability.
  • Study how algorithmic policies interact with social norms and intra-household dynamics (theory + empirical).
  • Test interventions where AI-assisted targeting or monitoring explicitly aims to improve gender equity, documenting mechanisms and unintended effects.

Concise recommendation: AI economists working on social protection should treat gender as a core modeling and evaluation dimension — from data collection through algorithm design and policy optimization — using causal, interpretable, and ethically governed methods that explicitly target equitable outcomes rather than only aggregate efficiency.