Empowering change: GEM Incubation Fund awardees catalyze AI for inclusive development
The Harvard Center for International Development's 2025 Global Empowerment Meeting put AI on the table for one clear purpose: more inclusive global development. Alongside the convening, CID opened the annual application cycle for the GEM Incubation Fund and introduced the 2025 awardees - a cohort translating conference energy into early-stage projects with policy reach and real-world impact.
Launched each year with GEM, the fund backs emerging research that values creativity, experimentation, and collaboration. This year's focus zeroes in on equity and inclusion. The selected projects apply AI and lessons from its fast evolution to priority areas: health, education, agriculture, climate adaptation, and crime - with a clear intent to expand opportunity in underserved communities.
What the 2025 cohort is building
- Health: decision support, earlier detection, and resource allocation that fits low-resource settings.
- Education: data-driven instruction, tutoring at scale, and teacher enablement over replacement.
- Agriculture: yield insights, pest and weather risk alerts, and market access for smallholders.
- Climate adaptation: local forecasts, resilient infrastructure planning, and community risk tools.
- Crime and safety: pattern analysis that improves prevention while respecting rights and privacy.
Methodologically, the work ranges from qualitative fieldwork and small pilots to AI simulations and system-level analysis. Different tools, same principle: build what's useful, ethical, and testable.
"This year's Incubation Fund cohort captures the ambition and urgency of GEM's theme," said CID Faculty Director Asim I. Khwaja. "Their work pushes us to think more expansively about what AI can enable, and who it can empower."
Why this matters for IT and development teams
- Start where the friction is: poor data quality, manual triage, and offline constraints are common; solve those first.
- Ship small, learn fast: pilots with clear metrics beat grand designs. Prioritize quick feedback loops.
- Build for low-connectivity: offline-first apps, lean models, and caching aren't optional in many regions.
- Privacy and consent by default: protect vulnerable groups with minimization, local processing, and clear consent flows.
- Human-in-the-loop: keep domain experts in review paths to prevent silent failure and drift.
- Measure fairness: track error rates by subgroup; publish evaluation protocols and mitigation steps.
- Interoperability: use open standards and APIs so governments and NGOs can adopt without vendor lock-in.
From insight to implementation
- Define the unit of change: is it a decision, a workflow, or a budget line? Make one metric the north star.
- Co-design with users: field interviews → paper prototypes → limited pilot → phased rollout.
- Right-size the stack: start with lightweight models; only escalate to bigger models if the use case proves it.
- Data pipeline first: schema design, lineage, and quality checks before model training.
- Guardrails: logging, policy checks, content filters, and red-teaming for safety-critical tasks.
- Governance: document model purpose, training data, limits, and update cadence; assign owners.
What to watch in 2025
- Hybrid systems that combine rules, retrieval, and smaller models to cut cost and improve reliability.
- Localization: models tuned to local languages, dialects, and norms with community input.
- Policy interface: clearer playbooks for procurement, evaluation, and responsible deployment in the public sector.
If you want more background on the convening and the fund, see the Center for International Development's page at Harvard Kennedy School: CID at HKS.
For teams building skills to support pilots like these, explore role-based AI training paths: Complete AI Training - courses by job.
Your membership also unlocks:
