Building Capacity for Responsible AI in Social Impact: Reflections from Tech4Dev's AI Cohort

Artificial intelligence has long been hailed for its potential to amplify impact across sectors.There has been growing recognition that when deployed responsibly and ethically, AI can be a force of social innovation. Drawing on this recognition, Tech4Dev launched its inaugural AI Cohort, supporting 7 NGOs building AI interventions at different stages of development, spanning use cases in healthcare and education.

DFL’s role in this AI Cohort involved working closely with the participating NGOs over 4 months, supporting teams to design, build, and scale their AI solutions using safe, responsible, and ethical approaches.

Our broad approach included analysing the AI tools being designed against key RAI principles, interrogating their impact thesis or theory of change (ToC), and supporting the development of Monitoring, Learning and Evaluation (MLE) frameworks. We provided this support through both 1:1 sessions as well as cohort-wide workshops, similar to workshops under DFL’s RAIL Fellowship.

With the cohort having culminated, we wanted to share our key learnings from this engagement, especially as this space of experimenting with AI models for social impact continues to expand within India and the wider Global South.

Trial-and-error as an Integral Feature of Responsible AI

Across the different journeys undertaken by the NGOs, the iterative nature of AI development stood out clearly. Several organisations that had already piloted their tools realised that they had taken certain assumptions about their intended users for granted, particularly around digital literacy, ease of access, and comfort with technology, and by extension, with AI systems. These insights prompted teams to revisit and redesign interfaces and interaction flows, ensuring that their tools could be adopted more safely and effectively in real-world contexts.

In other cases, testing surfaced a different set of risks altogether. When teams began stress-testing chatbots with out-of-scope or unsafe queries, new weak points were discovered, especially with respect to “harmful” outputs produced by the model. This prompted several organisations to introduce prompt-level guardrails to more clearly define conversational limits and reduce the risk of unsafe responses.

Across our work with the NGOs, each round of testing raised fresh questions: about data quality, edge cases, user behaviour, and safety. Crucially, every iteration also strengthened the interventions. Embracing trial-and-error helped create effective feedback loops, highlighting why practitioners should plan for multiple cycles of user testing and treat each cycle as a core part of their safety and evaluation strategy.

Integrating Responsible AI Early in the Development Lifecycle

For user-facing AI tools, integrating Responsible AI considerations from the outset is critical. In many contexts today, questions of responsibility tend to arise only once tools are approaching deployment or formal compliance requirements, rather than during earlier stages of design and development. By that point, certain user-facing risks or design choices may already be embedded in the system.

This challenge is particularly acute in underserved contexts, where alternative support systems and grievance redress mechanisms are limited. For instance, AI-powered chatbots are being deployed to answer women’s sexual and reproductive health queries, especially among low-income and low-literacy populations where medical access is limited and stigma around such themes remains high. However, because these systems are probabilistic and non-deterministic, they may hallucinate or provide inaccurate medical advice—such as recommending unproven treatments or unsafe protocols, exposing the user to the risk of bodily harm. Responsibility, therefore, cannot be treated as reactive. User-centred design, interdisciplinary collaboration, and early planning for monitoring and evaluation need to be embedded well before deployment.

Recentering the Social Impact of AI Interventions

How organisations define success for their interventions plays a central role in how they approach Responsible AI in their deployment and scaling strategies. In practice, success is often measured through short-term, technical indicators such as model accuracy or alignment with human judgement.

However, responsible deployment also depends on articulating broader social outcomes. This includes asking whether AI outputs are meaningfully usable by end users, whether they translate into improved services or decision-making, and whether benefits are equitably distributed across diverse users.

Longitudinal user research and more iterative pilots make it possible to examine these questions more closely, enabling closer alignment between evaluation frameworks and an intervention’s underlying theory of change, as well as greater visibility into unintended consequences and real-world impact over time.

All in all, the AI Cohort underscored that responsible, impact-oriented AI is less about one-off technical solutions and more about cross-disciplinary collaboration and sustained learning cycles.

The four-month engagement offered a valuable moment to surface questions, practices, and approaches that are central to this kind of work, even as learning and evidence-building around AI for social impact continues over time.

We are excited to see how future cohorts build on these conversations, and how the sector continues to learn through iterative, evidence-driven practice on responsible AI.

Upcoming Project Outputs

DFL is developing 2 knowledge products as a result of our engagement with the AI Cohort, due to be published in mid-2026.