AI will not save evidence ecosystems in the Global Majority, but relationships might

The recent launch of the working paper Responsible AI in Education: Evidence Synthesis and Use in Low- and Middle-Income Countries calls for a timely recalibration in the global conversation on artificial intelligence in education. At a moment when much of the discourse around generative AI centres on speed, automation, and model sophistication, the paper redirects attention to a more consequential question: how AI reshapes evidence ecosystems, and how evidence is interpreted, trusted, and used in high-stakes policy contexts.

From the vantage point of practice in the Global Majority, this reframing reveals a deeper structural challenge. In many education systems, the core constraint is not the absence of data or analytical tools. Ministries commission studies, donors fund evaluations, and dashboards proliferate. Yet despite this density of production, evidence rarely becomes a shared object of sustained deliberation. Particularly in complex domains such as student well-being, classroom practice, and school culture, evidence circulates but does not consistently shape collective reasoning.

The challenge lies less in technology and more in what might be called the relational infrastructure of evidence – the institutional spaces, norms, and professional capacities that enable researchers, policymakers, and practitioners to interpret evidence together.

In India, we have been piloting a GenAI-supported evidence platform as part of ongoing research–policy collaborations with state education departments across several states. The initiative integrates large-scale social and emotional well-being data from public schools, state administrative data, and insights from global studies. An online dashboard with an AI conversational interface allows policymakers and educators to pose natural-language questions across these datasets.

The technical objective was straightforward: to lower barriers to accessing complex evidence. However, easier access did not translate into greater influence. Evidence became more retrievable, but not more meaningful.

Uptake improved only when the AI-enabled dashboard was embedded within facilitated research–practice–policy sessions involving state officials, district administrators, researchers, and school leaders. These structured working sessions were integrated into ongoing state review and planning processes. Participants posed live questions to the dashboard, examined emerging patterns collectively, and paused to assess whether those patterns aligned with classroom realities, implementation constraints, and policy priorities.

Policymakers reflected on feasibility and alignment with state goals. Researchers interrogated methodological assumptions and data limitations. Practitioners grounded the discussion in everyday school experience. Through this process, contextual knowledge and professional judgment became integral to interpretation.

In these settings, AI did not function as an answer engine. It acted as a catalyst for inquiry. Meaning emerged through dialogue and shared reasoning rather than from automated outputs.

This experience underscores a central insight: evidence uptake is a relational achievement. AI can widen access to information, but only locally grounded collaboration converts access into influence.

The dashboard served as a shared reference point for diverse actors to reason together. Its value lay not in resolving disagreement but in sustaining structured conversation across professional boundaries. In contexts shaped by unequal histories of knowledge production, this distinction is significant. When AI is framed as neutral or definitive, it risks reinforcing epistemic dependency. When positioned as a support for collective interpretation, it can broaden participation in the use of evidence.

Our experimentation also revealed the limits of automation. Unreviewed AI-generated summaries sometimes omitted contextual qualifiers and presented tentative findings with greater certainty than the evidence warranted. We therefore embedded human review before any AI-generated synthesis entered policy discussions. This was not merely a procedural safeguard; it reflected the recognition that educational meaning is interpretive work requiring contextual knowledge and responsibility.

Human judgment is not noise in an evidence system; it is the medium through which evidence becomes meaningful.

Perhaps most revealing was how the platform exposed gaps in locally grounded education data. In several instances, participants discovered that questions central to their decision-making could not be answered because relevant longitudinal data were unavailable. The AI did not compensate for these gaps; it illuminated them. In doing so, it redirected attention toward strengthening data systems rather than relying on technological inference.

The working paper cautions against premature automation in high-stakes contexts. Practice adds a complementary proposition: the promise of AI lies less in replacing interpretation than in strengthening the relational infrastructure through which interpretation occurs. Investments that prioritise standalone tools without cultivating collaborative spaces risk deepening dependency rather than enhancing agency.

If AI is to contribute meaningfully to education systems in the Global Majority, success must be defined differently. It should not be measured by model sophistication or output volume, but by whether cross-stakeholder conversations become more rigorous, whether uncertainty is handled more honestly, and whether professional judgment is exercised more deliberately.

Artificial intelligence can expand access to information. It cannot, by itself, build trust, negotiate meaning, or sustain shared responsibility. Those capacities reside in relationships.

Humanising AI, in this context, does not mean making machines more empathetic. It means ensuring that human judgment, accountability, and collective reasoning remain central to decision-making.

Note: This piece draws on insights from the working paper Responsible AI in Education: Evidence Synthesis and Use in Low- and Middle-Income Countries, produced by the Building Evidence in Education (BE²) Secretariat. The report examines how AI is being introduced into education evidence workflows and outlines the conditions under which it can responsibly support interpretation while preserving human judgment and locally grounded evidence ecosystems.

Author

Sreehari Ravindranath is a psychologist and education researcher working at the intersection of social and emotional learning, well-being, evidence use, and education policy in the Global Majority. He currently serves as the Director of Research and Impact at Dream a Dream in India, where he leads large-scale research–policy collaborations to strengthen locally grounded evidence ecosystems. His work engages with questions of relational pedagogy, data governance, and the interpretation and use of evidence within complex education systems.