Working Paper recently convened fifteen organizations – including experts from ICPSR, the World Bank, Google Data Commons, Robert Wood Johnson Foundation, and NYU’s Center for Social Media and Politics – to discuss how the Model Context Protocol (MCP) is changing how social impact organizations access and trust their data.
The takeaway: MCPs paired with LLMs are dramatically outperforming standalone AI systems for generating reliable, source-grounded insights. Organizations like the World Bank are building cryptographic verification layers (Proof-Carrying Numbers) that reject AI outputs when they can’t be traced to authoritative sources. UChicago’s Genomic Data Commons shows that smaller models on curated data consistently beat frontier models on accuracy. But participants also raised hard questions about who gets left out when data pipelines remain underfunded, and how state propaganda is already shaping what LLMs “know” about the world. Our full summary is available here.
Full Text:
https://workingpaper.co/papers/2025.11.11_Working_Paper_MCP_Convening.pdf
Abstract:
As AI tools permeate enterprises, governments, and daily life, social impact organizations face a dual challenge: adopting new technologies while protecting under-resourced and marginalized communities from harm. Philanthropies and research organizations navigate troves of narrative reports and decades of individual expertise and shifting metrics to inform changes to complex systems from migration to vaccine hesitancy to democratic resilience. Large language models provide support, but the emergence of the Model Context Protocol (MCP) significantly expands the potential for reliable, source-grounded analysis that supports expert decision-making in the social sector.
Working Paper recently convened a Chatham House–style gathering of individuals from fifteen organizations to share early experiences building and applying MCPs for social impact organizations. This document summarizes key insights, participant discussions, and major themes from the presentations, and highlights implications for the field.