AI Digest for IDinsight — Feb 13

🌍 Field Applications 4 items

Featured

India's SAHI: A Radical AI-for-Health Framework

ICTworks · Feb 24, 2026

India's Strategy for AI in Healthcare is the first LMIC health-AI strategy to launch with an operational platform (BODH for benchmarking health AI models) and explicitly prioritize "innovation over cautionary restraint." It introduces a tripartite data-gap taxonomy that distinguishes AI-critical, AI-limiting, and AI-enhancing datasets — shifting investment conversations from "we need more data" to "which data deficits block which use cases." The framework also dedicates a full pillar to outcome-oriented procurement reform, directly addressing vendor lock-in and fragmented purchasing.

💡 Key Insight

SAHI's data-gap taxonomy is directly applicable to IDinsight's health and analytics work. Instead of the usual "collect more data" recommendation, it provides a framework for advising LMIC governments on where data investment yields the highest AI return — a concrete tool for our advisory engagements.

Anthropic–Rwanda 3-Year MOU for AI in Health & Education

Anthropic · Feb 17, 2026

First formal multi-sector government-AI partnership on the African continent. The 3-year MOU covers Claude integration for cervical cancer elimination, malaria and maternal mortality reduction, 2,000 Claude Pro licenses for educators, AI literacy training for public servants, and Chidi (a Claude-powered learning companion) deployed across 8 African countries.

India AI Impact Summit: A "Third Way" for the Global South

Rest of World · Feb 11, 2026

India's AI Impact Summit — the first hosted by a developing country — positions India as an alternative AI governance model between the US and China, framing AI as a public good. Multiple countries (Rwanda, Nigeria, UAE) are competing for AI leadership while development-sector AI narratives risk obscuring Big Tech dependency.

AI4D's Geographic Blind Spot: Ignoring Billions Beyond Africa

ICTworks · Feb 26, 2026

Comprehensive analysis finds AI-for-development research overwhelmingly focuses on Africa while ignoring South and Southeast Asia, Latin America, and MENA — where AI deployment rates are often higher. India leads globally at 59% company-level AI deployment; over 40% of ChatGPT traffic originates in middle-income countries led by Brazil, India, Indonesia, and Vietnam.

💬 Debates 4 items

Featured

Gender Bias in LLMs: Data Labelling and Qualitative Analysis

MERL Tech · Feb 9, 2026

Systematic tests of ChatGPT, Claude, and Gemini on simulated qualitative MERL data from a women's economic empowerment project. Claude performed strongest on intersectional analysis — identifying more dimensions of identity and flagging missing data. ChatGPT showed contextual distortion, misinterpreting legitimate concerns about donor KPIs as "bias against donors." Separately, Reliabl's research found 90% of image-generation outputs for "beautiful woman" depicted light-skinned women, even when ethnicity was implied.

💡 Key Insight

For IDinsight teams using LLMs for qualitative coding or analysis: bias is subtle rather than overt, requiring extensive validation. The study confirms that prompting helps mitigate some issues but cannot fix structural biases in training data. Human oversight and model comparison (e.g., via OpenRouter) remain essential.

Made in Africa AI Competencies for Evaluators

MERL Tech · Feb 10, 2026

Research digest identifies 7 core competencies for African MERL practitioners using AI, including AI literacy, decolonial awareness, participatory community approaches, and data sovereignty. Most practitioners use AI informally for literature reviews and drafting, but face Africa-specific gaps: poor representation of African knowledge sources in training data and limited African language support.

AI Is Shifting Power from Governments to Tech Companies

Rest of World · Feb 18, 2026

Essay by NUS Dean Simon Chesterman argues today's tech giants wield power comparable to the East India Company and that the public-private divide matters more than North-South in AI governance. Governments have been reluctant to regulate, and AI is increasingly enlisted as an instrument of state policy.

IDinsight Alum Critiques AI Funding Dynamics

LinkedIn · Feb 14, 2026

Shared by @Becca Gong Sharp

Tony Senanyake (CEO of Fortify Health, IDinsight alum) published a critique of development funding dynamics that focus too much on the "shininess" of AI tools and too little on ethical obligations to create systems change and real-world results. Becca's commentary underscores the need to be intellectually honest about when not to use AI tools.

🔧 Tooling 5 items

Featured

GABRIEL: GPT as a Measurement Tool

NBER Working Paper · Feb 2026

Shared by @Fahad Hasin

New NBER paper (Asirvatham, Mokski & Shleifer) presents GABRIEL, a software package that uses GPT to quantify attributes in qualitative data — e.g. how "pro-innovation" a speech is. Tested against 1,000+ human-annotated tasks across multiple domains, GPT was found to be generally indistinguishable from human evaluators. Results did not depend on the exact prompting strategy used, and the authors found no evidence of training data contamination.

🧪 Try it

Use GABRIEL to code qualitative survey responses or policy documents at scale for IDinsight evaluations — potentially reducing manual coding time while maintaining inter-rater reliability.

💡 Key Insight

This is the strongest evidence yet that LLMs can serve as reliable measurement instruments for qualitative data. For IDinsight's evaluation work, GABRIEL offers a concrete, tested tool for scaling qualitative coding — with the important caveat that the MERL Tech gender bias findings (see Debates) remind us that LLM-based coding still requires human validation, especially for culturally sensitive topics.

A Guide to Which AI to Use in the Agentic Era

One Useful Thing · Feb 18, 2026

Shared by @Tom Wein

Ethan Mollick's updated guide marks a "very large break with the past" — distinguishing Models, Apps, and Harnesses as three layers of AI tools. Highlights Claude Cowork for delegating multi-step office work, NotebookLM for research synthesis, and the shift from chatbot prompting to agent-style managing.

🧪 Try it

Upload project documents to NotebookLM to create a searchable knowledge base for your evaluation — great for synthesizing across concept notes, workplans, and meeting notes.

Google WaxalNLP: Open Dataset for African Speech Technology

Google · Feb 2026

Shared by @Tom Wein

Open speech dataset covering 22+ African languages including Swahili, Yoruba, Hausa, Igbo, Kikuyu, Lingala, Luganda, and Shona. Data ownership is retained by African partner institutions — a model for responsible dataset development.

🧪 Try it

Evaluate whether WaxalNLP languages overlap with IDinsight survey contexts for potential voice-based data collection or transcription improvements.

Claude Sonnet 4.6: Opus-Level Performance at Sonnet Pricing

Anthropic · Feb 17, 2026

New default model on Claude.ai and Cowork brings significant improvements in coding, data analysis, computer use, and long-context tasks. Matches Opus 4.6 on most real-world analytical work while being faster and consuming fewer usage credits.

🧪 Try it

Switch to Sonnet 4.6 for everyday tasks — it matches Opus on most analytical work at lower usage cost, preserving your limits for heavy-lift projects.

Four Modes of Working with AI

LinkedIn · Feb 26, 2026

Shared by @Tom Wein

Practical framework for structuring AI use: Compression (distill information and find the signal), Expansion (generate options beyond your defaults), Reflection (stress-test thinking with a no-ego devil's advocate), and Execution (produce drafts, summaries, formats).

🧪 Try it

Use Reflection mode as a "devil's advocate" for analysis plans and Terms of Reference — ask AI to challenge your assumptions before they're locked in.