Transcript: Episode 15

Welcome to Impact Signals, social impact at the scale of AI. I'm Charlie.

It's Friday, February 27th, 2026 — Episode 15 — and today we're tracking how the global humanitarian system is quietly being rebuilt around AI: new country-level partnerships, national disaster model repositories, and a United Nations scientific body that will shape policy for a generation.

We start in Rwanda. On February 26th, the Rwandan government signed a three-year memorandum of understanding with AI lab Anthropic — the company's first formal multi-sector government partnership on the African continent.

The focus is health outcomes: cervical cancer elimination, malaria reduction, and maternal mortality. Rwanda will deploy Claude AI models inside its public health and education systems. Critically, Anthropic is providing not just API access — but Claude Code, which is essentially development tooling, plus capacity-building training so Rwandan government developers can build and maintain their own applications.

Right. Rwanda's health ministry has been deliberate about this. They called it "sovereign AI" in the MoU framing — Rwanda retains data control and operational authority. Anthropic provides tools and training, then steps back.

So practitioners watching this should see it as a replicable template. Sovereign AI structure, specific health outcome targets, capability transfer built in. If this MoU produces measurable results on maternal mortality, expect other African health ministries to request similar arrangements. The three-year timeline gives them time to demonstrate.

Story two: Google.org launched a thirty-million dollar Impact Challenge on February 26th targeting nonprofits, social enterprises, and academic institutions deploying generative and agentic AI for public services.

Those partnering with governments on AI-enabled public services — health tracking, disaster resilience, economic infrastructure. Individual grant awards range from one to three million dollars. They're also offering an accelerator: pro bono access to Google AI experts and cloud computing credits. That accelerator component is as valuable as the cash.

Exactly. Google cited internal research showing eighty percent of public servants feel empowered by AI, but only eighteen percent believe their government is actually using it effectively. The gap between pilot and scale — what they called "pilot purgatory" — is what this fund is designed to close. And there's a parallel thirty million dollar science fund for climate resilience and health applications.

As of yesterday. This is a call to action for any organization with a government AI partnership that's ready to scale. We'll link the applications in the show notes.

Story three: The Philippines government launched something genuinely interesting on February 26th — the National Artificial Intelligence Center for Research and Innovation, or NAICRI.

The core product is something called DIMER — Democratized Intelligent Model Exchange Repository. Think of it as an app store for disaster response AI models. Local government units and NGOs can pull pre-built, field-tested models for typhoon tracking, flood detection, and damage assessment — rather than building them from scratch.

Right. NAICRI also includes a model called AI4RP — a self-correcting weather forecast system calibrated specifically for Philippine weather patterns. The Philippines has seven thousand one hundred islands — localized response data matters enormously.

And they've positioned it as exportable. ASEAN neighbors facing similar typhoon exposure will be watching this. For disaster management agencies globally, this is a template for national AI infrastructure investment.

Story four: This week, the United Nations formally named the forty members of its Independent International Scientific Panel on Artificial Intelligence — establishing what Secretary-General Guterres called the first IPCC-equivalent body for AI.

The International Panel on Climate Change — the IPCC — produced the scientific assessments that shaped the Paris Agreement. This new panel has an analogous mandate: produce independent, cross-national scientific assessments of how AI is transforming societies. Forty experts from thirty-seven nations, serving in a personal capacity, not as government representatives. Three-year mandate starting now.

The same way the IPCC did — authoritative reports that policymakers cannot easily dismiss. When the IPCC released its assessment that 1.5 degrees of warming was a critical threshold, that became a negotiating baseline for every climate conversation globally. This AI panel will likely produce similar threshold documents on bias, risk, and deployment standards.

Your deployments will eventually be assessed against frameworks this panel helps define. What counts as responsible AI in crisis response? What constitutes adequate safeguards when processing displacement data? These aren't settled questions — and this panel is where they get answered. Organizations that engage in the consultation process now can shape those definitions.

The panel has three years. Which by AI development timelines is both very long and very short.

Story five: The International Committee of the Red Cross — the ICRC — announced a pilot project this week that could fundamentally change how humanitarian organizations use AI.

The ICRC, working through something called the International Computation and AI Network — ICAIN — is partnering with ETH Zurich and EPFL, two of Switzerland's leading technical universities, to build Large Language Models specifically designed for humanitarian work.

Two core problems. First, commercial AI systems simply weren't trained on data from conflict zones and underrepresented regions — the Global South is systematically absent from their training data, which means outputs are biased against populations in the highest-need situations. Second, commercial models handle sensitive data insecurely — conflict-related protection data, displacement records, civilian harm documentation cannot be processed through general-purpose commercial systems.

That's the goal. And the ICRC's track record gives this credibility — ETH Zurich already built an AI supply chain planning tool that's deployed in twelve ICRC locations, including Africa and Ukraine, which saved the organization three-point-six million Swiss francs in medical supply costs in 2023. This is the same research partnership going deeper.

Watch the ICAIN network. If this pilot succeeds, the resulting models will likely be shared across the humanitarian ecosystem under open or collaborative licensing. Organizations that are plugged into ICAIN now will have early access. For anyone processing conflict-zone data or displacement information today, this represents a more secure and contextually appropriate alternative to commercial AI.

Story six requires some care. This is a policy story with real practitioner implications.

On February 27th, reporting from the Washington Post and Defense One confirmed that Anthropic has refused Pentagon demands to remove safety guardrails from its Claude models for military use. Specifically, guardrails preventing the AI from assisting with nuclear weapon scenarios and lethal autonomous operations.

This is the right question. Setting aside the political dimension — this conflict establishes a precedent about whether AI safety constraints are negotiable under government pressure. Defense officials reportedly threatened to invoke the Defense Production Act to compel compliance. If that kind of pressure succeeds in modifying an AI company's safety policies, it has downstream effects for every organization relying on those same models.

Exactly. Humanitarian organizations using commercial AI for protection work, for refugee status documentation, for civilian harm monitoring — they need those safety layers intact. What this standoff makes clear is that AI safety policy isn't just an abstract tech ethics debate. It's a governance question that will determine what tools practitioners can trust in the field.

One to track carefully. We're reporting on the policy precedent, not picking a side in the debate.

Story seven: A research platform called DIRE — Disease Informed Response Engine — deployed by UC San Diego, UNICEF, and the European Space Agency, is now generating dengue fever and malaria outbreak predictions for Brazil and Peru four to eight weeks in advance.

Satellite imagery combined with climate data inputs — vegetation indices, temperature trends, rainfall patterns — processed through machine learning models trained on historical outbreak data. Brazil recorded one-point-six million dengue cases in January 2026 alone. The model predicts where the next cluster is likely to form before cases spike.

Yes. Personnel, supplies, and testing capacity deployed before the outbreak peak rather than after. This is anticipatory action — not reactive response. ESA's satellite coverage makes the geographic input data available globally.

Any region with endemic vector-borne disease and ESA satellite coverage — which is most of sub-Saharan Africa, Southeast Asia, South Asia — could have an equivalent system built with similar methodology. The DIRE framework could be a template.

Before we wrap, the events radar — five deadlines that matter in the next thirty days.

Yes — today, February 27th, is the early bird registration deadline for the AAAI AI plus HADR Symposium, April 7th through 9th in Burlingame, California. That's the AAAI spring symposium specifically for AI and humanitarian assistance and disaster relief researchers. If you work in this space, this is the room you want to be in. hadr dot ai has the registration link.

ITU AI for Good is hosting a free virtual webinar on AI and the world of work — coordinated with the ILO. That same day, the IndiaAI Innovation Challenge closes applications — up to one crore rupees in funding plus a two-year government implementation contract for AI solutions addressing public health or MSME governance.

And March 16th through 17th in Glasgow — the AI Standards Hub Global Summit, co-organized with the UN Office of the High Commissioner for Human Rights. This is where AI technical standards and human rights frameworks meet. Hybrid, available online.

That's Impact Signals for Friday, February 27th, 2026. If this briefing is useful, subscribe wherever you get your podcasts. Check us out at impactsignals.ai and share it with someone working in the field.