#119: Impact Signals #119 — GiveDirectly's AI Flood Cash Missed the Farmers Who Could Not Charge a Phone, Microsoft Puts 5 Million Dollars Behind California's Wildfire Cameras, Sikkim Wires Its 16 Most Dangerous Glacial Lakes
AI for Impact Daily Briefing, September 02, 2026
Top Stories
GiveDirectly's AI Flood Cash Reached 4,600 People in Ibaji; The Farmers It Missed Could Not Keep a Phone Charged
In 2024 GiveDirectly and the International Rescue Committee received a 4.6 million dollar Google.org grant to deliver anticipatory cash ahead of floods in Nigeria, and GiveDirectly ran its pilot in Ibaji, a farming district in Kogi State that floods every year between July and October. Using Google's historical flood data, wealth indicators and input from the Nigeria Hydrological Services Agency, it narrowed Ibaji's 10 wards to the six most at risk; nearly 39,000 people pre-enrolled by phone short code, almost half were disqualified for self-reporting an address outside those wards, 939 failed Smile ID biometric checks, and more than 13,000 of the remaining 18,000 were dropped because they could not be confirmed as living in an eligible area. When Google Flood Hub and JBA Global Resilience hydrology forecast 20 percent of a community's area under water, verified residents received 105 dollars to evacuate and 210 dollars later to rebuild. The New Humanitarian's investigation, published August 31 and based on interviews with at least 10 flood-affected farmers, mostly in Unale ward, found people excluded because a submitted name did not exactly match government or bank records, and others cut because they could not pay for solar or generator charging and so missed verification calls in a district with no grid electricity and little mobile coverage; some took on debt expecting aid that never came. Payments also went to relatively wealthy people, to people outside the target area, and to multiple members of one household. GiveDirectly retracted its original claim of 4,600 households, said the programme verified 4,600 individuals, and updated its website to acknowledge the design excluded people with limited phone access and poor households that had already moved to higher ground. Why it matters: This is the first field-level audit of one of the most cited AI anticipatory-action pilots, and it separates the model from the plumbing: the forecast was not where people were lost, the enrolment, exact-match verification and call-back layers were. GiveDirectly says the design involved deliberate tradeoffs to reach more people faster and that it made the right ones; the same model is now being studied in Kenya and Mozambique, and a randomised controlled trial covering more than 100,000 families is planned for the flood season in Bangladesh's Jamuna River Basin, which makes Ibaji the baseline that trial will be measured against.
Microsoft Puts 5 Million Dollars Behind ALERTCalifornia's 1,300-Camera Wildfire Network, Split Between Engineering and Cloud Credit
Microsoft is investing 5 million dollars in ALERTCalifornia, the UC San Diego programme that runs more than 1,300 high-definition cameras across fire-prone terrain and uses AI to flag smoke and fire before anyone calls 911. Innovation News Network, reporting September 2, breaks the figure into 2 million dollars of in-kind technology development and a 3 million dollar Azure grant, with Microsoft's AI for Good Lab working on detection and its Public Safety Industry Team supplying cloud infrastructure, data pipelines and continuity during peak demand. The cameras rotate through 360 degrees roughly every two minutes and the system analyses millions of images a day, distinguishing smoke from clouds, fog and changing light. When ALERTCalifornia announced the collaboration on February 3, 2026, the network stood at 1,200 cameras and edge computing had lifted the frame rate from one frame every 20 seconds to one per second, accelerating detection by 10 to 30 minutes over the previous platform. The network remains open source and publicly viewable. Why it matters: The show reported on August 17 that ALERTCalifornia's cameras catch about half of California's wildfires before the first 911 call; what is new is the first disclosed dollar figure for the Microsoft relationship and its shape, engineering time plus cloud credit, attached to a network that has grown by roughly 100 cameras since February. Cloud credit is the line that pays for processing millions of images a day, the cost that has limited public-university camera networks elsewhere. ALERTCalifornia says the money will harden the platform and support expansion.
Coefficient Giving Raises Its 2026 Global-Health Commitment From 175 Million Dollars to 1 Billion, Sized to an AI Windfall That Has Not Arrived
Coefficient Giving, the San Francisco grant maker formerly known as Open Philanthropy and funded by Good Ventures, the foundation of Dustin Moskovitz and Cari Tuna, has raised its 2026 commitment to GiveWell's recommended global-health organisations from 175 million dollars to 1 billion, a nearly sixfold jump for a funder that has directed about 7 billion dollars since 2014. The first major grant under the new figure, announced last week with GiveWell, is 276 million dollars for roughly 90 million insecticide-treated bed nets in the Democratic Republic of Congo, which GiveWell estimates could prevent about 67,000 deaths, most of them children under five. The Chronicle of Philanthropy reports September 1 that Coefficient's leaders time the surge to a possible windfall from the AI boom: in a recent blog post the group wrote that if OpenAI, Anthropic and others go public, the approach could lay groundwork for more tech donors to pick up the baton, and that grantees need time to plan, hire and expand into new regions. One estimate cited, from Stripe's Nan Ransohoff, puts potential new annual philanthropic spending at 37 to 100 billion dollars if OpenAI and Anthropic list. Hala Hanna of MIT Solve told the Chronicle that consolidated decision-making power among the tech wealthy deserves scrutiny, since a small number of unelected individuals could reshape a funding category like global health almost overnight; David Goldberg of Founders Pledge cautioned against assuming tech donors share one set of priorities. Why it matters: A single funder has moved close to a billion dollars into evidence-ranked global-health delivery within one year and has said in writing that the reason is anticipated AI-company liquidity, making this the first large grant flow explicitly sized to AI wealth that has not yet been realised. Whether the windfall arrives or not, GiveWell-recommended organisations now hold a 2026 budget line that did not exist in August, and the debate over who directs that money is now attached to a concrete number rather than a forecast.
Yokosuka Runs the Same Earthquake Scenario Twice, With and Without Generative AI in the Disaster Headquarters, and Times the Difference
NTT East and the city of Yokosuka, on the Miura Peninsula south of Tokyo, began a trial on August 31 to test whether generative AI can support staff decisions inside a municipal disaster headquarters. In the first phase, held the same day, officers from the city's crisis management division worked an identical scenario, the hours after a Miura Peninsula fault-zone earthquake, twice: once without AI support and once with it. The application, built by NTT East's disaster-prevention research institute, is grounded in Yokosuka's regional disaster-prevention plan and in Disaster Response Operations BOOK 47, a guide that organises disaster work into 8 fields and 47 task types, developed with associate professor Munenori Numada of the University of Tokyo's Disaster Management Training Center; it proposes response priorities, candidate partner agencies and draft actions. Three measures are compared: time to answer per item, the appropriateness of the chosen counterpart and action, and whether a response is correctly fanned out to multiple agencies at once. The partners will also cross-check results against each officer's years of posting and disaster experience. The work sits under a May 2025 wide-area disaster cooperation agreement between NTT East and four cities and one town on the peninsula, which is surrounded by sea on three sides with limited road links. Why it matters: Most municipal generative-AI disaster pilots report that a tool was introduced; this one is a controlled comparison with named metrics and a staffing variable, which is what would let a city say whether AI support changes decision speed or accuracy, and for whom. The output the partners have committed to is a list of the tool's usefulness and its limits for headquarters work, with the explicit target of supplementing the experience gap as veteran disaster staff retire.
After the Bhote Koshi Flood, Sikkim Puts Water-Level Probes and Weather Stations on Its 16 Category-A Glacial Lakes
Sikkim holds 40 of the 189 high-hazard glacial lakes identified across the Indian Himalayas, and 16 of them are rated Category A, the most vulnerable class, by India's National Disaster Management Authority. With the Nepal-Tibet flood of August 26 still unfolding across the border, the state has stepped up glacial lake outburst flood safeguards: pressure-probe water-level sensors and automated weather stations are now installed at high-vulnerability sites including Kerang near Gurudongmar Lake and Dolma Sampa, and NDTV reported on September 1 that sensors are in place at the 16 Category-A lakes. The state has constituted a steering committee chaired by the chief secretary, a multidisciplinary task force for ground mitigation, and a Sikkim Commission on Glacier Hazard headed by Dr Akhilesh Gupta, a former central-government adviser, with the NDMA, the Ministry of Home Affairs, the Brahmaputra Board, the Ministry of Jal Shakti and the Central Water Commission as partners. Three sites are prioritised for engineering: a flood-retention structure at Dolma Sampa in the Lhonak complex, now in design; a solar-powered pumping proposal to draw down Shako Chho; and a protective moraine-plug wall in the Gurudongmar complex. Five field expeditions between 2024 and mid-2026 used electrical resistivity tomography, bathymetric mapping and differential GPS to characterise the lakes. South Lhonak, in the same complex, is the lake whose outburst in October 2023 destroyed the Teesta III dam and killed dozens. Why it matters: This is the concrete Indian response to the Nepal disaster the show has covered since August 27: instrumentation on the 16 lakes India itself ranks most dangerous, plus drawdown and retention engineering at the three complexes with the worst history. The sensors give downstream Teesta valley towns a measured water level rather than an inferred one, and the Dolma Sampa retention structure is the first hard-engineering answer to the South Lhonak failure of 2023. Whether a warning reaches people fast enough remains open; the ICIMOD gauge record the show cited on August 27 showed a nine-metre rise in thirty minutes.
Upcoming Events & Opportunities
CyberSafe Foundation with SANS Institute, AI Security Fellowship 2026 (Funding)
- Amount: fully funded nine-month training and certification pathway (SANS OnDemand courses, GIAC AI Platform Security and GIAC AI Security Automation Engineer); no tuition cost to fellows
- Deadline: September 10, 2026, 12:00 PM WAT
- Eligibility: women over 21, resident in an African country, English-speaking, at least three years of hands-on cybersecurity experience, able to commit to the full programme
- Programme dates: selection notifications September 30, 2026; orientation October 21, 2026; training begins October 26, 2026; completion July 30, 2027
- Apply: cybersafefoundation.org
NetHope Global Summit 2026
- October 26 to 29, 2026 in person; November 3, 2026 virtual programme, with main-stage sessions broadcast live from Amsterdam October 27 to 29
- Location: Amsterdam, and virtual
- The annual convening of the NetHope nonprofit-technology consortium; registration deadline not stated on the organiser page
- Register: nethope.org
Active Disaster Monitoring (GDACS/OCHA)
- Nepal and Tibet:** Bhote Koshi and Trishuli glacial-collapse flood, onset August 26, ongoing. GDACS Red alert; 955 deaths and 3,458 displaced in the GDACS record as of September 2, with news reporting on September 1 and 2 putting the toll above 1,000 and roughly 3,900 to 4,800 still out of contact, figures still moving. GDACS describes about 1 square kilometre of ice and rock collapsing in the upper Lhende Khola near the border, a seismic signal equivalent to magnitude 5.2, a temporary dam that then burst, and evidence of further lakes forming upstream. Search and rescue continues, including for workers trapped in hydropower tunnels at Rasuwagadhi; Nepal has said it is open to foreign rescue teams; fresh flood alerts were issued as river levels rose again on September 2. Source: gdacs.org.
- India:** Flooding, onset August 9, ongoing. GDACS Orange flood alert; 20 deaths and 240,570 displaced in the GDACS record as of its August 30 update. Source: gdacs.org.
- Northwest Pacific:** Tropical cyclone SAUDEL-26, August 18 to September 2 per GDACS, ongoing. GDACS Orange alert; maximum wind speed 213 km/h; 13.05 million people exposed to Category 1 winds or higher across Japan, China and neighbours. Source: gdacs.org.
- China:** Flooding, onset July 31, ongoing. GDACS Orange flood alert; 1 death and 309,701 displaced in the GDACS record as of its September 1 update. Source: gdacs.org.
- China:** M5.0 earthquake, August 28, 05:13 UTC, depth 10 km. GDACS Orange alert; 15.8 million people within 100 km. Ongoing GDACS monitoring. Source: gdacs.org.
- Somalia:** M6.0 earthquake, August 27, 145 km north of Caluula. USGS alert level green, offshore, no damage reports. Monitoring only. Source: earthquake.usgs.gov.
- Note: only major or ongoing-major disasters are featured; low-severity alerts are excluded per the major-only bar.
Sources: See individual stories above for full attribution.