BRENNEN

Industries / Government, education and community / Nonprofits

Disaster relief

Aid groups get food, water, shelter, medical care and cash to people hit by wars, earthquakes, floods and famines, as fast as they can.

7

Milestones achieved, from 1863 to 2025.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1863

    A committee for the wounded

    In February 1863 the body that became the International Committee of the Red Cross (ICRC) met for the first time in Geneva. One of its five founding members, Henry Dunant, had published A Memory of Solferino the year before, calling for better care of wounded soldiers. In August 1864 the committee persuaded governments to adopt the first Geneva Convention, which obliged armies to care for wounded soldiers of any side and made a red cross on white the emblem of medical services.

    icrc.org

    Credit

    Henry Dunant and the ICRC's founding committee

  2. 1996

    Rwanda's aid is judged

    After the 1994 genocide in Rwanda, donors led by Denmark's foreign ministry paid for a joint evaluation of the international response, published in March 1996. Its study of the aid effort, led by John Borton of ODI, a London research institute, faulted UN coordination, poor preparation for the refugee influx into Goma in July 1994, and some international NGOs (non-governmental organizations, here private aid groups) whose "unprofessional and irresponsible behaviour" had probably cost lives. Aid groups answered with voluntary standards: the Sphere Project's Humanitarian Charter and Minimum Standards, begun just before the report, and ALNAP, a network for learning from evaluations, set up in 1997.

    LessonWriting in 2014, Borton noted that during the 2010 Haiti earthquake response it was "still possible" for unprofessional groups to set up operations, so voluntary standards alone did not close the gap.

    odihpn.org

    Credit

    the Joint Evaluation of Emergency Assistance to Rwanda (Danish Ministry of Foreign Affairs; ODI for the humanitarian study); the Steering Committee for Humanitarian Response (Sphere)

  3. 2006

    A fund ready before disasters

    On 9 March 2006 the UN launched the Central Emergency Response Fund (CERF), approved by the General Assembly in December 2005, so that UN agencies could get money in the first days of an emergency instead of waiting for donors to answer an appeal. It was set up to provide up to $500 million in grants and loans, and by launch day more than 30 governments had pledged over $200 million, by the UN Foundation's count.

    unfoundation.org

    Credit

    UN General Assembly; UN Office for the Coordination of Humanitarian Affairs (OCHA); Secretary-General Kofi Annan, who had called for the fund a year earlier

  4. 2015

    Six homes in Haiti

    After the 2010 earthquake in Haiti, the American Red Cross raised nearly half a billion dollars. In June 2015 ProPublica and NPR reported that it had built six permanent homes in all of Haiti, while it said it had provided homes to more than 130,000 people; that figure counted construction training, rental aid and several thousand transitional shelters. The Red Cross said it had scaled back its housing plans because it could not get land rights.

    LessonA count of "people helped" says little until it says what each person received.

    propublica.org

    Credit

    Justin Elliott (ProPublica) and Laura Sullivan (NPR)

  5. 2020

    Cash before the flood

    In 2020 WFP, the UN's food agency, sent one-off cash transfers to households in Bangladesh before a forecast river flood peaked, so the payment was set off by the forecast rather than by the damage. Researchers at the Centre for Disaster Protection compared recipients with similar households that got nothing: recipients were 36% less likely to go a day without eating, and more likely to evacuate (by 12%) and to move livestock to safety (by 17%). In September 2026 WFP's evaluation office and the World Bank's Development Impact Evaluation unit (DIME) reported a randomized trial in Bangladesh in which cash sent before floods peaked did significantly better than the same cash sent one or two months after the peak.

    anticipation-hub.orgwfp.org

    Credit

    WFP Bangladesh; Ashley Pople, Stefan Dercon, Ruth Vargas Hill and Ben Brunckhorst (Centre for Disaster Protection); WFP Office of Evaluation and World Bank DIME

  6. 2022

    Cash at record scale

    After Russia's invasion of Ukraine, aid groups delivered US$1.7 billion as cash and vouchers between February 2022 and October 2023, 46% of it as multipurpose cash that people could spend as they chose. ALNAP's 2026 review of the aid system quotes this as the "fastest and largest cash programming scale-up in history". The same review reports a CALP Network study of US$11.4 billion of aid spent from 2014 to 2025: 71–77% of a cash or voucher budget reached recipients directly, against 54–56% for food aid handed out in kind.

    LessonCash was also easy to cut: cash and voucher funding fell 11% in 2025, to US$7.3 billion, by ALNAP's figures.

    alnap.org

    Credit

    aid agencies in Ukraine; the CALP Network (the cost study); ALNAP (the review)

  7. 2025

    Aid funding drops 35%

    Donor funding to UN-coordinated aid appeals peaked at US$31.6 billion in 2022, then fell every year, and by 35% between 2024 and 2025 alone, to US$16.6 billion. Only 35% of the 2025 target was met, the largest gap in the history of the appeals, and only 56% of a cut-down, "prioritised" target of US$29.5 billion. These are ALNAP's figures in its State of the Humanitarian System 2026, which describes 2025 as the year the United States closed its aid agency, USAID.

    LessonThe means of learning shrank with the money: USAID's evaluation teams closed, and by mid-2026 one count found seven evidence and data platforms shut since the cuts and eight more interrupted or uncertain.

    alnap.orgalnap.org

    Credit

    ALNAP (the figures)

Watch for

  1. Forecasts that pay out early

    GiveDirectly, a charity that gives cash directly to people in poverty, says it used Google's AI flood forecasts and local knowledge to pick six flood-prone wards in Kogi State, Nigeria. It says it paid 3,250 people between 29 August and 6 October 2024, weeks before the flood peaked on 21 October, and more than 4,600 in all. From its own surveys, it reports that recipients' incomes "more than doubled" and food insecurity "dropped by 90%". Google says the International Rescue Committee used the same forecasts in its trigger for cash to 3,000 households in Adamawa State.

    Achieved whenAn outside evaluator compares recipients with similar households that were not paid, as the Centre for Disaster Protection did in Bangladesh in 2020, and publishes the result.

    givedirectly.orgblog.googlethenewhumanitarian.org

    Claimed by

    GiveDirectly, June 2025 (with Google, February 2025). In August 2026 The New Humanitarian reported, from interviews with flood-hit farmers, that some of the poorest residents were left out: GiveDirectly said 939 applicants failed its identity checks and that it could not reach 3,000 applicants by phone, and after questions it corrected "4,600 households" to "4,600 individuals". GiveDirectly calls paying by area "a deliberate tradeoff: speed and coverage over precision targeting"

Open

  1. 01

    Score flood triggers

    A public scorecard for flood "anticipatory action" plans, in which money or supplies go out when a forecast crosses a set threshold. For each plan, record when its trigger fired or stayed silent, the forecast it used, and whether and when the flood came, checked against free radar satellite images that show water through cloud. Report hits, misses, false alarms and the warning time actually achieved.

    The brief
    Credit · unclaimed
  2. 02

    Lessons that repeat

    A public table of the recommendations made in independent evaluations of major relief operations since 1996, with each one linked to later evaluations that find the same problem again. An AI model reads the reports and groups recommendations that say the same thing; a person checks every group.

    The brief
    Credit · unclaimed

Open 01 · the brief

Score flood triggers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public scorecard for flood "anticipatory action" plans, in which money or supplies go out when a forecast crosses a set threshold. For each plan, record when its trigger fired or stayed silent, the forecast it used, and whether and when the flood came, checked against free radar satellite images that show water through cloud. Report hits, misses, false alarms and the warning time actually achieved.

Why it is still open

Trigger rules are "usually buried in PDFs and Word files", which "makes it hard to compare, test and improve triggers across countries and organisations", according to an open catalogue started in 2025 that encodes the rules but does not score how they performed. The skill tests found cover one country at a time and use past forecasts rather than real activations; a 2025 Netherlands Red Cross study of Mali's flood trigger says its assessment is "limited by the quality of impact data".

github.comre.public.polimi.it

Built with what exists

The Anticipation Hub's Trigger Database and the open trigger catalogue for the rules; forecasts from Google Flood Hub and GloFAS, the EU's Global Flood Awareness System; Sentinel-1 radar images through Copernicus or Google Earth Engine for flood extent; an AI model such as Claude or Gemini to pull trigger rules and activation dates out of plan PDFs and reports.

  1. First step

    Encode three documented cases, WFP's 2020 Bangladesh activation, GiveDirectly's 2024 Kogi payments and the Mali Red Cross trigger, and pull the satellite flood maps for the days each fired.

  2. Done looks like

    At least 30 flood triggers in at least five countries scored for 2020 to 2026, each with counts of hits, misses and false alarms and a median warning time in days; every scored activation checked by a second person; data and code public.

Open 02 · the brief

Lessons that repeat

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public table of the recommendations made in independent evaluations of major relief operations since 1996, with each one linked to later evaluations that find the same problem again. An AI model reads the reports and groups recommendations that say the same thing; a person checks every group.

Why it is still open

The 1996 Rwanda evaluation criticized unprofessional aid groups, and the leader of its humanitarian study wrote in 2014 that the same thing was still possible in Haiti in 2010. ALNAP's 2022 "Lessons of Lessons" paper went back by hand through its own 20 lessons papers from 2003 to 2021 and, in its director's words, found recommendations on handing work to local groups and on accountability to the people aid serves "almost word for word the same as they were 20 years ago"; it summarized lessons papers, not the evaluations under them. The closest automated work found, a 2024 study of 631 UNICEF evaluation reports, sorted findings by outcome for one agency and did not trace recommendations across agencies or decades.

odihpn.orgalnap.orgebi.ac.uk

Built with what exists

ALNAP's HELP library of humanitarian evaluations; pdftotext to turn PDFs into text; an AI model with a long context window, such as Claude or Gemini, to extract and group recommendations (see the Note "Seven more flagged"); a shared spreadsheet for the human checks.

  1. First step

    Take the 1996 Rwanda humanitarian study and two evaluations of the 2010 Haiti response, extract their recommendations by hand and with a model, and compare the two lists.

  2. Done looks like

    At least 100 evaluations of at least 20 major responses from 1996 to 2026 coded; at least 1,000 recommendations extracted; on a random 200, the AI's extraction and grouping agree with two human coders at least 85% of the time; the 20 most repeated recommendations published with the reports that made them.