Industries / Government, education and community / Nonprofits
Program delivery
Charities and aid groups run services for people directly: food, cash, small loans, training, advice and places to meet.
7
Milestones achieved, from 1889 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1889
A house among neighbors
Jane Addams and Ellen Gates Starr leased a large house at Halsted and Polk Streets in Chicago, moved in, and ran services for the immigrant district around it. By its second year Hull-House drew 2,000 people a week to kindergarten classes, clubs and evening courses, and it later added a public kitchen, a gymnasium and an employment bureau, by the account of the Nobel committee, which gave Addams the 1931 Peace Prize. Addams took the idea from Toynbee Hall in London, which she had visited.
Credit
Jane Addams and Ellen Gates Starr
1967
A bank for food
In Phoenix, a mother told John van Hengel, a soup-kitchen volunteer, that she fed her children from soup kitchens and grocery-store dumpsters. He proposed a "food bank," where people and companies with excess food could deposit it and people in need could withdraw it, and his parish, St. Mary's Basilica, gave him $3,000 and an empty building. St. Mary's Food Bank calls itself the first food bank in the world; van Hengel went on to found America's Second Harvest, now Feeding America, and the Global Food Banking Network.
Credit
John van Hengel; St. Mary's Basilica, Phoenix
2006
Small loans win a Nobel
The Norwegian Nobel Committee split the 2006 Peace Prize between Muhammad Yunus and Grameen Bank of Bangladesh, which over three decades had turned lending to poor people with no collateral into a model copied in many countries. The committee called such small loans, known as microcredit, "an important liberating force," especially for women.
LessonSix randomized trials published together in 2015 found "modestly positive, but not transformative, effects" from wider access to microcredit.
Credit
Muhammad Yunus; Grameen Bank
2010
The merry-go-round pump
The PlayPump used a children's merry-go-round to pump water into a tank, with billboards on the tank meant to pay for upkeep. Its backers, among them PlayPumps International, the Case Foundation and USAID, planned to raise $60 million to install more than 4,000 across sub-Saharan Africa. By 2010 WaterAid, UNICEF and Mozambique's government had called the device too costly and hard to repair, many pumps had been swapped for hand pumps, and PlayPumps International had given its remaining stock to Water For People, as Circle of Blue reported.
LessonThe pumps were installed before communities were asked whether they wanted them, and before anyone settled how they would be repaired.
Credit
Trevor Field (inventor); PlayPumps International; the Case Foundation
2015
Graduating from extreme poverty
BRAC, founded in Bangladesh in 1972, began in 2002 to give its poorest members a productive asset such as livestock, plus food money, training and coaching, for 12 to 36 months; BRAC counts 2.4 million Bangladeshi families who have since left extreme poverty. Randomized trials of the same package, run by different groups for about 11,000 households in Ethiopia, Ghana, Honduras, India, Pakistan and Peru, found cost-effective gains in consumption and well-being that lasted at least a year after support ended.
Credit
BRAC (Fazle Hasan Abed); Banerjee, Duflo, Karlan, Udry and colleagues (the trials)
2016
Cash with no strings
GiveDirectly gives poor households money and sets no rules on how it is spent. In a randomized trial in rural Kenya, household spending rose from $158 to $193 a month nine months after transfers began (in dollars adjusted for local prices), and recipients reported large gains in psychological well-being. GiveDirectly says it has delivered more than $1 billion to over 2 million people since 2009.
Credit
GiveDirectly; Johannes Haushofer and Jeremy Shapiro (the trial)
2025
Aid answers, tested by language
Linguists from the charity Respond Crisis Translation, working with the research nonprofit Taraaz, put questions that refugees and asylum seekers ask to GPT-4o, Gemini 2.5 Flash and Mistral Small, in 655 evaluations. Answers in Arabic, Farsi, Pashto and Kurdish Sorani were less accurate and less useful than answers in English. In English, Gemini sometimes told a person with serious symptoms to see a doctor; in Pashto, Arabic and Kurdish it never did.
LessonA chatbot checked only in English can fail in the languages aid groups most need.
Credit
Respond Crisis Translation; Taraaz (Roya Pakzad)
Watch for
An AI help desk
The International Rescue Committee says Signpost, its digital help desk, has reached more than 20 million people in 20 countries in 12 languages, and that an early pilot of an AI version, SignpostChat, can make responses 92% faster. No outside measure of the accuracy of its answers was found.
Achieved whenAn outside evaluator publishes accuracy and safety results for SignpostChat's answers, graded by native speakers of its working languages.
Claimed by
International Rescue Committee, November 2026
A chatbot alongside cash
GiveDirectly offered 832 cash recipients in rural Rwanda a ChatGPT-based chatbot on WhatsApp next to one-time transfers of about $1,000. From November 2025 to April 2026 they sent it 21,169 messages, voice notes, photos and menu taps, including questions on family conflicts, sick children and market prices. GiveDirectly says it will next test cash alone against cash plus chatbot, and will not scale the tool without measurable gains.
Achieved whenThe cash-against-cash-plus-chatbot comparison reports economic results, checked by outside researchers.
Claimed by
GiveDirectly, June 2026
Open
01
Audit help listings
A check of a large directory of food pantries or social services, such as one region's 211 listings or a statewide pantry list. An agent reads each provider's own website and social posts and flags where the listed address, hours or eligibility rules differ; a person confirms each flag with a short call. The output is an error rate for each field and a list of corrections sent to the directory.
The briefCredit · unclaimed02
Test aid chatbots in Kinyarwanda
A public test set of about 400 questions in Kinyarwanda of the kind cash-transfer recipients ask: program rules, market prices, where to find services, how to plan a purchase. Native speakers who work in aid write the questions and vetted answers; no real recipients' messages are used. Several AI models answer each question, and two native speakers grade each answer for accuracy and usefulness. Health questions are scored only on whether the answer sends the person to a clinic.
The briefCredit · unclaimed
Open 01 · the brief
Audit help listings
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A check of a large directory of food pantries or social services, such as one region's 211 listings or a statewide pantry list. An agent reads each provider's own website and social posts and flags where the listed address, hours or eligibility rules differ; a person confirms each flag with a short call. The output is an error rate for each field and a list of corrections sent to the directory.
Why it is still open
The outside checks found are small. Researchers who visited 18 food resources listed in Kaiser Permanente's resource locator near one Washington clinic found five that did not match their listing or were unavailable. Michigan's auditor general reported in June 2026 that the contractor running Michigan 211 said it had missed the required level of database accuracy in over 85% of the quarters it reported; those are the contractor's own figures. A 2026 study of a Kansas pantry chatbot describes such information as fragmented and outdated. No larger outside audit with a published error rate was found.
Built with what exists
A web agent such as Browser Use (see the Note "Deleting 80% to start over"); Python and a spreadsheet; Open Referral's open data standards for service listings, so corrections reach directories in a form they can load.
First step
Pull 50 pantry listings for one county and check each by hand against the pantry's own posts, to set the error categories.
Done looks like
500 listings in one region checked, with the share wrong for address, phone, hours and eligibility, each error confirmed by a person; the check repeated three months later to measure how fast listings go stale; method and data published.
Open 02 · the brief
Test aid chatbots in Kinyarwanda
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of about 400 questions in Kinyarwanda of the kind cash-transfer recipients ask: program rules, market prices, where to find services, how to plan a purchase. Native speakers who work in aid write the questions and vetted answers; no real recipients' messages are used. Several AI models answer each question, and two native speakers grade each answer for accuracy and usefulness. Health questions are scored only on whether the answer sends the person to a clinic.
Why it is still open
GiveDirectly, which runs a chatbot pilot in Rwanda, says Kinyarwanda support across major models is "inconsistent and often poor." The closest tests found cover refugees' questions in Arabic, Farsi, Pashto and Kurdish Sorani, or cultural knowledge in 15 African languages (Afri-MCQA), not the practical questions aid recipients ask.
Built with what exists
The Multilingual AI Safety Evaluation Lab, Taraaz's open-source platform for grading model answers across languages; the Afri-MCQA set as a check on general language skill; AI models such as Claude, Gemini and GPT through their programming interfaces; a shared spreadsheet for graders.
First step
Ask ten Kinyarwanda-speaking aid workers to write ten questions each with model answers, and run them through two models.
Done looks like
At least four models scored on all 400 questions, with error rates by type, the agreement between the two human graders, and how closely an AI grader matches them; questions, answers and scores published under an open license.