What thousands of health workers know, made visible, usable, and impossible to ignore
The Insights Team is the research and analysis group of The Geneva Learning Foundation. We study what health and humanitarian workers tell us about their own practice, return what we learn to the communities who shared it, and put their findings in front of the ministries, agencies, researchers, and partners who set strategy and allocate resources.
Our evidence comes from the Foundation’s own programmes and events, which reach more than 120 countries in English, French, and Spanish. Each Teach to Reach gathering generates between 1,500 and 3,000 first-person accounts of practice. A single course generates hundreds of project submissions, peer reviews, and forum exchanges. Since 2026, an AI-led research process reads all of it, with our researchers deciding what to ask, what it means, and what we are willing to claim.
What we produce
| Output | Description |
|---|---|
| Collections of shared experiences | First-person accounts organized by country and theme, returned to contributors within weeks |
| Insights reports | Thematic analysis of hundreds or thousands of accounts, with methods and limitations stated, published with DOIs in an open repository |
| Field intelligence | Analysis that treats a course or event as a research instrument, asking what practitioners collectively know that the literature does not |
| Replication bundles | Data, code, analytical framework, reliability measurements, and a fingerprint for every file, so any finding can be independently reproduced |
| Live insights sessions | Sessions where contributors present their own findings, with global partners in the audience, released afterwards as recordings and podcasts |
| Custom inquiry | Investigation of a specific question registered by a partner, from question design through publication and use |
Who we work with, and what they get
National and sub-national programme managers. Current, granular evidence about how implementation is going in their own districts, including obstacles that do not survive the trip up the reporting chain.
Global agencies and technical partners. Direct feedback on whether guidance survives contact with a health facility, from the people implementing it, before the next revision. Our events carry no ceiling on attendance and no travel budget, so participation is not limited to those who can be flown in.
Organizations that allocate resources. Early warning about what is failing and why, documentation of what local actors already achieve with the resources they have, and a means of checking whether an investment changed anything on the ground.
Researchers. First-person qualitative data at a scale rarely assembled, documented methods, replication bundles, and the option to register a question before analysis begins.
Health and humanitarian workers. What thousands of peers in comparable conditions have already tried, fast enough to be useful, and recognition for knowledge their own systems rarely credit.
Why the Insights Team exists
Health workers hold knowledge that no one else has, because they are present every day. A midwife in Cameroon told us that when floods come, the men who can paddle canoes ferry women to the health facility. A disease control officer in Ghana reported that harmattan dust fills his clinic with respiratory cases and proposed distributing masks. A nurse in the Democratic Republic of the Congo described a women’s solidarity fund paying for maternity stays when roads became impassable. None of this appears in the literature. All of it is actionable, and testable.
Information in global health mostly travels one way. Local actors report coverage, stockouts, and case counts upward, and guidance written elsewhere comes back down. The experience that explains the numbers is discarded as anecdote.
Our mission follows from that gap, and has two parts.
Give back what we learn. The communities who trust us with their experience are the first audience for every analysis we produce. When a health worker in a conflict zone spends thirty minutes writing about a family she persuaded, she has done unpaid work for the global good. Returning the value of that work quickly, and in a form she can use, is a condition of asking for it at all. This principle has governed the design of our research from the start.
Make these voices count where decisions are made. We name contributors, quote them word for word, and invite them to present their own findings to the institutions that shape their working conditions. The aim is not visibility for its own sake. It is response: guidance, strategy, and funding that change because of what practitioners reported. Most organizations use evidence to run an existing programme slightly better; the more valuable use is to question whether it is the right programme. We have seen that happen, when global experts who came to present stayed to listen and went home to redesign what they were offering.
Charlotte Mbuh, who began as a sub-national health worker in Cameroon and now directs our work, stated the premise at COP28: “What we know, we know because we are here every day.” Dr Kate O’Brien, who leads immunization at the World Health Organization, described the implication: “The lessons and learning are coming from people who are working in the community. That is where vaccination actually happens.”
How we work
An AI-led process, with researchers accountable
Reading every contribution used to be impossible, which forced a choice between sampling the data and losing rare accounts, or limiting how much we collected. Since 2026, machine readers work through every submission, every review, and every reflection, applying an analytical framework our researchers build and lock before reading begins.
What this changes:
- Full coverage. Nothing is sampled. An unusual observation from one nurse in one district reaches the final report.
- Relevant timing. A health worker facing an outbreak this month can read what colleagues learned last month, rather than waiting years for a publication.
- Verbatim quotation. Software checks every quoted sentence against the original submission, character by character. If it cannot find the words in the source data, we do not publish them as a quotation. In July 2026 this check stopped two completed reports until three sentences that a drafting model had smoothed were restored to the words their authors wrote.
- Numbers with stated limits. We measure how reliably our analysis assigns meaning to text, publish those measurements including the poor ones, and state in each report how far a figure can be pushed. Where the method undercounts a theme, the percentage is published as a floor rather than a finding.
- Hypotheses fixed in advance. Where we hold a hypothesis, we record what would disprove it and lock it before analysis. The data have contradicted us more than once.
- Independent reproducibility. Reports ship with the data, the code, and the analytical framework, so a finding can be re-run rather than trusted.
- Human decisions, documented. Researchers determine the questions, the consent conditions, the interpretation, and the claims. The boundary between their judgment and machine work is written down and open to inspection.
We publish our methods, our reasoning, and our failures, including a production run that went wrong and what we changed as a result. The code is open source.
Where the data comes from
The Foundation’s community includes more than 4,000 locally led organizations alongside individual practitioners. Teach to Reach, the event series that anchors it, grew from 2,604 participants in 2021 to more than 24,000. About half of participants work for government and half for civil society. Most serve remote rural areas, nearly half serve the urban poor, a quarter work with refugees or displaced populations, and one in five work in areas of active armed conflict.
Participation on this scale rests on trust built over a decade. A global survey on climate and health drew 6,436 responses, most from sub-national staff, with the majority reaching us through this community, which is what ensured coverage of the most climate-vulnerable places. When 734 practitioners each produced a peer-reviewed case study on overcoming vaccine hesitancy, for a report prefaced by Heidi Larson of the Vaccine Confidence Project, they did so knowing where their words would be used.
From insight to use
Our reports feed courses, planning tools, and implementation support, so findings become something a practitioner can apply in her own district within weeks. One report on climate change and health contained enough material that we built a certificate programme around it.
Independent researchers at the Centre for Change and Complexity in Learning, University of South Australia, examined an early exercise of ours: eight in ten participants used the ideas their peers had shared, more than nine in ten of those found them useful in their work, and two thirds of the ideas cited in their subsequent plans came from peers at a different level of the health system. A district officer in Kenya described the effect: “I found one idea in the Ideas Engine helpful. Implementing it actually helped us improve the trust of caregivers.”
Working with us
- Bring a question. We will say whether this community can answer it, and what an answer would and would not establish.
- Fund a line of inquiry, from question design through publication and use.
- License datasets for research, or replicate a published analysis from its bundle.
- Join a live session and hear contributors present their own findings.
- Partner on data collection design, so the questions asked reflect what you and the community both need to know.
Contact the Insights Team through the Foundation’s partnerships team.
Methods, limits, and permissions
Languages. Contributions come from anglophone, francophone, and increasingly hispanophone participants. Reports appear in English, with French translation for major outputs.
Signals, not prevalence. A nurse who notices more heatstroke has produced a signal, not a prevalence estimate, and we do not present one as the other. Our claim is that thousands of such signals form patterns with a speed and granularity conventional studies cannot match, and that they generate hypotheses worth testing.
Known limitations. Much of our data is self-reported and unverified, and every report says so. Some contributors are self-selecting practitioners committed to helping peers, who may not be representative. Where our analysis skews, we publish the direction of the skew and restrict how the figure may be cited. Validation of our methods has so far been internal, which is why the code and replication bundles are open.
Use and interpretation. Practices shared by peers require adaptation between contexts, and that judgment rests with whoever adapts them. Including an experience in a report is not an endorsement of it, and we advise against generalizing from a single case.
Consent and attribution. Every contribution we use carries its author’s permission, checked case by case, with clear separation between material that may be quoted with attribution, material that may be quoted anonymously, and material that may not be used at all. Silence is never treated as consent.
Quality review. Practitioners vet insights for accuracy and credibility before publication, and peer feedback on datasets is used to improve them.