The Geneva Learning Foundation has opened a conversational way into three years of frontline evidence on climate change and health. You can now ask questions, in plain language, of the reports and accounts that thousands of health workers built, and get answers drawn only from what they actually said. Learn more about The Geneva Learning Foundation’s work on climate change and health…
You can open it here: Chat with the climate change and health evidence.
What it is
The interface is a chat built on Google’s NotebookLM.
We have loaded it with the body of work this review describes: the 2023 eyewitness report from more than 1,200 health workers, the 2026 insights report on local action from Teach to Reach 11, and the analysis and recommendations that surround them.
- You type a question.
- The tool answers using only those sources, and it shows you where in the documents each part of its answer came from.
What you get is a conversation with a bounded set of evidence.
A search box would hand you a list of links to sort through. A general chatbot would answer from the whole internet, with no way to tell where its claims came from. This interface stays inside the documents we gave it, and every claim it makes can be traced back to the health worker or the report that is its source.
How it works
Open the link, type a question, and read the answer. Each answer carries small numbered citations that point to the exact passage in the underlying documents, so you can check the source and read the fuller context.
A few questions show the range of what it can do.
- A funder might ask what local solutions health workers have already built to keep maternity care running during floods, and what those solutions cost.
- A researcher might ask how health workers describe the link between waste, flooding, and malaria in urban settings, and which countries those accounts come from.
- A national planner might ask what health workers say they need from district budgets, in their own words.
- A health worker might simply ask whether anyone else has faced the problem they are facing this week, and what those colleagues tried.
The tool answers each one from the evidence, names its sources, and leaves the reader free to go and read the original.
Why it matters for health workers first
The Geneva Learning Foundation makes a commitment to the health workers who share what they know. What we learn from them, and what they learn from each other, we give back.
The reports are one form of giving back. This interface is another, and in some ways a more useful one. A 158-page report is a serious thing to read on a phone, on a slow connection, between patients. A question is not.
A community health worker in Kinshasa or Kano can now ask the collective experience of tens of thousands of peers a direct question and get a direct answer, with the sources named, without reading the whole report to find the one account that speaks to their situation.
This keeps faith with how the knowledge was made. Health workers wrote these accounts to learn from one another. The chat interface lets them do exactly that, on demand, in the moment a problem arises rather than months later. The knowledge returns to the people who created it, in a form they can use while they work.
Why it matters for everyone else
For researchers, the interface is a fast way to interrogate a large qualitative dataset and locate the accounts worth studying in full, with citations that lead straight back to the primary text. It does not replace reading the reports. It makes the reports navigable.
For policymakers and country planners, it answers the question that global frameworks keep asking and cannot answer on their own: what is actually happening at community level, what is blocking action, and what local solutions already work. The Belem Health Action Plan and the Lancet Countdown both depend on knowing this. Here the knowing is one question away, grounded in the testimony of the people closest to the impacts.
For funders and partners, it is a way to test the evidence behind the case before committing to it. Ask what works, ask what it costs, ask who is already doing it, and read the answer against the sources.
What it is honest about
The interface answers only from the documents loaded into it. It will not invent a statistic or reach for a fact from elsewhere, and when the evidence does not contain an answer, that itself is worth knowing.
The accounts it draws on are self-reported by health workers and, as the reports state, not independently verified. The tool inherits that status, and it is a feature rather than a flaw. The point of this body of work is to let frontline experience be heard and examined as evidence, and the chat interface is built so that anyone can examine it directly, trace every claim to its source, and decide for themselves what it shows.
The evidence is no longer locked inside a long document that few will finish. It is open to a question. The people who built this knowledge can ask it. So can the people whose decisions will determine whether it is ever acted upon.
For researchers: how to cite the evidence in TGLF’s climate and health knowledge base
A growing number of researchers want to draw on what frontline health workers have documented about climate change and health, and then hesitate at the citation. The knowledge is real, the sources are fixed and published, and there is now settled guidance for citing both the chat interface and the reports it draws on.
Two things you can cite, and they are not the same
The chat interface is built on Google’s NotebookLM, and it differs from a general chatbot in the one way that matters for citation. It answers only from a fixed, declared set of sources, the TGLF reports loaded into it, and it shows you the passage behind each answer.
That means you are almost never citing the tool alone. You are using the tool to locate a claim that lives in a formally published report. So there are two distinct citable objects, and good practice often cites both.
The first is the underlying report. When the interface points you to a finding, the authoritative source is the report itself, and the report is a formal publication with named authors, a publisher, a date, and a registered DOI. Cite it exactly as you would cite any grey-literature report or dataset. No special handling is required, because nothing about the source is novel once you are citing the document.
The second is the chat session, when the conversation itself is what you are referring to, for instance when you quote a synthesis the tool produced or want a reader to be able to reproduce your query. Here you follow the current guidance for citing generative AI.
How to cite the report
The reports in the knowledge base are published on Zenodo with DOIs, which makes them straightforward to cite in any style. The two anchor sources are below.
In APA, the 2026 report reads: Jones, I., Njua Mbuh, C., Steed, I., and Sadki, R. (2026). Teach to Reach 11: Local action to mitigate the impact of the climate crisis on health. The Geneva Learning Foundation. https://doi.org/10.5281/zenodo.18246203
The 2023 eyewitness report reads: Jones, I., Mbuh, C., Sadki, R., Eller, K., and Rhoda, D. (2023). On the frontline of climate change and health: A health worker eyewitness report. The Geneva Learning Foundation. https://doi.org/10.5281/zenodo.10204660
When you quote a specific health worker, name them as the source within the report, exactly as the report does, and cite the report as the container. The contributors are named, and attributing their words to them by name is the respectful and accurate practice.
How to cite the chat session
The major style authorities updated their guidance in 2024 and 2025, and the rules are now stable enough to follow with confidence.
The Modern Language Association revised its guidance on 13 August 2025. MLA does not treat the AI tool as an author. You describe what was generated, name the tool in the container element, name the model in the version element, name the company as publisher, give the date the content was generated, and give a stable, shareable URL where one exists.
A works-cited entry for the climate knowledge base would read along these lines:
“Summary of how health workers kept maternity care running during floods” prompt. NotebookLM, Google, 16 June 2026, notebooklm.google.com/notebook/70ed4dee-d717-4b97-925c-b9d678db1127/preview.
The American Psychological Association treats the company that made the tool as the author, names the tool, and links to it. APA also gives a direct instruction that fits this case well: when an AI tool surfaces secondary sources, cite both the tool and those secondary sources, after confirming the sources are genuine. With NotebookLM that confirmation is easy, because the underlying reports are right there with DOIs.
An APA reference for the session reads:
Google. (2026). NotebookLM [Large language model]. https://notebooklm.google.com/notebook/70ed4dee-d717-4b97-925c-b9d678db1127/preview
In Chicago, describe the interaction and the tool in a note, give the date, and link to the session.
In every style, the principle is the same. State which prompt produced the content, name the tool and model, give the date, and provide the most stable link available. Then, wherever the content traces to a report, cite the report as well.
Do not reject a new kind of source by reflex
Some researchers will pause at the idea of citing a knowledge base of frontline testimony, on the instinct that it is not the kind of evidence that belongs in a reference list. That instinct deserves examination rather than obedience.
Qualitative evidence from practitioners is not new to science. Interview studies, focus groups, ethnographies, and field reports have long been cited as evidence, and they are no less self-reported than the accounts here. What is new is only the scale at which this testimony was gathered and the conversational means of reaching it. The novelty is in the access, not in the epistemic category.
A health worker’s documented account of how a flood disrupted antenatal care is data of exactly the kind that qualitative health research has always used. That it can now be reached through a chat interface changes how you find it, not what it is.
The interface itself is built to meet the standard that makes evidence citable. It is bounded, its sources are declared, and every claim traces to a passage you can read in full. That is a higher standard of traceability than a general language model offers, and it is the same standard a careful reader applies to any secondary source.
If you remain reluctant, cite the reports directly
For a researcher who is still not ready to cite an AI-mediated source, the path is short and entirely conventional. Use the chat interface as a finding aid, locate the report and the page the claim comes from, read it in context, and cite the report.
The reports are peer-contributed, formally published, and carry DOIs. Citing them requires no methodological innovation and no editorial indulgence from a journal. In that case the chat interface has still done its work. It helped you find the evidence faster. The citation that lands in your manuscript is a published report with a permanent identifier, indistinguishable in form from any other report you would cite. The knowledge base lowers the cost of discovery without asking you to lower your standards.
Either way, the evidence that thousands of health workers built is now within reach of the literature, and it can enter that literature with full and proper attribution.
Learn more
- Modern Language Association. (2025, August 13). How do I cite generative AI in MLA style? https://style.mla.org/citing-generative-ai-updated-revised/
- American Psychological Association. How to cite ChatGPT and other generative AI. APA Style blog. https://apastyle.apa.org/blog/how-to-cite-chatgpt