- Use the web search and in-depth research features of Claude and Gemini for monitoring or desk review.
- Assess the reliability of cited sources and check every important statement.
- Build a documented context analysis from recognised public sources.
Keep the course at home: free PDF handout with exercises, key takeaways and quiz answers.
1Three research modes
You can get information from an assistant in three ways, of very different reliability:
- From memory (no search): the model answers from its training. Fast, but without a verifiable source and limited to the knowledge date. Reserve it for general notions.
- Web search: the assistant queries a search engine, reads a few pages and answers with links. Useful for current affairs and precise facts; quality depends on the pages found.
- In-depth research (Deep Research-type features in Gemini, Research mode in Claude, depending on the plan): the assistant plans a search, consults dozens of sources and produces a structured report with citations, in a few minutes.
In addition to these modes there is work on your own sources: you gather reliable documents (agency reports, analyses by recognised bodies) and upload them to NotebookLM or a Claude Project. This is often the most reliable method for serious analysis.
2Assessing cited sources
A citation is not a guarantee. Check:
- Does the link exist and does it really say what the assistant claims? Open it. A real source is sometimes cited for a statement it does not contain.
- Who publishes? UN agencies, governments, recognised humanitarian analysis bodies (for example ReliefWeb for aggregation, OCHA, ACAPS, IPC, IOM's Displacement Tracking Matrix), research institutes, established media, or an anonymous blog?
- When? A humanitarian situation changes fast; a figure from last year may be wrong today.
- What method? Estimate, survey, projection? With what geographical coverage?
- Independence: is the source a stakeholder (party to the conflict, interested company)?
Rule: any statement that goes into an official document (proposal, report, advocacy) is checked at the primary source, not in the AI's summary.
3Effective monitoring
Monitoring (funding, policies, the situation in a country, technical developments) is a very worthwhile use:
- define precise questions: "Which new calls for proposals for nutrition in West Africa have been published in the last 30 days on donors' official websites?" rather than "What's new?";
- impose sources: "Use ReliefWeb, UN agency websites and donors' official websites as a priority";
- require a format: table date / source / title / one-line summary / link;
- keep a shared monitoring log in which each item is checked before circulation.
Some plans allow scheduled tasks or regular summaries; they must remain under the responsibility of a person who checks before sharing.
4Context analysis
For a context note (new area of operation, proposal preparation, briefing a newcomer), a four-step method:
- 1. Framework: ask for a structure (politics, general security, economy, demography, humanitarian needs by sector, actors, access, seasonality).
- 2. Sources: run in-depth research, then select the reliable sources yourself.
- 3. Synthesis from these sources only, with citations.
- 4. Validation by colleagues in the country and local partners, who know what no online source says.
Online sources on a local context are often scarce, old or written from outside. Local actors, national NGOs and communities are essential sources: AI does not replace them.
5Limits and pitfalls
- Uneven coverage: high-profile crises are well documented; "forgotten crises" much less so. AI reflects this inequality.
- Languages: sources in local languages are found less often. Explicitly ask for sources in the country's language.
- Paid or closed content: much humanitarian data sits on restricted-access platforms; AI cannot see it, and must not access it without authorisation.
- Repetition effect: a figure repeated by ten websites often comes from a single primary source, sometimes an old one.
- Sensitive topics: for political or conflict-related topics, compare several sources with different perspectives and keep neutral vocabulary.
Example. The assistant states "2.3 million people in need" and cites three press articles. Tracing back, all three cite the same Humanitarian Needs Overview from the previous year. The right reflex: look for the latest version of that document and cite the primary source directly, with its date.
6Citing correctly
In a professional document:
- cite the primary source (the body that produced the figure), with title, date and link;
- do not cite "according to Claude" or "according to Gemini" as the source of a fact: the assistant is not a source, it is a tool;
- give the date of access for web pages;
- follow the required citation style (donor, journal, organisation).
You can ask AI to format your references in a given style, then check each element (author, title, year, link).
Exercise: checking an in-depth research summary
Situation. In-depth research on "the humanitarian situation in the Doro region" produces a 6-page report with 24 citations. You are preparing a two-page note for your management. Checking a sample, you find: 3 links no longer open; 2 links lead to pages that do not contain the cited figure; 6 citations come from the same press article; the main figure on food insecurity is 18 months old; no source is in the national language; no local actor is mentioned.
Task. What do you do before writing the note?
Answer. 1) Do not reuse the 5 problematic citations (3 dead links, 2 non-matching) without finding the real source. 2) Trace back to the primary source of the 6 citations from the same article. 3) Look for the latest official food insecurity analysis (for example the latest IPC analysis or equivalent) and replace the 18-month-old figure, or clearly state that it is old. 4) Supplement with sources in the national language and with exchanges with local partners and colleagues in the country. 5) Write the note from verified sources only, with the date of each figure and a clear mention of information gaps. 6) Given the error rate in the sample (5 problems among the citations checked), check all the citations used, not just a sample.
Key points
- Memory, web search, in-depth research, or your own sources: increasing reliability in this order for serious work.
- A citation is not a guarantee: open the link, check the publisher, the date, the method.
- Any statement in an official document is checked at the primary source.
- Local actors and communities remain sources that AI does not replace.
- The assistant is not a source: cite the body that produced the information.
Ask the other participants: they are taking the same course as you.
💼 Programme management jobs
Put this course into practice: the latest jobs in this field.
- Outreach WorkerOm Prakash Network of People Living with HIV/AIDS (OPNP+) · India
- VILLAGE PROJECT COST ANALYSIS CONSULTANTSpark MicroGrants · Malawi
- InterpreterHCR / UNHCR · Israel
- Field WorkersICMR Project - Jamia Hamdard · India
- ترینر برق (Re Announced)Afghans Better Life Organization (ABLO) · Afghanistan
Further reading: course references
- ICRC and Brussels Privacy Hub (2020). Handbook on Data Protection in Humanitarian Action (2nd edition). www.icrc.org
- Inter-Agency Standing Committee (IASC) (2023). Operational Guidance on Data Responsibility in Humanitarian Action. interagencystandingcommittee.org
- UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence. www.unesco.org
- United Nations (Chief Executives Board) (2022). Principles for the Ethical Use of Artificial Intelligence in the United Nations System. unsceb.org
- European Union (2024). Regulation (EU) 2024/1689 on artificial intelligence (AI Act). eur-lex.europa.eu
- European Union (2016). General Data Protection Regulation (GDPR). eur-lex.europa.eu
- CHS Alliance, Groupe URD and Sphere (2024). Core Humanitarian Standard on Quality and Accountability. www.corehumanitarianstandard.org
- Sphere Association (2018). The Sphere Handbook: Humanitarian Charter and Minimum Standards in Humanitarian Response. spherestandards.org
- OECD (2019). Better Criteria for Better Evaluation: revised DAC evaluation criteria and principles for use. www.oecd.org
- National Institute of Standards and Technology (NIST) (2023). AI Risk Management Framework (AI RMF 1.0). www.nist.gov
- Anthropic (2026). Claude documentation and help centre (features, privacy, plans). support.anthropic.com
- Google (2026). Help centre for Gemini Apps and Gemini for Google Workspace. support.google.com/gemini
- Anthropic (2026). Claude prompting guide (developer documentation). docs.anthropic.com
- OCHA (2026). ReliefWeb: humanitarian reports and maps. reliefweb.int
Weekly quiz
5 demanding questions to check your learning. Answer them all to complete the week.