Artificial intelligence, explained for humanitarian research

Software that performs tasks we normally associate with human reasoning: recognising patterns, predicting outcomes, generating language, or making decisions under uncertainty.

In humanitarian work, AI usually means statistical models trained on past records that produce a guess about something new. It is not a thinking machine. It is a pattern matcher, and its quality is bounded by the quality of the data you give it.

The branch of AI where a model improves at a task by being shown examples instead of being programmed step by step.

A model that forecasts cross-border refugee arrivals from past ACLED conflict events and UNHCR registration data is doing machine learning. It learns the relationship; an engineer does not encode it by hand.

A model trained on huge amounts of text that can read, write, summarise, translate, and answer questions in natural language.

LLMs power tools that draft situation summaries from OCHA reports or translate Arabic-language testimony into English. They are confident speakers and unreliable narrators: the words read well, but the facts must be checked against a primary source.

The collection of examples a model is shown so it learns the pattern you want it to reproduce.

If a model is trained mostly on English text from the Global North, it will be confidently wrong about Sudanese place names, Pashto dialects, and Syrian governorate boundaries. Training data is destiny.

The moment a trained model is used to produce an output, as opposed to the training phase when it is being built.

Each time a journalist asks our daily briefing tool to summarise yesterday’s ReliefWeb feed, that is inference. Inference is cheaper than training but it still costs electricity, water, and money.

When a generative model produces a statement that sounds plausible but is not supported by its source material or by reality.

A hallucinated casualty figure in a humanitarian briefing is not a quirky bug. It is a fact that can travel into a policy document. Every AI-assisted page on this site is cross-checked against UNHCR, IOM, IDMC, ACLED, or OCHA before publication.

An architecture where the model is forced to read a specific set of trusted documents before it answers, instead of relying on what it memorised during training.

Our country briefings use a RAG pattern: the model is shown only the latest verified primary sources for that country, then asked to synthesise. It cannot freely invent statistics because it is grounded in a closed source set.

A systematic skew in a model’s outputs caused by the data it was trained on, the labels it was given, or the choices its builders made.

A face-recognition system that performs worse on darker skin tones is biased. A displacement forecast that ignores undocumented IDPs in urban areas is biased. Bias in humanitarian AI is rarely visible until someone is left out of an aid distribution.

The physical building that houses the servers, cooling systems, and power infrastructure that AI runs on.

Every model query travels to a data centre somewhere on the planet. Those buildings draw measurable shares of national electricity and freshwater supplies, which is why AI’s climate cost is not abstract.

A model whose trained parameters are published so anyone can download, inspect, and run it on their own hardware.

Open-weight models matter for humanitarian work because they can be run inside a sanctioned country, on a field laptop, or behind a firewall without sending sensitive case data to a third-party server.