What Artificial Intelligence Actually Is, In Plain Language

A working definition

Artificial intelligence, in the form most people now encounter, is software that has been trained on examples to produce a useful guess about something new. That is a deflationary definition, and it is the right place to start. The chatbot that drafts a press release, the model that flags a suspicious bank transaction, the system that recommends what you watch next on a streaming service, and the tool that translates a witness statement from Arabic to English are all doing the same fundamental thing. They have seen many past examples, they have extracted statistical regularities from those examples, and they apply those regularities to a new input.

It is worth being blunt: today's AI is not thinking. It is not conscious. It does not understand the words it produces. It is extraordinarily good at imitating patterns it has been shown, which is why its output often feels like understanding. The distinction matters because it tells you where the technology is reliable and where it is not.

How a model is actually built

Building a modern AI model has three rough phases.

First comes the data collection phase. Engineers assemble a very large body of examples. For a language model that might mean trillions of words of text scraped from the public web, books, code repositories, and licensed datasets. For an image model it might be billions of image and caption pairs. For a humanitarian application it might be a much smaller, much more careful set of validated records from sources like the UNHCR Refugee Data Finder, the IOM Displacement Tracking Matrix, or the IDMC Global Internal Displacement Database. If you are new to how the sector defines its core terms, our explainer on what counts as a humanitarian crisis walks through who is counted and why.

Second comes the training phase. The model, which begins as a structure with billions of randomly initialised numbers, is shown the examples one batch at a time. After each batch the numbers are nudged slightly so that the model's prediction for that batch gets a little closer to the right answer. Repeat this hundreds of billions of times and the model becomes startlingly good at the narrow task it was trained on. This phase consumes the famous quantities of electricity and silicon you read about in the press.

Third comes inference. The trained model is deployed, and end users send it inputs. Each input costs a little energy and a little money to process. Inference is what happens every time a journalist asks our daily briefing tool to summarise a ReliefWeb situation report.

Why training data is destiny

If you remember one thing about AI, remember this: a model is a mirror of its training data. A language model trained mostly on English text from North American and European sources will be confidently fluent in English and confidently wrong about the names of Sudanese towns, the spelling conventions of Pashto, or the administrative boundaries of Syrian governorates. A model trained on medical images from one hospital network will quietly fail when shown images from a clinic in a country it has never seen.

This is not a fixable bug. It is a structural property of the technology. The only way to make a model good at a thing is to show it good examples of that thing. That is why the humanitarian sector cannot simply lift a general-purpose chatbot off the shelf and trust it to summarise a UNHCR report. The model has to be either grounded in primary sources at query time, or specialised by additional training, or both.

The two families you keep hearing about

Most public conversation about AI today is really about two families of models.

Discriminative models answer the question: given this input, which of these labels fits best? The spam filter, the credit-scoring model, the satellite-imagery classifier that flags newly destroyed buildings in Gaza, and the model that predicts which displaced households are most likely to fall below a food-security threshold are all discriminative. They produce a label, a score, or a probability.

Generative models answer a different question: given this input, produce a plausible new output of the same kind. Chatbots, image generators, voice cloners, and the tools that draft summaries of long reports are all generative. They produce new content rather than labels. Generative models are the noisy revolution of the last few years, and they are the ones whose mistakes are hardest to catch because their mistakes look like fluent prose.

What AI is reliably good at right now

It is reliably good at five things: classifying inputs against a known taxonomy, drafting first passes of structured text, translating between widely spoken languages, summarising long documents into shorter ones, and finding patterns in numerical data that a human analyst would take weeks to surface manually. Each of those capabilities has a clear humanitarian use. A model can triage incoming case notes by urgency. It can draft the first version of a daily situation summary. It can translate testimony recorded in one language into another so it can be reviewed faster. It can compress a 200-page Humanitarian Needs Overview into a 500-word briefing. It can flag the camps in a dataset whose population trajectory is anomalous compared to the rest.

What AI is unreliable at, and dangerous around

It is unreliable at producing facts on demand. A generative model that is asked "how many internally displaced people are in Sudan" will produce a confident, fluent answer that may be entirely fabricated. It does not know the number. It knows what an answer to that question typically sounds like. This is why every AI-assisted page on this site is cited inline to a primary source and why we publish an editorial standards page describing exactly how AI is used in our work.

It is also unreliable in any context where the cost of an error is borne by someone who was not consulted. An aid-allocation model that under-counts urban IDPs is not a technical curiosity. It is a decision that withholds assistance from people who needed it. A face-recognition system used at a registration point that performs worse on darker skin tones is not a quirk of the training set. It is a system that delays or denies status determinations for a specific group of refugees. Our guide to understanding displacement explains why these populations are systematically under-counted to begin with.

How to read AI claims without being taken in

When a vendor, a government, or a research paper makes a claim about an AI system, four questions usually expose how much of the claim is real.

What was the training data, and who is missing from it? What is the system being asked to do at inference time, and is that the same as what it was trained on? What is the error rate, broken down by the groups whose lives the system will affect? What happens, in practice, when the model is wrong: does a human catch the error before it causes harm, or does the error become the decision?

If those questions cannot be answered, the system is not ready for use in any context where its mistakes will matter. That is true whether the system is deciding which loans to approve, which neighbourhoods to police, or which households on a returnee list to verify first.

Where to go next

If you want to understand the specific kind of AI that powers modern chatbots and summarisation tools, the next piece in this series is on large language models. If you are more interested in the physical cost of running these systems at scale, the data-centre piece walks through electricity and water. If you want to see how we apply these tools to displacement data without sliding into the failure modes above, the humanitarian AI piece is the practitioner-facing one.

The honest summary is that AI is a useful, fallible, electricity-hungry tool. Used carefully, with primary sources in the loop, it speeds up work that would otherwise be impossible at the volume the humanitarian sector now needs. Used carelessly, it manufactures plausible nonsense at industrial scale. The interesting question is no longer whether to use it. The interesting question is what discipline you put around it before you do.

Sources

• UNHCR Refugee Data Finder, the official refugee population dataset cited throughout this piece.
• IOM Displacement Tracking Matrix (DTM), the primary source for movement and mobility figures.
• IDMC Global Internal Displacement Database, the reference dataset for internally displaced persons.
• ReliefWeb, OCHA's humanitarian information portal used as a primary source for situation reports.
• Stanford 2024 AI Index Report, background reading for the state of model capabilities referenced above.