Model

Starter
AI Foundations
Policy officers

Also called: AI model, models.

Detail

In five seconds

What it is

A parameterised mathematical function learned from data that maps inputs to outputs such as labels, scores, predictions or generated text.

In plain words

A model is a trained pattern that takes in data and gives back an answer, score or generated content.

An everyday example

A city team receives thousands of paper comments from a public consultation. To help new staff sort them, an experienced organiser creates a guide with clues: words about buses often go in transport, comments about rent often go in housing, and comments about clinics often go in health. A new staff member uses that guide on fresh comments and puts each one into a tray. If a comment mentions several issues or uses unusual wording, the guide can still lead to the wrong tray. The organiser can revise the guide later after seeing mistakes.

Where the story stops working

A real AI model is not a short written guide made of clear rules. It usually consists of many numerical parameters learned from data, and its internal patterns may be hard for a person to inspect directly.

Remember thisA model is the trained mathematical function inside an AI system. It maps new inputs to outputs, but it is neither the whole system nor a guarantee that those outputs are correct.

This term is contested

The basic idea is widely shared, but the boundary of what counts as a 'model' can vary across statistics, machine learning, software practice and policy discussions.

Understand it

Why it matters

If you hear that an organisation is 'using AI', the practical question is often what model it is using, what inputs it takes, what outputs it gives, and how those outputs are checked. That affects accuracy, bias, privacy, procurement, accountability and fitness for purpose.

How a model turns input into output

After viewing this, a reader should be able to say that a model is a trained component that receives input data, applies learned patterns and returns an output.

01
Training examples
The process starts with data examples, often paired with the correct answer or target.
Draw several small cards on the left; each card contains short text and a coloured label.
02
Adjustment
Training changes the model's internal numerical settings so it matches patterns in the examples more closely.
Place a box with small circles or sliders moving toward new positions; label it 'adjust parameters'.
03
Trained model
After training, the model stores learned patterns in its parameters.
Show a larger central rounded rectangle labelled 'trained model' with many small dots inside.
04
Use on new input
A new item goes into the model, which produces an output such as a label, score or draft text.
Draw a new comment card entering the model from below or left, with an arrow leading to a right-side output badge reading 'housing'.

Show a left-to-right flow with four boxes connected by arrows. Box 1: 'Examples for training' with small icons for comments and labels. Box 2: 'Training adjusts numerical settings' with sliders or dots shifting into place. Box 3: 'Trained model' shown as a central block labelled 'learned pattern'. Box 4: 'New input' feeding into the model and then to 'Output' such as 'housing'. Use simple labels above arrows: 'learn from examples' and 'apply to new case'.

Hovering or tapping each step highlights the relevant box and arrow, and reveals one sentence of explanation beneath the diagram.

Describe this without the diagram

The diagram shows a four-stage flow. First, there are training examples, such as public comments paired with correct labels. Second, a training stage adjusts many numerical settings so the system fits patterns in those examples. Third, the result is a trained model, which stores those learned patterns. Fourth, a new input goes into the trained model, and the model produces an output such as a category label, a score, a prediction or generated text. The main point is that the model is the trained pattern-matching part between input and output, not the whole software system.

With reduced motion: A static left-to-right diagram with four labelled boxes and arrows: training examples, adjustment during training, trained model, and new input leading to output. No animated movement; the same explanatory text appears below the selected step.

In AI, a model is the part that takes something in and produces something out. The input might be text, an image, a sound clip or a row of numbers. The output might be a category, a probability, a forecast or a new piece of content.

A model does not appear on its own. It is usually built by training, which means adjusting many numerical settings using examples from a dataset. A dataset is an organised collection of data. After training, those settings capture patterns found in the examples.

When you use the model, it applies those learned patterns to new input. For example, it might read a public comment and estimate whether it is about transport, housing or health. In a generative system, it might produce draft text based on patterns in its training and the prompt it receives.

People often confuse a model with an algorithm or with software. An algorithm is a method or procedure. Software is the whole application or system people use. A model is the trained pattern-matching component inside such a system.

The word can mean slightly different things in statistics, machine learning and everyday discussion. The core idea is similar: a simplified mathematical representation that maps inputs to outputs. But the details vary, so it is worth asking what kind of model someone means.

How “A sorting guide for consultation comments” maps onto the real thing

In the story	In the AI system
the organiser's sorting guide	the trained model
paper comments	input data
labelled trays such as transport, housing and health	output categories
revising the guide after mistakes	updating or retraining the model

A worked example

In a country office, an information management officer tests a text-classification model that tags non-sensitive implementing-partner progress reports against workplan themes before a quarterly review.

Starting from: Report excerpt: “The project repaired water points and trained local maintenance committees.”

During training, the model receives earlier report excerpts with workplan labels assigned by programme officers.
An optimisation method adjusts the model’s parameters to reduce the difference between its predicted labels and the staff-assigned labels.
During inference, the new excerpt is converted into a numerical representation that the fitted model can process.
The model assigns a probability to each label. Basic services receives the highest probability, so the tool proposes that tag.

Result: The progress report is provisionally tagged under basic services. This is a prediction based on learned patterns, not a determination that the report supports only that theme. The information management officer reviews the proposed tag before it is added to the file.

In a United Nations workplace

At an inter-agency secretariat, an evaluation officer uses an approved internal drafting tool to turn non-sensitive findings into three paragraphs for a statement for delivery. The draft still has to reach the front office for clearance by close of business.

The model maps the officer’s prompt and source text to a sequence of tokens using parameters learned during training. It can produce fluent wording quickly, but it does not verify the findings or know which qualifications will matter in clearance. Every claim must be checked against the evaluation report.

Before you try thisPersonal data, confidential evidence and unpublished findings must not be entered into a tool that has not been approved for that material, because doing so could expose protected information. Approval of the tool does not clear its output. The officer must check for invented claims, omitted caveats and changes in meaning before circulating the draft.

Where you may hear it

A vendor demo in a procurement unit before an internal tool is selected
An inter-agency working group reviewing an automated triage tool before a pilot
An ICT governance board considering whether a model can connect to an internal system
A country office briefing before a chatbot pilot opens to staff

Avoid confusion

Often confused with

Common misconceptions

Common misconception: A model is the same thing as the whole app or chatbot.
Correction: The app includes interfaces, databases, safety controls and other software. The model is only the trained component that maps inputs to outputs.
Common misconception: A model contains facts in a reliable way, like a database.
Correction: A model stores patterns in numerical parameters, not a checked list of facts. That is why it can produce useful output but also mistakes or invented details.
Common misconception: Once trained, a model keeps improving just by being used.
Correction: Some systems may be updated, but a model does not automatically become better from ordinary use unless people collect data and retrain or tune it.
Common misconception: A bigger model is always better.
Correction: Larger models can do more, but they may cost more, use more resources, be harder to govern and still perform poorly on a specific local task if the data or setup is weak.

Quick check

Which description best matches a model in AI?

Go deeper

The simple version

A model is a trained pattern that takes in data and gives back an answer, score or generated content.

The technical version

A model is a parameterised mathematical function learned from data. During training, optimisation methods adjust parameters to reduce error on examples; during inference, the fitted function maps new inputs to outputs such as class probabilities, numeric predictions or token sequences.

Reviewed external resources

Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This link gives a policy and ethics framework for AI and suits policymakers and public-sector readers who need governance context rather than technical detail.
Intermediatelink checked 2026-07-31
EncyclopaediaStatistical model (opens in a new tab)Wikipedia
This link gives a broad background on statistical models and suits readers who want general context beyond AI-specific usage.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This link gives short glossary-style definitions of machine learning terms and suits beginners who want a quick technical orientation.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Model (opens in a new tab)YouTube search
This is a YouTube search results page, so it offers a range of video explanations from different creators and suits readers who prefer to compare teaching styles.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Model (opens in a new tab)Reddit search
This is a Reddit search results page, so it shows community questions and debates about the term and suits readers who want to see common confusions and informal explanations.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

A model is the trained mathematical function inside an AI system. It maps new inputs to outputs, but it is neither the whole system nor a guarantee that those outputs are correct.