Learn AI /AI Foundations /Machine learning

Machine learningML

Starter
AI Foundations
Policy officers
Technical readers

Also called: ML.

Detail

In five seconds

What it is

A way of building computer systems that learn patterns from examples so they can make predictions, classifications or recommendations on new cases.

In plain words

Machine learning lets a computer improve at a task by finding patterns in examples instead of following only fixed hand-written rules.

An everyday example

A teacher has many pairs of short paragraphs from students in Nairobi, Lima and Bangkok. Some pairs express the same idea in different words, and some do not. Instead of writing one rule for every phrase, the teacher studies many marked examples and starts noticing patterns such as shared meaning, similar sentence structure and matching key terms. Later, the teacher can make a quick judgement on a new pair of paragraphs that was not in the practice set. This works as an intuition, not a perfect match. A machine learning system does not notice meaning in the same rich way a teacher does, and it depends heavily on how the text is represented in data and how the model was trained.

Where the story stops working

The teacher can use broad human context, world knowledge and intent. A machine learning system usually works through mathematical patterns in the data representation, so it may miss obvious meaning for people or rely on misleading cues.

Remember thisMachine learning fits a model to examples, then uses that model to predict, classify or recommend for new inputs. Its outputs remain estimates shaped by the data, objective and testing.

This term is contested

The basic idea is widely agreed, but the boundary between machine learning, broader AI, statistics and pattern recognition is not always drawn the same way by all communities.

Understand it

Why it matters

Non-specialists often need to approve, buy, govern or use systems that claim to use AI. If you understand machine learning, you can ask better questions about data quality, error rates, fairness, oversight, security and whether the tool actually fits the task.

How machine learning turns examples into a usable model

After viewing this, a reader should be able to say that machine learning uses example data to train a model, then applies that trained model to new inputs.

01
Examples in
Collect past examples relevant to the task, sometimes with correct answers attached as labels.
Draw 6 to 8 small document cards entering the first box; some cards include visible tags such as 'spam' and 'not spam'.
02
Training
The system adjusts the model so its outputs fit the examples more closely.
Show a central box with three to five slider controls or dials changing position, with a caption 'adjust parameters'.
03
Model ready
The result is a trained model: a reusable pattern-matcher shaped by the training data.
Draw a sealed container or chip-like rectangle labelled 'trained model', visually distinct from the training box.
04
New case out
When new data arrives, the trained model produces a prediction, classification, recommendation or score.
Show two new document cards entering from the left of the final stage and three output pills on the right labelled 'class', 'score', and 'ranking'.

A left-to-right flow diagram with four boxes connected by arrows. Box 1 shows a stack of example records labelled with simple tags. Box 2 shows a training stage with sliders being adjusted. Box 3 shows a boxed model icon labelled 'trained model'. Box 4 shows new incoming records producing outputs such as 'class', 'score' or 'ranking'. Use a consistent colour for data, a second colour for the training stage, and a third colour for outputs. Add a small note under the final box: 'Check performance on new data; outputs can still be wrong.'

On hover or tap, each step reveals one short sentence and highlights the arrow to the next stage. A toggle switches between 'labelled examples' and 'unlabelled examples' to show that some machine learning uses labels and some does not, without changing the overall flow.

Describe this without the diagram

The diagram shows four stages in order. First, example data goes in, sometimes with correct labels attached. Second, a training process adjusts a model to fit patterns in those examples. Third, the output of training is a trained model. Fourth, new data goes into that trained model and comes out as an output such as a class, score or ranking. A note reminds the reader that the output still needs testing and human review because trained models can make mistakes.

With reduced motion: A static four-box flowchart appears with all labels visible at once: examples in, training, trained model, and new case out. The optional labelled-versus-unlabelled toggle changes the icon labels without animation.

In machine learning, the main input is data: recorded examples such as text, images, numbers or clicks. Each example may come with a label, which is a target answer used during training, such as “spam” or “not spam”.

During training, a model is adjusted so its outputs match the examples more closely. A model is a mathematical structure with many settings, often called parameters, that can be tuned. The system does not store a hand-written rule for every case. Instead, it looks for patterns that often appear together.

After training, the model is used on new data it has not seen before. It produces an output such as a category, a score, a ranking or a likely next word. The output can be useful, but it can also be wrong, biased or overconfident, so people still need to test, monitor and review it.

Machine learning is not the same as all artificial intelligence, and it is not the same as ordinary software with fixed rules. It is one broad approach within AI. Deep learning is a more specific family inside machine learning, not a synonym for all of it.

How “Sorting paragraphs that say the same thing in different words” maps onto the real thing

In the story	In the AI system
Marked paragraph pairs used for practice	Training data with examples and correct labels
The teacher noticing repeated patterns	The model adjusting to statistical patterns in data
A new pair of paragraphs	New input used after training
The quick judgement about whether they match	The model output, such as a classification or score

A worked example

An information management officer at a regional hub is testing an approved tool that assigns routine facilities requests to categories before they appear on the service desk worklist.

Starting from: De-identified past requests, each labelled with the category previously assigned by service desk staff.

The text is converted into a representation the model can process, while names, contact details and other unnecessary identifiers are excluded.
Some labelled requests are used to fit the model. Others are held back to test how it performs on examples it did not learn from.
During training, the model adjusts its parameters to reduce differences between its predicted categories and the staff-assigned labels.
A new request says that a meeting room remains too warm even after the controls were adjusted. The fitted model assigns scores to the available categories.
The highest-scoring category is added to the worklist as a suggested route. A service desk colleague reviews low-confidence or unusual cases.

Result: The tool has learned statistical patterns connecting request text with earlier labels. It predicts a useful category for a new request, but it does not understand the building or prove that the suggested route is correct.

In a United Nations workplace

An ICT officer in a country office is fact-checking the technical paragraph of talking points that the regional bureau needs for informals. The file, on a proposed machine learning service for routing public enquiries, is due back by close of business.

The officer can distinguish a system trained on examples from one following fixed rules, then flag the questions the talking points should not glide past: which examples were used, what objective was set, how performance was tested on new enquiries, which languages were covered, and when a person reviews the output.

Before you try thisThe officer should not paste sample enquiries containing names, contact details or case information into an external tool that has not been approved for those data. Public-facing inboxes can still contain personal or protection-sensitive material, and a demonstration is not a basis for disclosure.

Where you may hear it

A vendor demo before an ICT governance board
A first reading of terms of reference in a procurement unit
Clearance of a briefing note on a proposed digital service
An inter-agency working group reviewing a data pilot

Avoid confusion

Often confused with

Side-by-side

AI vs machine learning vs deep learning
Are these three words for the same thing?

Common misconceptions

Common misconception: Machine learning means the computer writes its own goals and works on anything.
Correction: Machine learning systems are usually built for specific tasks and optimise for objectives chosen by people, such as predicting a label or ranking results.
Common misconception: If a system uses machine learning, it must be more accurate than rule-based software.
Correction: Not always. For stable tasks with clear rules, a rule-based system can be simpler, safer and easier to audit. Machine learning is useful when patterns are too complex to write out directly.
Common misconception: A trained model simply memorises all the training data.
Correction: Some models can memorise parts of data, which is a real risk, but the main aim is to capture patterns that generalise to new cases rather than store every example exactly.
Common misconception: Machine learning and deep learning are the same thing.
Correction: Deep learning is one subset of machine learning. Many machine learning methods do not use deep neural networks.

Quick check

Which description best matches machine learning?

Go deeper

The simple version

Machine learning lets a computer improve at a task by finding patterns in examples instead of following only fixed hand-written rules.

The technical version

Machine learning uses data to fit or construct a model for an objective, then applies the resulting function or procedure to produce outputs for new inputs.

Reviewed external resources

Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This link gives an international ethics and governance framework for AI use, and it suits policy, oversight and procurement audiences.
Intermediatelink checked 2026-07-31
EncyclopaediaMachine learning (opens in a new tab)Wikipedia
This link gives a broad overview of the term, its history and common types, and it suits readers who want a general starting point.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This link gives short glossary-style definitions of related machine learning terms, and it suits readers who want quick clarification of vocabulary.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Machine learning (opens in a new tab)YouTube search
This link is a YouTube search results page for video explanations of machine learning, and it suits readers who prefer to compare several visual introductions rather than read one article.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Machine learning (opens in a new tab)Reddit search
This link is a Reddit search results page for community discussions about machine learning, and it suits readers who want to see common questions, debates and informal explanations with caution about reliability.
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

Machine learning fits a model to examples, then uses that model to predict, classify or recommend for new inputs. Its outputs remain estimates shaped by the data, objective and testing.