presentsUNAIVERSE
The UN AI timeline

    Learn AI /How Machines Learn /Supervised learning

    Supervised learning

    Detail

    In five seconds

    What it is

    A method for training a model on examples paired with target outputs, so it can predict outputs for new inputs.

    In plain words

    You show the system many examples with the right answers, and it adjusts itself so it can give likely answers for new examples.

    An everyday example

    Someone learning to do the laundry is shown garments paired with the correct pile: whites, dark colours, delicate wash or hand wash. When they put a wool jumper in the ordinary wash, they are corrected and adjust the rule they are using. After enough labelled examples, they can sort garments they have not seen before. Testing them on a fresh basket shows whether they learned useful patterns rather than memorising where yesterday’s clothes went.

    Where the story stops working

    This analogy can mislead because a student may consciously reason about mistakes and use general world knowledge. A supervised learning model usually does not have that kind of explicit reasoning. It changes numerical parameters based on many examples. Also, in the analogy the teacher provides a perfect answer sheet, but in real projects labels can be incomplete, noisy or disputed.

    Remember thisSupervised learning learns from examples with known answers, then uses those patterns to predict answers for new cases.

    Understand it

    Understand it

    Why it matters

    Many systems that people meet in daily life use supervised learning: spam filters, document sorting, speech recognition, fraud screening and medical image support tools. If you understand the basic method, you can ask better questions. What labels were used? Who created them? Are they accurate? Does the model work well on new cases, not just old ones? Could the labels reflect past bias or inconsistent human judgement? These questions matter because a supervised model can look reliable while still failing for some groups, situations or languages. Knowing the term helps you judge claims, buy software more carefully, and keep a human in the loop where mistakes could cause harm.

    How supervised learning turns labelled examples into predictions

    After viewing this diagram, the reader should be able to say that supervised learning uses examples with known correct outputs to adjust a model, then applies the trained model to new inputs.

    1. 01

      1. Labelled examples go in

      Each training example includes an input and its known correct output, such as a message plus the label “spam”.

      Draw three paired cards on the left. Each pair has an input box connected to a label tag. Use simple icons or short text, not decorative images.

    2. 02

      2. The model makes trial predictions

      At the start, the model may guess poorly because its internal settings have not been tuned yet.

      Place a model box in the centre-left. Show arrows from the input parts of the example cards into the model. Add faint sample outputs leaving the model.

    3. 03

      3. Compare with the known labels

      The system checks how far the trial predictions are from the correct outputs.

      Add a comparison box between the model outputs and the label tags. Show a mismatch indicator such as small red and green markers, with text ‘compare’.

    4. 04

      4. Adjust the model

      Training changes the model’s internal numerical settings so future predictions better match the labels.

      Draw a feedback arrow from the comparison box back into the model. Label it ‘update parameters’. Use one clear loop arrow only.

    5. 05

      5. Use the trained model on new input

      After training, a new input without a known label can be sent through the model to get a prediction.

      On the right, show a fresh input card entering a darker ‘trained model’ box, with one output card labelled ‘prediction’.

    Show a left-to-right flow with two phases. First, labelled examples enter training, where predictions are compared with known labels and the model is adjusted. Second, a new unlabelled input enters the trained model and produces a prediction. Read from left to right.

    Hover or tap each step to highlight only the relevant arrows and boxes; in reduced motion mode, use instant highlight changes with no animated transitions.

    Describe this without the diagram

    The diagram is a left-to-right flow in two phases. On the left are labelled examples, where each example has an input and a known correct output. These inputs go into a model, which produces trial predictions. A comparison step checks the trial predictions against the known labels. A feedback arrow sends the difference back to the model, showing that training updates the model’s internal parameters. On the right, a new unlabelled input goes into the trained model and comes out as a prediction. The main idea is that correct answers are available during training but not for the new case being predicted.

    With reduced motion: A static five-step flow chart with all stages visible at once: labelled examples, model, comparison, feedback update, and trained model making a prediction for a new input. Colour and line weight, not motion, show the two phases.

    Supervised learning is a common way to train an AI model from examples. Each example pairs an input with a target output, often called a label. A label is the training reference, but it may be wrong, noisy or disputed. For instance, an email could be the input, and “spam” or “not spam” could be the label.

    During training, the model produces trial outputs and compares them with the labels. A training method repeatedly adjusts internal numerical settings, called parameters, to reduce the measured error.

    This does not create a hand-written rule list. Instead, the fitted parameters define a mapping from inputs to outputs. If training goes well, that mapping also works on relevant examples the model did not encounter during training.

    The trained model can then predict a category, such as “approve” or “reject”, or a number, such as an estimated delivery time. Unsupervised learning finds structure without target outputs. Self-supervised learning creates targets from raw data as part of training. Reinforcement learning uses feedback and accumulated reward across a sequence of actions.

    How “Learning to sort the wash” maps onto the real thing

    In the storyIn the AI system
    Practice questions and answer sheetLabelled training examples with inputs and correct outputs
    The student checking mistakes after markingThe model comparing its predictions with labels during training
    Changing how similar questions are answered next timeAdjusting internal parameters to improve future predictions
    A new examination paperNew unseen data used at inference time

    A worked example

    A reporting tool is being trained to assign paragraphs from implementing partner updates to outputs in a results framework.

    Starting from: Past cleared paragraphs, each paired with the output assigned by a reporting officer. Separate records are reserved for testing and are not used during training.

    1. The model reads the paragraph as input and predicts one of the output labels used in the results framework.
    2. A loss function measures how far the prediction is from the label assigned by the reporting officer.
    3. The training process updates the model’s parameters to reduce that error across the labelled examples.
    4. The model is then evaluated on the reserved records, including results by working language and for less common labels, to check whether it generalises beyond familiar wording.
    5. For a new paragraph, the application presents the highest-scoring output label to the reporting officer for confirmation or correction.

    Result: The tool proposes a results-framework output rather than making the final reporting decision. Its usefulness depends on the quality and consistency of the past labels, its performance on unseen records and continued human review.

    In a United Nations workplace

    At a duty station where reports arrive in three working languages, a national programme officer is finalising the quarterly progress report for a donor. A proposed tool would assign paragraphs from implementing partner updates to the relevant results-framework output, based on paragraphs that reporting officers previously labelled during clearance.

    Supervised learning explains how the tool learns the link between each paragraph and its assigned output. The officer should ask how the labels were created, whether performance was tested separately in each working language, and whether rare outputs or unfamiliar wording produce more errors. The tool should support, not replace, reporting judgement.

    Before you try thisThe officer must not upload unredacted partner reports to a tool that has not been approved for that material. Personal, protection-related or security-sensitive details could be exposed or retained. Past labels may also reflect inconsistent reporting practice, so low-confidence and disputed classifications require human review.

    Where you may hear it

    • A vendor demo during an information technology governance board session
    • A procurement unit’s review of terms of reference for an AI-enabled service
    • Clearance of a concept note for automated document triage
    • A monitoring meeting reviewing model performance on new records

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • If a system uses supervised learning, it has been taught the correct rule for every case.

      Supervised learning usually does not store a complete rulebook. It fits patterns from labelled examples. If the examples are limited, noisy or unrepresentative, the model may still make mistakes on new cases.

    • More labelled data always guarantees a good model.

      Quantity helps, but quality and relevance also matter. Labels can be wrong, categories can be inconsistent, and the training data may not match the real setting where the model will be used.

    • Supervised learning means the model is supervised by a human every time it makes a prediction.

      The supervision mainly happens during training, because correct outputs are provided for examples. During later use, the model often predicts without a human giving the answer at that moment, though human oversight may still be needed.

    Quick check

    Which situation best matches supervised learning?

    Go deeper

    Go deeper

    The simple version

    You show the system many examples with the right answers, and it adjusts itself so it can give likely answers for new examples.

    The technical version

    Supervised learning fits a function from input features to target labels using labelled training data. During training, the model computes predictions, measures error with a loss function, and updates parameters to reduce that error across examples. The learned mapping can support classification, where the target is a category, or regression, where the target is a numerical value. Good practice checks performance on separate validation or test data to estimate generalisation rather than memorisation.

    Reviewed external resources

    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

    Supervised learning learns from examples with known answers, then uses those patterns to predict answers for new cases.