Machine learning is a branch of artificial intelligence where systems learn patterns from data, rather than being explicitly programmed with fixed rules for every scenario. In supervised learning, a model is trained on labelled examples (input paired with a known correct output) to learn patterns it can then apply to new, unseen data — this underlies many everyday applications, from email spam filters to recommendation systems, though the quality and biases in training data directly shape how well and how fairly the resulting model performs.
Example
A spam filter trained on thousands of emails labelled "spam" or "not spam" learns to recognise patterns — like certain phrases or sender behaviours — associated with spam, allowing it to make reasonably accurate predictions about new, previously unseen emails.
Key terms
Machine learning:
A branch of AI where systems learn patterns from data rather than fixed rules.
Supervised learning:
Training a model using labelled examples of input paired with known correct outputs.
Training data:
The data set used to teach a machine learning model.
Questions
1. Machine learning is a branch of:
Artificial intelligence
A field entirely unrelated to computing
Only traditional, hard-coded programming
A term with no connection to data
2. Machine learning systems learn patterns from:
Data
Nothing at all; they require no data
Only fixed, hard-coded rules
A completely random process with no data involved
3. Supervised learning trains a model using:
Labelled examples with known correct outputs
Completely unlabelled data with no known outputs
No data of any kind
Only images, with no other data type
4. Training data is:
The data set used to teach a machine learning model
A term unrelated to machine learning
Only used after a model has already been fully built
Something that has no effect on a model's performance
5. A spam filter is an example of:
A machine learning application
Something entirely unrelated to machine learning
A purely manual, non-automated process
A system with no connection to data or patterns
6. Machine learning differs from traditional programming because it:
Learns patterns rather than following fixed, explicit rules for every scenario
Always requires fixed rules for every single possible scenario
Has no connection to data of any kind
Is identical to traditional programming with no meaningful difference
7. Recommendation systems are an example of:
A common machine learning application
Something with no connection to machine learning
A purely random, unlearned process
A system that never uses any data
8. Why might a spam filter trained on thousands of labelled emails be able to identify spam patterns a simple, fixed rule couldn't catch?
Learning from many examples can reveal subtle or evolving patterns that would be difficult to anticipate and hard-code manually
A fixed set of rules always catches spam more accurately than any learned pattern could
Machine learning models never actually identify any patterns beyond what a fixed rule already covers
Labelled training data provides no additional benefit over simply writing fixed rules
9. Why is the quality of training data considered critical to how well a machine learning model performs?
A model learns directly from the patterns present in its training data, so flawed or limited data can lead to flawed predictions
Training data quality has no bearing on how a machine learning model ultimately performs
Machine learning models always perform equally well regardless of the quality of their training data
Models never actually learn anything from the data they are trained on
10. Why might a machine learning model trained mostly on one type of data produce biased or inaccurate results for underrepresented groups?
The model may not have learned sufficient patterns relevant to groups that were underrepresented in its training data
Training data composition never has any effect on how a model performs for different groups
Machine learning models always perform identically well for every group regardless of training data composition
Underrepresentation in training data has no connection to potential model bias
11. Why is "labelled" data (input paired with a known correct output) specifically required for supervised learning?
The model needs known correct answers during training to learn which patterns are associated with which outcomes
Supervised learning never actually requires any labelled data of any kind
Labels have no relevance to how a supervised learning model is trained
Unlabelled data is always exactly as useful as labelled data for supervised learning
12. Why might testing a machine learning model on new, unseen data (rather than only the data it was trained on) be an important step?
It reveals whether the model has genuinely learned generalisable patterns, rather than just memorising the specific training examples
Testing on new data provides no useful information about how well a model has actually learned
A model always performs identically on its training data and on completely new data
Testing on unseen data is never necessary once a model has been trained
13. Why might it be misleading to describe a machine learning model as making decisions with the same kind of understanding a human would?
The model identifies statistical patterns in data, which is different from genuine comprehension or reasoning about meaning
Machine learning models always have exactly the same kind of understanding and reasoning ability as a human
There is no meaningful difference between statistical pattern recognition and human understanding
This distinction has no relevance to how machine learning systems are described or evaluated
14. Why might biased training data lead to real-world consequences when a machine learning model is deployed in a sensitive application, like loan approval decisions?
Patterns learned from biased historical data could result in the model unfairly disadvantaging certain groups in its predictions
Biased training data never has any real-world impact once a model has been deployed
Sensitive applications like loan approvals are never affected by the quality of a model's training data
Machine learning models used in sensitive applications are always guaranteed to be completely fair
15. Why might machine learning models require ongoing monitoring and retraining after deployment, rather than being considered "finished" once trained?
Real-world patterns can change over time, meaning a model's performance may degrade if it isn't updated with new, relevant data
A trained machine learning model always continues performing perfectly forever with no further attention needed
Retraining a deployed model is never actually necessary or beneficial in any circumstance
Real-world patterns and data never change in ways that would affect a model's ongoing performance
16. Why might understanding the basics of machine learning be increasingly important for informed citizenship, not just for computer scientists?
These systems increasingly influence decisions in areas like finance, hiring and media, making basic understanding relevant to evaluating their impact
Machine learning has no relevance whatsoever to daily life outside of specialised computer science contexts
Only computer scientists ever need to understand anything about how machine learning systems work
Machine learning systems never actually influence any real-world decisions relevant to ordinary people
17. A photo-tagging app that learns to recognise faces from many labelled example photos is using:
Supervised learning
A process with no connection to machine learning
Only fixed, hard-coded rules with no learning involved
A method unrelated to labelled training data
18. Voice assistants that improve their speech recognition accuracy over time as they process more audio data are an example of:
A machine learning application
A system with no connection to learning from data
A purely rule-based system with fixed, unchanging logic
Something unrelated to artificial intelligence
19. Why might a self-driving car's machine learning system need to be trained on an enormous variety of driving scenarios, rather than just a few common ones?
Limited training scenarios could leave the system unprepared for unusual or rare real-world situations it might encounter
A small number of common scenarios is always sufficient to prepare a system for every possible real-world situation
The variety of scenarios in training data has no bearing on how well a self-driving system performs
Self-driving systems never actually need any training data related to driving scenarios
20. Why might explaining why a machine learning model made a particular decision (its "interpretability") be especially important in high-stakes fields like medical diagnosis?
Understanding the reasoning behind a decision helps professionals evaluate whether to trust and act on the model's output
Interpretability has no relevance in any high-stakes field, including medical diagnosis
Medical professionals never need to understand or evaluate how a model reached its output
Machine learning models always provide complete, automatic explanations for every decision they make
21. Why might overfitting — where a model performs very well on training data but poorly on new data — be considered a significant problem in machine learning?
It indicates the model has essentially memorised specific examples rather than learning patterns that generalise to new situations
Overfitting always means a model will perform equally well on both training data and completely new data
Performing well on training data always guarantees strong performance on new, unseen data as well
Overfitting is not actually considered a meaningful concern in machine learning practice
Answer key (parent copy)
1. Artificial intelligence
2. Data
3. Labelled examples with known correct outputs
4. The data set used to teach a machine learning model
5. A machine learning application
6. Learns patterns rather than following fixed, explicit rules for every scenario
7. A common machine learning application
8. Learning from many examples can reveal subtle or evolving patterns that would be difficult to anticipate and hard-code manually
9. A model learns directly from the patterns present in its training data, so flawed or limited data can lead to flawed predictions
10. The model may not have learned sufficient patterns relevant to groups that were underrepresented in its training data
11. The model needs known correct answers during training to learn which patterns are associated with which outcomes
12. It reveals whether the model has genuinely learned generalisable patterns, rather than just memorising the specific training examples
13. The model identifies statistical patterns in data, which is different from genuine comprehension or reasoning about meaning
14. Patterns learned from biased historical data could result in the model unfairly disadvantaging certain groups in its predictions
15. Real-world patterns can change over time, meaning a model's performance may degrade if it isn't updated with new, relevant data
16. These systems increasingly influence decisions in areas like finance, hiring and media, making basic understanding relevant to evaluating their impact
17. Supervised learning
18. A machine learning application
19. Limited training scenarios could leave the system unprepared for unusual or rare real-world situations it might encounter
20. Understanding the reasoning behind a decision helps professionals evaluate whether to trust and act on the model's output
21. It indicates the model has essentially memorised specific examples rather than learning patterns that generalise to new situations