- Define and explain Training Data in your own words
- Use key terms such as training data accurately
- Apply what you have learned to new examples and questions
- Avoid the common mistakes learners make with this topic
This lesson focuses on Training Data: understand why the quality and quantity of data decides what AI learns.
Understand why the quality and quantity of data decides what AI learns.
Key ideas
Learning replaces hand-written rules
Nobody could write rules to recognise every cat photo, so instead a model is shown thousands of labelled examples and adjusts its internal numbers until it predicts well. This is training; the model's skill is then tested on new examples it has never seen. If it only memorises the training set — overfitting — it fails on anything new.
Biased data in, biased decisions out
A hiring tool trained on a decade of a company's past hires will learn to prefer the kind of people the company used to hire — baking old discrimination into new software. Because the bias hides inside millions of tuned numbers, it is hard to spot without deliberately testing for fairness across groups.
Key term — training data: The examples a model learns from — thousands of labelled photos, texts or records. Its quality and balance shape everything the model does.
A team trains a cat photo classifier on 10,000 photos — but 9,500 show white cats. It scores 98% in testing, then fails on users' ginger and black cats. Explain why.
The training data is unbalanced: the model sees white cats 19 times more often than all other colours combined. The model learns the easy pattern 'white and fluffy means cat' instead of learning true cat features like ears and whiskers. Testing used photos from the same collection, so the test set was unbalanced too — the 98% score measured skill on white cats, not cats in general. On real users' varied photos, the shortcut fails: ginger and black cats do not match the learned pattern.
Answer: The classifier learned colour, not 'cat-ness', because unbalanced training and testing data rewarded the shortcut. Fix: train and test on diverse, representative photos.
- Assuming AI is neutral because maths is neutral Correction: models inherit the biases of their training data — neutrality must be tested for, never assumed.
- Trusting a high accuracy score without asking what it was tested on Correction: always ask whether the test data was balanced and representative — 98% on easy data can hide total failure elsewhere.
Practice
In traditional programming humans write explicit rules; in machine learning the program finds its own patterns by training on data.
Its training data contains no night-time patterns, so it never learned the features of dark images — it can only recognise what it has seen.
Overfitting is when a model memorises training examples instead of learning general patterns; it shows as high training accuracy but poor accuracy on unseen test data.
False — remaining details like postcode or school can act as proxies for protected characteristics, so bias can survive.
Quick check
Which of these best defines "training data"?
You use an AI chatbot to draft your history essay. What is the responsible way to handle this?
A university uses AI to screen applications, trained on ten years of past admissions. Name one fairness risk.
- Training Data: understand why the quality and quantity of data decides what AI learns.
- Learning replaces hand-written rules: Nobody could write rules to recognise every cat photo, so instead a model is shown thousands of labelled examples and adjusts its internal numbers until it predicts well.
- bias: A systematic unfairness in a model's decisions, usually inherited from unrepresentative or prejudiced training data.
- Watch out for: assuming AI is neutral because maths is neutral