Good Data, Bad Data

By the end of this lesson you can:

• Tell good training data from bad.
• Name three data problems: too little, unbalanced, wrongly labelled.
• Explain how bad data leads to a biased, unfair AI.

Last lesson we designed clean categories. Now we check the examples inside them.

Here are photos labelled "cat": a tabby cat, a black cat, a dog, a ginger cat. Which label is wrong?

An AI learns only from its data. If the data is poor, the AI will be poor too. Computer scientists call this "garbage in, garbage out".

Three common data problems

• Too little — only a few examples, so the pattern is weak.
• Unbalanced — far more of one kind than another, so the AI favours it.
• Wrong labels — examples in the wrong box, teaching mistakes.

From bad data to bias

Bias is unfairness in an AI's answers. It often comes straight from unfair data.

If a face model is trained mostly on one group of people, it works worse for everyone else — not because anyone meant harm, but because the data was not balanced.

Arjun trains a "cat vs dog" model. Look at his data and predict the failures.

The amounts look balanced. But there's a sneaky problem.

• The model might learn a shortcut: "indoors = cat, grass = dog".
• Show it a dog indoors and it may say "cat".
• Show it a cat in a park and it may say "dog".

Plus one mislabelled photo (a dog tagged "cat") would quietly teach more mistakes.

Use your worksheet.

An AI is only as good as its data. Too little, unbalanced, or wrongly-labelled data makes a weak, biased AI — garbage in, garbage out. Checking data for balance and correct labels is part of building AI responsibly.