Train an AI (and watch it get things wrong)

Suggested age: 11-16Length: 15 minutes

A real, tiny AI running inside this page. You give it fruit examples and it learns to tell them apart. Give it skewed examples and it learns skewed.

Try this: load only heavy apples and light oranges. Then test a light apple and watch it get it wrong.

1. Give it examples

Each fruit has two measurements: how much it weighs and how red it is. Pick examples and tell the AI what they are.

2. Train it

The AI looks for a line that separates the two groups. It does not know what an apple is: it only looks for where to cut.

3. Test it

Now show it a fruit it has never seen and watch what it answers.

Teacher sheet

Goal. See with your own eyes that an AI knows nothing about fruit: it only repeats the pattern of the examples it was given.

Questions to open the discussion

If we only give it big apples, what happens to a small apple?
Does the AI "know" what an apple is?
Where else could this happen, with people instead of fruit?

The idea that should stick. The AI did not learn what an apple is: it learned which examples it was shown. If the examples are biased, the result comes out biased, and the machine has no way of noticing.

Chris Meniw Foundation Inc. · Free to use in classrooms. No permission or attribution needed.