A Simple Classroom Activity for Teaching AI Bias and Fairness

Learning goal and materials

Students will explain how missing variety in training data can cause uneven performance and propose a data change that could improve the system.

Prepare paper cards or a local drawing tool, two safe labels such as stars and hearts, and a simple classifier approved by the school. Do not use faces, voices, names, handwriting identification, grades, or behavioural records.

Round 1: Create an intentionally narrow dataset

Make the star examples large, upright, and drawn with thick lines. Make the heart examples varied in size, angle, and line thickness. Train the classifier without telling students why the sets are different.

Ask students to predict which new drawings will be easy or difficult. Test large upright stars, then small, rotated, thin, or unusually shaped stars. Record every result.

Discuss what the model had a chance to learn

Ask which types of stars appeared in training and which appeared only during testing. The model did not decide to dislike small stars. The available examples made some patterns easier to learn than others.

Define dataset bias here as a systematic problem in the examples or collection process that can contribute to uneven outcomes. Connect this to people carefully: when AI affects people, missing groups or contexts can produce serious consequences.

• Which group of drawings had more variety?
• Which test cases failed most often?
• Would adding more copies of the same large star solve the problem?
• What new examples should be added, and why?

Round 2: Redesign and retest

Add small, rotated, thin, and differently shaped stars. Keep the original test set unchanged so students can compare results. Retrain, test, and count improvements as well as remaining errors.

Emphasize that balance is not only equal quantity. Twenty nearly identical examples may provide less useful variety than ten carefully different ones.

Move from accuracy to fairness

Accuracy asks how often the model is correct overall. Fairness asks how outcomes and errors are distributed and what those differences mean for affected people. A high overall score can still hide repeated failure in one context.

UNICEF’s child-centred AI guidance calls for inclusion, fairness, non-discrimination, transparency, and accountability. Students can practise those ideas by documenting the dataset, testing different cases, and refusing to use the model beyond its safe purpose.

Keep exploring

• Read the broader AI ethics guide
• Use the AI literacy classroom checklist
• Build a private-data-free first classifier