Ethics and AI Bias Awareness

AI learns from its training data, so one-sided data gives unfair results. We check accuracy for each group and use AI responsibly.

Key ideas

Bias happens when training data has far more examples of some groups than others.
A model usually makes the most mistakes on the group with the fewest training examples.
Test a model separately on each group: one overall accuracy can hide a group it treats badly.
A face-unlock model trained mostly on daytime photos may fail more often at night.
Informed consent means people know what data is collected and agree before it is used.
Responsible use: check AI answers, say when AI helped, and report unfair results.

Remember

Share of a group	group count ÷ total × 100%
Accuracy for a group	correct in the group ÷ tested in the group × 100%
Gap between groups	difference of the two percentages, in percentage points
Example	quiet rooms 17 of 25 = 68%, noisy rooms 13 of 25 = 52% → 16 points

Watch out

Comparing the numbers right instead of the percentages when the groups are different sizes.
Testing a model only on the examples it was trained on.
Thinking more data always fixes bias: it must include every group fairly.

Worked example

A model was tested on 25 recordings from quiet rooms (17 right) and 25 recordings from noisy rooms (13 right). How many percentage points apart are its accuracies on the two groups?

Working: recordings from quiet rooms: 17/25 = 68%; recordings from noisy rooms: 13/25 = 52%

Answer: 16