Fair and safe AI: be a responsible AI maker

Introduction

You've built AI that finds faces, reads expressions, tracks hands, spots objects and even recognises people. That's real power, and real AI makers have to ask two big questions:

Is it fair? Does it work equally well for everyone?
Is it safe? Does it respect people's privacy?

In this lesson you'll become an AI tester. You'll check your own projects for fairness and make a set of rules for using cameras responsibly.

Unfair: like an umpire who only ever watched one team, a model trained on mostly one kind of face makes far more mistakes with everyone else. That's not the people's fault; it's the training data's.

Working principle

Bias is when an AI works better for some people or situations than others, usually because of what was (or wasn't) in its training samples.
You found this yourself: your rock, paper, scissors model worked best on the hands you trained it with.
Testing fairly means trying a model with many different people and conditions and writing down where it fails.
Privacy means people get to decide who sees and keeps pictures of them. A face is personal data.
Online AI blocks (like text recognition) send pictures over the internet. Offline ones keep them on your computer.

What you'll need

One of your earlier projects: the emotion mirror, the crown or your rock, paper, scissors game
Two or three volunteers (family or friends) who agree to help
A notebook and pencil for your test results

Working in real life

Part 1: The fairness test. Pick one project and test it in different conditions. Make a table like this in your notebook and tick each time it works (try each one 5 times):

Test	Works (out of 5)
Bright room
Dim room
With glasses
Different person 1
Different person 2
Far from the camera

Where did it do worst? Why do you think that happened? How could you make it fairer? (For your own model: add more varied samples and retrain!)

Part 2: Make the AI maker's promise. Discuss these questions with your class or family, then write your own rules:

Should a camera project film people who don't know about it?
Who should be allowed to see the faces you saved in your doorbell project?
What should you do with photos when your project is finished?
Is it OK for an AI to decide who is allowed into a room, if it sometimes makes mistakes?

Try it

Write your AI maker's promise on a card, with at least four rules. Here are some ideas to start you off:

I always ask before filming or saving someone's face.
I delete faces and photos when I don't need them any more.
I test my AI with lots of different people before I trust it.
I tell people when a camera project is switched on.
I never use AI to make fun of someone.

Check what you learnt

What is bias in AI?
How can you make your own trained model fairer?
Your friend says no to being in your face recognition project. What do you do?

Conclusion

AI can be unfair when its training data doesn't include everyone. Testing with lots of people helps you find out.
Faces and photos are personal data: ask first, keep them safe, delete them when you're done.
Good AI makers think about people, not just code.

What next?

You're ready for the final challenge: Final project: build your own AI helper.