Machine Learning in the Real World

Introduction

Machine learning is already part of daily life. Knowing where it is used, and where it can go wrong, helps you use it wisely and ask good questions.

Where you meet it

• Recommendations: videos, music and products suggested to you
• Spam and fraud detection in email and banking
• Voice assistants and translation
• Photo tools that find faces or improve pictures
• Maps that predict travel time
• Medical research that helps spot patterns in scans, always checked by professionals

How bias gets in

A model learns from past data, and past data reflects human choices.

• If a hiring model learns from past decisions that were unfair, it can repeat the unfairness.
• If a face tool was trained mostly on some groups, it may work worse for others.
• A model can use a feature that stands in for something sensitive, such as a postcode that hints at background.

Privacy

Training data often contains personal information. Responsible teams collect only what they need, protect it, and follow privacy laws. As a user, check what an app asks to collect.

Keep humans in the loop

For decisions that affect people, such as health, jobs, loans or safety, a model should support a person, not replace their judgement. Good systems also explain what they do, can be challenged, and are monitored after launch because the world changes and models can become out of date.

Questions worth asking about any model

1. What data was it trained on, and who is missing from it?
2. How was it tested, and how often is it wrong?
3. What happens to people when it is wrong?
4. Can a human review or change the result?

Try it yourself

Pick an app that makes suggestions or decisions. Answer the four questions above as best you can, and note which ones you cannot answer.

Summary

• Machine learning powers recommendations, filters, maps and more.
• Bias in data can become bias in a model.
• Privacy, testing and human oversight make systems safer and fairer.