lesson
30 min

Bias, Accuracy, and Limits

How unfair data becomes unfair predictions, and when a prediction should never decide a person's future.

What you will be able to do

By the end of this lesson, you can describe how biased data leads to unfair results and decide when a prediction should not drive a decision.

Bias in, bias out

Bias in machine learning means a model treats some groups worse than others in a way that is not fair. It usually comes from the training data, not from a bad intention.

If the past was unfair, data from the past carries that unfairness. A model trained on it learns the unfair pattern and repeats it, fast and at scale.

"Accurate" can still hurt people

A company may say its model is "95 percent accurate." Ask two questions:

  1. Accurate for whom? A model can do well on average and badly for one group.
  2. What happens to the people it gets wrong? A wrong movie suggestion is fine. A wrong fraud flag that freezes someone's rent money is not.

When a prediction should not decide

Use machine learning to sort, suggest, or flag. Let a trained person decide when the result affects:

  • Jobs, school admission, or discipline
  • Housing, loans, or public benefits
  • Policing, courts, or immigration
  • Health care

People affected by an automated decision should be able to ask why and to appeal to a person. If no one can explain the decision, that is a warning sign.

Questions to ask about any prediction tool

0 of 5 done

Module 3 check

Quiz: Module 3: Machine Learning

  1. 01What is training data?

    Choose one.

  2. 02In the Python example, what happened when you added a new training example?

    Choose one.

  3. 03Why did Amazon's experimental hiring tool downgrade women?

    Choose one.

  4. 04A model is "95 percent accurate." What should you ask? Choose all that apply.

    Choose all that apply.

  5. 05Which use of machine learning needs a trained person to make the final decision?

    Choose one.

Ask. Answer. Build together.