AI Bias and Fairness

# CHAPTER 17

AI Bias and Fairness #

1. Introduction #

2. Learning Objectives #

• Explain why AI algorithms are not inherently objective.

• Identify the sources of bias in training datasets.

• Recognize real-world examples of algorithmic discrimination.

• Understand techniques used to reduce bias in AI models.

3. Beginner-Friendly Explanation #

4. Real-World Examples #

• Amazon's Recruiting AI (2018): Amazon built an AI to review resumes. It trained on 10 years of Amazon hiring data. Because the tech industry was male-dominated, the AI learned to penalize any resume that contained the word "women's" (e.g., "Women's Chess Club Captain"). Amazon had to scrap the project entirely.

• Facial Recognition: Several commercial facial recognition systems were trained primarily on images of lighter-skinned individuals. When deployed, these systems had terrible accuracy rates for darker-skinned individuals, leading to false arrests.

5. Sources of Bias #

1. 1. Historical Bias: The data perfectly captures the world as it was, but the world was flawed or discriminatory.

1. 2. Representation Bias: The training data does not accurately represent the population. (e.g., Training a global self-driving car only using pictures of roads in sunny California. It will fail in the snow).

1. 3. Measurement Bias: The way the data was collected was flawed. (e.g., Relying on smartphone GPS data to measure crowd sizes, completely ignoring low-income populations who may not own smartphones).

6. The Danger of "Proxy" Variables #

7. Mitigating Bias (How to Fix It) #

• Curating Balanced Datasets: Actively over-sampling underrepresented groups. If your dataset is 80% men and 20% women, collect more data on women until it is 50/50 *before* training.

• Algorithmic Fairness Constraints: Modifying the loss function. You mathematically penalize the AI during training if it generates a different False Positive rate for one demographic compared to another.

• Continuous Auditing: Constantly testing the deployed model with diverse data to ensure its accuracy remains uniform across all groups.

8. Step-by-Step: Bias Audit #

1. 1. Train the model.

1. 2. Test the model's accuracy on the general validation dataset.

1. 3. Slice the validation dataset into demographics (e.g., Men, Women, Minority A, Minority B).

1. 4. Calculate the accuracy for *each* slice individually.

1. 5. If the accuracy for Minority B is 20% lower than the general accuracy, the model fails the audit and must be retrained with better data.

9. Mini Project #

1. 1. Miles driven per year

1. 2. Make and Model of the car (e.g., Minivan vs Sports Car)

1. 3. Number of speeding tickets

10. Best Practices #

• Diverse Teams: The best way to prevent Representation Bias is to ensure the team of data scientists building the AI is diverse. Diverse teams are more likely to notice when a dataset excludes certain populations.

11. Common Mistakes #

• Blind Trust in Math: Never assume an algorithm is fair just because it is a computer. A computer is a mirror of its creator's data.

12. Exercises #

1. 1. Explain why simply deleting the "Race" column from a dataset does not guarantee that an AI loan-approval system will be racially fair.

13. Coding Challenges #

Function audit_model_fairness(model, validation_data)
    general_accuracy = test_accuracy(model, validation_data)
    
    // Filter data for underrepresented group
    minority_data = filter(validation_data, group == "Minority_X")
    minority_accuracy = test_accuracy(model, minority_data)
    
    If (general_accuracy - minority_accuracy) > 0.05:
        Print "FAIL: Model exhibits severe bias against Minority_X!"
        Return False
    Else:
        Print "PASS: Model performs uniformly."
        Return True

14. MCQs with Answers #

15. Interview Questions #

• Q: How would you address Representation Bias in a dataset before training a model?

• Q: Explain why algorithms are not inherently objective or neutral.

16. FAQs #

17. Summary #

18. Next Chapter Recommendation #