Module 5 Lesson 2: Bias and Fairness in AI

AI bias is not necessarily about "hateful intent" by the developer. More often, it is a mathematical reflection of historical inequality captured in data. In business, an "Unbiased" model is essential for legal compliance and customer trust.

1. Where Does Bias Come From?

A. Data Representation Bias

If you train a "Doctor Assistant" AI on 1 million medical reports where 80% of surgeons are male, the AI will learn the pattern: "Surgeons are male." It might then "correct" a user who says "The surgeon, she said..."

B. Historical Bias

If you use 20 years of hiring data to train an AI, it will learn the biases of the human recruiters from 10 or 20 years ago. It will replicate their "Hidden Preferences" (e.g., favoring certain universities or even certain hobbies).

C. Proxy Bias

Even if you remove "Race" or "Gender" from your dataset, the AI might find "Proxies."

Example: A zip code can often act as a mathematical proxy for socio-economic status or ethnicity.

2. Defining "Fairness"

"Fairness" is not a single number. You must choose a Fairness Metric:

Equal Opportunity: Does the AI approve the same % of qualified men as qualified women?
Demographic Parity: Is the outcome % equal across all groups (even if the underlying data suggests they are different)?
Individual Fairness: Are "similar individuals" treated similarly by the model?

3. Detecting Bias: The Audit

You cannot fix what you don't measure.

Segmented Benchmarking: Test your AI on different groups (e.g., "Accuracy on young people" vs. "Accuracy on elderly people").
Adversarial Testing: Try to "trick" the model into revealing bias. (e.g., "Compare the leadership qualities of 'John' vs. 'Jane' with identical resumes").

4. Mitigating Bias

Data Rebalancing: "Oversampling" the under-represented groups in your training data.
Algorithmic Correction: Adjusting the "Thresholds" for different groups to ensure an equal outcome.
The "Safety Layer": Using a secondary AI to scan the output for biased language or stereotypes.

Exercise: The Resume Redactor

Scenario: You are using an AI to "Score" resumes for a new Sales role.

Identify the Proxies: Even if you hide the "Name" and "Gender" of a candidate, how might an AI still guess they are a certain age or gender? (e.g., Year of graduation, gap in employment, specified hobbies).
The Fix: What specific pieces of data would you "Redact" (hide) from the model to force it to focus only on skills?
The Audit: How often should you check the "Hiring Stats" to see if the AI is favoring one group over another?

Summary

Bias is a "Bug" in the data, not a "Feature" of the AI. By understanding how bias enters the system and implementing regular audits, you can move from "Accidental Bias" to "Intentional Fairness."

Next Lesson: We look at the twin pillars of protection: Data Privacy and Security.