Bias and Fairness

Based on ISO 24027 (2021)

Section 1: Fundamental Concepts of Bias and Fairness

Key Definitions

• Bias: A systematic difference in the treatment of certain objects, people, or groups compared to others. Important note: bias isn’t inherently negative.
• Fairness: Treatment, behavior, or outcome that respects established facts, beliefs, and norms without unjust discrimination.

Understanding Bias Types

1. Positive Bias

   • Example: Providing extra support to disadvantaged regions
   • Can be intentional and beneficial

2. Neutral Bias

   • Example: Self-driving car recognizing mailboxes more accurately than garbage bins
   • Not inherently problematic

3. Context Dependency

   • Age discrimination case study:
     • Unfair: Rejecting qualified job candidates based on youth
     • Fair: Age restrictions for alcohol purchases
     • Cultural: Young people giving seats to elderly on public transport

Section 2: Hidden Bias and Real-World Applications

CV Screening Case Study

1. Initial Approach

   • Remove gender from classification features
   • Focus on work experience, employment duration, performance consistency

2. Hidden Problems Discovered

   • Seasonal/temporary work patterns
   • Forced entrepreneurship
   • Childbirth-related career gaps
   • Result: Unintended discrimination against women

Important Insight

• Fairness issues can exist without bias (Example: system rejecting all candidates)

Section 3: Fairness Metrics and Classification

Classification Examples

1. Simple Classification (Iris Dataset)

   • Clear ground truth available
   • Objective verification possible

2. Complex Classification (Medical School/Firefighter)

   • Multiple assessment criteria
   • Ethical considerations in defining success

Confusion Matrix Analysis

• Groups typically divided into:
  • Privileged (1)
  • Non-privileged (0)

Key Metrics:

1. Equal Opportunity: TPR₀ ≈ TPR₁
2. Equalized Odds: TPR₀ ≈ TPR₁ and FPR₀ ≈ FPR₁

Section 4: Types of Data Distortions

Selection Bias

1. Sampling Bias: Non-random sampling
2. Coverage Bias: Incomplete population coverage
3. Non-response Bias: Systematic refusal to participate

Cognitive Biases in AI

1. Confirmation Bias

   • People perceive confirming evidence more strongly
   • Creates filter bubbles in content recommendation

2. Group-based Biases

   • In-group Favoritism
   • Out-group Homogenization

3. Automation-related Biases

   • Over-reliance on automated systems
   • Selective Adherence: accepting only confirming results

Section 5: AI Transparency

Transparency Levels (IEEE P7001)

1. For Users (Levels 0-5)

   • Level 1: Basic system information
   • Level 3: Immediate explanations
   • Level 5: Continuous behavior explanation

2. For Public/Bystanders (Levels 0-5)

   • Level 1: System identification
   • Level 3: Purpose and contact information
   • Level 5: Data governance

3. For Validation/Certification (Levels 0-5)

   • Level 1: System specifications
   • Level 3: High-level design documentation
   • Level 5: Full source code and training data

Explainability Approaches

1. Inherently Transparent Solutions

   • Rule-based systems
   • Formal methods
   • Logic-based reasoning

2. Black-box Solutions

   • Post-hoc explanations
   • Saliency maps
   • Surrogate systems (LIME)