Ethics & Bias
Responsible AI Development
26 SLIDES Part 4: Applications
The AI That Rejected All Women: Amazon's hiring AI (2014-2018) learned from 10 years of mostly male engineer resumes. Result: "women's chess club" on a resume meant -5 stars automatically.
Prerequisites
- Understanding of how ML models learn from data
- Week 6: Pre-trained models and their training data
- Awareness of societal impact of AI systems
Overview
Build responsible NLP systems. Detect and mitigate bias, ensure fairness.
Learning Objectives
- Identify sources of bias in NLP systems (data, model, deployment)
- Apply bias detection methods (WEAT, demographic parity)
- Understand fairness metrics and their trade-offs
- Evaluate responsible AI frameworks and guidelines
- Design mitigation strategies for biased systems
Key Topics
Bias detection
Fairness metrics
Debiasing methods
Responsible AI
Key Concepts
Training data biasHistorical biases encoded in text corpora
Representation biasUnderrepresentation of minority groups
WEATWord Embedding Association Test for measuring bias
Demographic parityEqual outcomes across groups
Equalized oddsEqual error rates across groups
Model cardsDocumentation for responsible AI deployment
Key Visualizations
Ai Ethics Landscape
Bias Sources Flowchart
Fairness Metrics Comparison
Ethics