Crouching Harm, Hidden Bias—Addressing issues of fairness....
summary
Walks through concrete failure modes: historical, representation, annotation, measurement, selection, aggregation, optimization, interpretation biases and real examples—COMPAS unfairly flags Black defendants, Amazon's resume tool favored men, word embeddings preserve gender stereotypes, and ten public datasets show up to ~5% labeling bias. Recommends operational fixes: mandatory algorithmic impact assessments, continuous MLOps monitoring, model cards and data sheets for provenance, vendor audits, and notes ~60 countries crafting AI rules (EU drafts impose extraterritorial fines up to millions of euros).
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:35Outline
- 4:53What is Bias
- 8:47Sources of Bias
- 17:20How do we prepare
- 19:43Highlevel approaches
- 22:12Algorithmic impact assessment
- 23:32Model cards
- 24:42Tools and resources
related talks
38mSocial Implications of Bias in Machine Learning
38mSocial Implications of Bias in Machine Learning
31mFlagging *and* Fixing Bias in ML
21mFairness and bias in AI
54mEthical Machine Learning
35mBuilding Inclusive Machine Learning Models
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
51mDigital Discrimination: Cognitive Bias in Machine Learning