When AI Failures Reflect Our Own Social Biases
summary
Examines real-world AI failures (Microsoft Tay’s antisemitic outputs, Google image tagging Black people as gorillas, Amazon’s gender-biased resume model, COMPAS fairness tradeoffs, and removal of ~600,000 problematic ImageNet labels) and concrete fixes. Covers data-level interventions (preprocessing, fairness metrics, benchmarks), model tools (LIME for local explanations, GANs for augmentation and their tendency to amplify underrepresentation), adversarial attacks/transferability, and prototype/critique visualizations.
AI-written summary · about
chapters (8)
- 0:00AI bias introduction
- 3:32Algorithmic bias defined
- 11:51Training data bias
- 17:37Bias detection tools
- 25:23Model training bias
- 31:25Bias in production
- 37:24Adversarial limits
- 39:30Collective responsibility
related talks
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
47mEthics & AI: Identifying & preventing bias in predictive models
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
37mThe AI 'Black Box'
49mAn AI with an Agenda
52mSome pitfalls in AI
38mSocial Implications of Bias in Machine Learning
41mMeasures and Mismeasures of Algorithmic Fairness