The AI 'Black Box'
summary
Shows concrete failures and fixes: Google Photos mislabels Black people as gorillas (2015), Joy Buolamwini’s Gender Shades finds much higher error rates for darker-skinned women, a hate-speech detector produces 46% false positives on African-American language, and COMPAS scores overpredict risk for Black defendants. Proposes remedies: build diverse datasets, use hybrid or white‑box models, apply post‑hoc methods (pixel perturbation) to inspect decisions, and adopt toolkits like IBM’s AI Explainability 360. Notes workforce gap with only 22% of AI professionals female.
AI-written summary · about
chapters (22)
- 0:00Introduction and speaker background
- 1:28AI decisions affecting our lives
- 3:13Bias in AI algorithms and social impact
- 7:35Focus on AI ethics over hype
- 8:21Real-world cases of AI bias
- 12:41Hate speech detection and racial bias
- 13:26Predictive policing algorithms unfairness
- 15:17Diversity in datasets for fair AI
- 16:02Gender gap in AI and computer science
- 18:11Understanding human bias in AI
- 19:15Democratizing AI for broader participation
- 20:17Black box models in machine learning
- 22:03Governments distrust in AI systems
- 24:13Explainable AI and transparency
- 25:40Concept of explainable artificial intelligence
- 27:46White box vs black box models
- 28:31Approaches to explainable AI
- 30:41Post-hoc explainability techniques
- 32:10IBM's AI Explainability 360 toolkit
- 32:52Open data and AI in Brazil
- 34:17AI for transparency in public spending
- 35:24Call to action for ethical AI development
related talks
41mMeasures and Mismeasures of Algorithmic Fairness
47mEthics & AI: Identifying & preventing bias in predictive models
41mWhen AI Failures Reflect Our Own Social Biases
38mSocial Implications of Bias in Machine Learning
56mGlobal Bias in AI: When Western Data Shapes the World
45mBias in Artificial Intelligence
28mEthical AI
21mFairness and bias in AI