An AI with an Agenda
summary
Seven concrete failure modes for deployed ML: phantoms of false correlation (finds spurious links), biased sample data (historical mortgages, Amazon recruiting failed to debias), overly simplistic maximization (pricing bots collude upward), simulation surprises (robot gripper hacks, HalfCheetah gait), lack of contextual understanding (Delphi yields racist/sexist judgments), representation disparity, and the feedback devil that amplifies policing or product gaps. Mitigations include scrubbing or excluding sensitive columns, adversarial de‑biasing, IBM AI Fairness Toolkit and What‑If, distributionally robust optimization, representative sampling, and control‑engineering of feedback loops.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:30Why does this matter
- 3:01Consequences
- 7:23Correlation
- 8:35Conditional Probability
- 9:35Data vs Sample Data
- 14:26Bias
- 16:59What can you do
- 23:24Pricing algorithms
- 28:05Shadow of understanding
- 35:44What can we do
- 39:55Distributionally Robust Optimization
- 43:58Feedback
related talks
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
41mWhen AI Failures Reflect Our Own Social Biases
52mSome pitfalls in AI
29mResponsible AI Challenges and Recommendations
30mHow to Lie with AI: Understanding Bias, Ethics, and the Hidden Risks in ML
50mEthical Machine Learning
26mHow to develop useful and unbiased AI