When AI goes wrong and how to fix it fast
summary
Algorithms fail in visible and subtle ways — Tay went toxic, an autonomous Uber misclassified a pedestrian, Google Photos labeled people as gorillas — so build an AI Red Team split into a public-facing incident unit and a technical QA squad that implements data versioning and lineage, long-lived inference logs, CI/CD for ML with shadow testing and blue/green rollouts, synthetic or purchased datasets for retraining, and targeted relabeling that can yield 20%+ accuracy gains while triage fixes (like blocking offensive labels) stop immediate damage.
AI-written summary · about
chapters (7)
- 0:00Start
- 1:53Examples of Nightmare Examples
- 5:54Explainable Ai
- 6:53Explainability
- 8:55Ai Red Team
- 9:21The Customer Public-Facing Team
- 26:33Integrated Tests
related talks
31mAvoid mistakes building AI products
41mWhen AI Failures Reflect Our Own Social Biases
52mSome pitfalls in AI
26m(Un)ethical Artificial Intelligence: How to Keep the AI Fair for Everyone
49mAn AI with an Agenda
42mIs your AI plotting to kill you? Why AI and Machine Learning is dangerous
43mDebugging AI
56mArtificial Intelligence? - more like Artificial Stupidity!