Data Governance in the AI age
summary
Claims that untraining an LLM is practically impossible except by full deletion, that poisoning or “poison pill” retraining causes model degradation, and that RAG helps but is not bulletproof. Cites concrete risks: a 40 cents dark web ticket for hacks, historical scale like 5,000 data sources at PepsiCo, the need for ETL cleansing, sensitivity labels (no SSNs/CCNs in training), continuous real‑time lineage and auditing, role‑based intent controls, and use of Phoenix framework for governance.
AI-written summary · about
chapters (13)
- 0:00Data governance importance
- 1:50Old vs AI governance challenges
- 3:54AI’s data appetite
- 6:45Poison pill and model decay
- 8:50Untrainable models risks
- 10:56Five pillars of AI governance
- 13:03Real-time auditing and agentic AI
- 15:11Monitoring framework choices
- 17:01Data classification start
- 18:47Enterprise Data Foundation
- 20:32Policy lineage oversight
- 21:58Humans in the loop
- 22:59AI governance rules
related talks
31mReliable Data for Large ML Models: Principles and Practices
49mGetting to True Predictive Risk: Will Data Accuracy Thwart AI’s Potential?
44mStreamlining Your Approach to Data
36mAvoiding 5 missteps that undermine your AI readiness and success
01h 05mNavigating the Security and Privacy Landscape of Modern AI
28mSecurity of Machine Learning
09mMLOps-Native Data Governance for Generative AI Pipelines
33mMachine Learning in the age of increasing Data Privacy Consciousness