CXOTalk Top Data Scientists Explain Bad Data, Poisoned Datasets, and Other AI Killers | CXOTalk #896
62 snips
Oct 9, 2025 Join Dr. David Bray, a tech policy expert at the Stimson Center, and Dr. Anthony Scriffignano, a data science leader, as they dive into the hidden threats of bad data and poisoned datasets in AI. They discuss the Five Ms framework for identifying AI failures and why organizations often rush into AI adoption without proper vetting. Learn about the risks of generative AI, the importance of critical thinking and ethical oversight, and how to recognize malicious data campaigns that can undermine your AI systems.
AI Snips
Chapters
Books
Transcript
Episode notes
Data Truth Has A Lifespan
- Truth in data decays and math doesn't care about temporal validity, so models can regress on stale facts.
- Organizations must ask hard provenance and agency questions before allowing teams to build AI tools.
Smaller Specialized Models Over Mega LLMs
- The future likely favors many smaller specialized models communicating rather than one mega-LLM doing everything.
- Active inference and agentic systems can model continuous environments and coordinate domain-specific intelligence.
Nation-State Data Poisoning Example
- David described a reported Russian campaign to teach LLMs falsehoods across free societies about history and events.
- He emphasized that poisoned training signals are hard to undo because models don't forget easily.
