Getting to True Predictive Risk: Will Data Accuracy Thwart AI’s Potential?
summary
Panel argues predictive risk needs verified data not just bigger models, showing concrete failures and fixes: LLMs favor common tokens (example peak at 42) and can be poisoned by buying expired domains to inject training data—effective at about one in 10,000 poisoned points and done for $60. Practical fixes include per-application RAG applications, human sampling, provenance and SBOM-like tracking, and adoption of NIST/OWASP/ISO guidance plus policy-driven vetting and testing.
AI-written summary · about
chapters (20)
- 0:00Welcome
- 1:05Predictive risk defined
- 2:28Panel on predictive risk
- 5:19AI and Gen AI role
- 7:28AI modeling challenges
- 9:13AI standards evaluation
- 13:28Data accuracy in AI
- 15:17Addressing data accuracy challenges
- 23:28Data poisoning risks
- 24:34Real-world AI risks
- 26:21Improving data accuracy
- 27:24Human oversight importance
- 30:54AI policies for orgs
- 34:06Monday steps
- 36:35AI adoption frameworks
- 37:58AI governance frameworks
- 39:01Q&A on AI trust
- 41:49Overcoming data sharing barriers
- 42:51AI data integration challenges
- 47:05Summary
related talks
56mData, AI, Hype and Fear - Recipe for Disaster?
52mThe Risky Business, of AI Illiteracy
24mData Governance in the AI age
49mHype and Reality: How to Evaluate AI/ML in Cybersecurity
24mTechnical Tracks
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
44mAdvancing Cybersecurity with AI and Machine Learning: Opportunities and Challenges
01h 01mSecure AI: Integrating OWASP Principles in Machine Learning Model Development