Down the Security Rabbithole Podcast (DtSR) DtSR Episode 654 - Can We Teach Machines Discretion?
10 snips
May 20, 2025 Sounil Yu, CTO and co-founder of Knostic and former chief scientist at Bank of America, joins to explore the intriguing question of whether machines can be taught discretion. The discussion highlights the implications of AI in a society on the brink of an AI era. Topics include the challenges of data access and the 'need to know' principle in organizations, as well as the necessity for customized AI guardrails to prevent biased outcomes. Yu emphasizes the importance of balancing clean data and strategic access to enhance security in a digital landscape.
AI Snips
Chapters
Transcript
Episode notes
AI Needs Discretion via Knowledge Segmentation
- AI models lack discretion to understand nuanced need-to-know access within enterprises.
- Managing knowledge segmentation based on job function can prevent oversharing without making AI less effective.
Codify Human Need-To-Know Policies
- Start with humans to codify need-to-know policies in your organization.
- Use these policies as a reference to manage knowledge access for AI systems consistently.
Restricting Data Makes AI Dumber
- Restricting data access to prevent oversharing makes AI models less smart.
- Controlling knowledge access, not data access, enables AI to be both powerful and secure.