Robust AI Security and Alignment: A Sisyphean Endeavor?

Abstract

This article establishes information-theoretic limitations for robustness of artificial intelligence (AI) security and alignment. Knowing these limitations and preparing for their challenges is essential for responsible adoption of AI. Broader implications for cognitive reasoning limitations of AI systems are also proven.