On Controllability of AI

Full control of advanced AI may be theoretically impossible: The paper presents formal arguments and cross-domain evidence suggesting that artificial general intelligence and superintelligence cannot be completely controlled, challenging the assumption that safety measures alone can guarantee safe AI deployment.

Uncontrollability has serious implications for humanity's future: The authors argue that if advanced AI systems cannot be fully controlled, this has significant consequences for how we should approach AI development and deployment, particularly regarding existential risks to humanity.

Current AI safety approaches may need fundamental rethinking: The paper suggests that research focused solely on controlling AI behavior may be insufficient, implying that alternative safety strategies—such as alignment, transparency, or governance frameworks—deserve greater attention and resources.

The problem spans multiple domains: By drawing supporting evidence from multiple fields, the paper indicates that AI controllability challenges are not isolated technical issues but reflect broader principles that apply across different contexts and applications.

Urgent need for proactive governance: The findings suggest that waiting until advanced AI is developed to address controllability issues may be too late, highlighting the importance of establishing safety protocols, regulatory frameworks, and international cooperation before superintelligent systems are created.