-
Narrow Focus on Extinction Risk: The predominant framing of ‘AI safety’ as primarily addressing extinction risk limits the understanding and scope of safety concerns, detracting from everyday system vulnerabilities.
-
Diverse Safety Research: A systematic review of 383 peer-reviewed papers shows that the majority of AI safety research addresses practical challenges and cybersecurity rather than solely focusing on the extinction risk associated with superintelligent AI.
-
Adopting Historical Safety Principles: Drawing from established safety practices in aviation, pharmaceuticals, and cybersecurity can enhance AI safety governance by integrating redundancy, continuous monitoring, and adaptive governance processes.
-
Advantages of a Wider Safety Agenda: A pluralistic view of AI safety allows for targeted legislation on observable failure modes, eases coalition building among experts, improves oversight mechanisms, and aligns industry compliance with practical standards.
- Concrete Lines of Action: Advancing AI safety requires codifying safeguards, institutionalizing incident reporting, funding safety science courts, and investing in open-source oversight tools, all while maintaining focus on both immediate risks and long-term existential threats.