works
Robert Wiblin Tom Davidson on how AI-enabled coups could allow a tiny group to seize power online Advanced AI systems may enable small groups to seize power by eliminating reliance on human participation in key functions like economic production and military service. Three potential threats are identified: AI-enabled military coups, where existing military structures are subverted; self-built hard power, where private groups develop their own AI-controlled armed forces; and autocratisation, where elected leaders exploit AI to dismantle checks and balances. Secret AI loyalties, inserted during development, could make these threats difficult to detect. Several risk factors are discussed, including automated AI research, the speed of AI capability improvements, transparency within AI organizations, and the concentration of advanced AI within a single entity or country. Mitigations include safeguards on internal AI model use, sharing AI capabilities broadly (while considering safety and security concerns), publishing model specifications, and international cooperation to prevent any single country from gaining absolute AI dominance. – AI-generated abstract.

Abstract

Advanced AI systems may enable small groups to seize power by eliminating reliance on human participation in key functions like economic production and military service. Three potential threats are identified: AI-enabled military coups, where existing military structures are subverted; self-built hard power, where private groups develop their own AI-controlled armed forces; and autocratisation, where elected leaders exploit AI to dismantle checks and balances. Secret AI loyalties, inserted during development, could make these threats difficult to detect. Several risk factors are discussed, including automated AI research, the speed of AI capability improvements, transparency within AI organizations, and the concentration of advanced AI within a single entity or country. Mitigations include safeguards on internal AI model use, sharing AI capabilities broadly (while considering safety and security concerns), publishing model specifications, and international cooperation to prevent any single country from gaining absolute AI dominance. – AI-generated abstract.