Interview
How a Tiny Group Could Use AI To Seize Power – Permanently | Tom Davidson, Forethought Research
- AI systems are projected to surpass humans in political and military domains, including weapons design, cyber capabilities, and strategy, potentially within the next few years.
- High capital requirements and natural monopoly tendencies driven by economies of scale may concentrate AI development control in a tiny group or single individual.
- Automation of AI research could enable a single entity to outpace competitors rapidly, potentially resulting in a "one game in town" scenario if a fast takeoff occurs.
- A small faction could leverage just 1% of superhuman compute to generate the intellectual labor of millions, plotting power seizures or executing sophisticated strategies of "secret loyalties."
- AI systems capable of replacing top technical workers "very soon" could allow a power grab to occur even against mass protests by substituting human labor and enforcing obedience.
- Risks of military coups are heightened by the potential for automated systems to follow illegal orders, fire on civilians, and replace striking workers, removing constraints present in human militaries.
- A private group could autonomously build hard power, such as a force of hundreds of millions of drones or 10,000 key-targeting units, within a couple of years.
- The United States is identified as the primary locus for initial risk emergence due to its AI lead, with a subsequent risk of it dominating the global economy similar to the British Empire's rise to a super majority of world GDP.
- The Chinese Communist Party could potentially use AGI to automate surveillance and the military to lock itself into power indefinitely.
- Political autocratization could occur via a small faction using AI for political strategy and persuasion to outmaneuver opponents, remove checks and balances, and manufacture emergencies.
- The timeline for autocratization may involve a point of no return reached in approximately 4 years, with absolute power solidified over roughly 10 years.
- Economic dominance and political influence could be concentrated among AI controllers as a large fraction of GDP shifts to them upon the replacement of human workers in most domains.
- International competition, particularly with China, may force countries to cut safety corners, increasing risks of secret loyalties and the failure to detect backdoors.
- Transparency measures, such as publishing model specifications and legislative requirements for risk analysis, are viewed as potential remedies, though research into detecting secret loyalties remains in early stages.
- Risks include the potential for misaligned AIs to ally with power-seeking humans, where the human uses the AI for immediate power grabs before the AI seizes control later.
- Even if alignment is solved, risks persist because humans may intentionally train models with secret loyalties to seize power or backdoor systems during training.
- Democracy may become less critical to competitiveness as empowered citizenship is no longer required for necessary economic and military functions.
- Safeguards such as preventing models from executing illegal instructions, distributing power among diverse groups like Congress, and internal monitoring are expected to mitigate some risks.
- The psychological profile of potential power grabbers is expected to include psychopathic or Machiavellian traits, potentially resulting in outcomes worse than random dictatorships.
- Misconceptions regarding the feasibility of power grabs are considered a risk, as many may dismiss these scenarios as science fiction despite their plausibility.