RAND proposes a strategy of flexibility for the USA in the era of superintelligence

RAND Corporation published an extensive document describing how the USA should proceed to secure a geopolitical advantage on the uncertain path to superintelligence. According to RAND, the key recommendation is a strategy called “Freedom of Action” – keeping options open rather than committing to one specific course of action, because the document describes the future development of AI as highly uncertain. The goal of this strategy is to “build and preserve the ability to secure a geopolitical advantage for the USA and ensure humanity’s survival with its agency intact” during the transition to superintelligence.

The document describes four key components of the strategy: building an ecosystem of people and AI (investment in AI safety, tools to preserve human agency, sharing the benefits of the technology, preparing society for disruption, shaping incentives in the AI ecosystem); building a security architecture for AI (tools to understand the most advanced systems at home and abroad, including visibility into computing infrastructure, technologies for verifying agreements, regulatory expertise); adapting existing security institutions to the AI era; and investing in the ability of citizens, companies and governments to respond to change.

The paper also defines seven archetypal strategies divided into three groups. The coexistence group includes dominance (the USA leads development and suppresses others), co-development (the USA leads a consortium including China with shared governance and shared computing power) and preparedness (informal coordination, voluntary commitments, phased deployment). The denial group includes a moratorium (a verifiable global halt to development above certain thresholds), deterrence (the USA halts its own development and uses instruments of national power to prevent or destroy foreign programs) and societal continuity (creating geographically dispersed, self-sufficient settlements resilient to AI threats if all preventive strategies fail). The third approach, acceleration, relies on markets and rapid iteration as a source of safety. According to the document, the choice between strategies is influenced by five uncertainties – for example, how close the danger is, whether coexistence between humans and AI is feasible at all, or whether a decisive strategic advantage can be achieved.

The author of the Import AI newsletter (Jack Clark, co-founder of Anthropic) adds his own commentary that the main message is the need for substantial government spending to prepare for advances in AI – even if only to preserve freedom of choice – and that, in his view, the current American strategy more closely resembles the acceleration model without corresponding investment in safety measures.