RAND urges US to keep options open on superintelligence

RAND's paper arrives amid growing debate over whether the US is pursuing the right approach to advanced AI. The paper's author argues that current US strategy largely resembles the 'acceleration' archetype — one of seven the report identifies — which he compares to focusing entirely on making a car go faster while ignoring seatbelts, headlights, or brakes. The paper lays out a detailed framework of uncertainties that should shape strategy, from how close dangerous AI capabilities are to whether humans and AI systems can find a stable equilibrium.

The report's central recommendation is not to pick one strategy but to invest in preserving options across all of them. It lays out four ingredients for this approach: building a human-AI ecosystem that includes safety investment and tools for human agency; creating an AI-security architecture with compute visibility and verification technology; adapting legacy national security institutions; and building citizen, firm, and government capacity to respond. The paper also identifies seven archetypal strategies grouped into coexistence, denial, and acceleration families.

Whether any of this translates into actual policy hinges on political calculations the paper acknowledges are outside its scope. The five uncertainties it describes — danger proximity, coexistence feasibility, restraint feasibility, decisive strategic advantage, and suppression feasibility — are described as open questions rather than settled ones. The paper's value may lie less in its prescriptions and more in giving decisionmakers a shared vocabulary for debating a problem where the report's own author admits it is not clear exactly what is optimal.