A Strategy for Preserving Options on the Path to Advanced AI

Nathan Labenz speaks with RAND researchers Joel Predd, Ben Boudreaux and Matt Chessen about their proposed US strategy for an uncertain path to advanced AI. Their claims are a policy framework, not established facts about future superintelligence.

The guests say superintelligence could have major economic, security and societal consequences. They also stress uncertainty about timing, capabilities and whether competing governments can cooperate.

Chessen says their central recommendation is to preserve freedom of action: avoid committing so early to one approach that other options become unavailable before evidence improves.

The authors compare strategic archetypes including dominance, joint development, preparedness, moratoriums and deterrence. They argue each has potential benefits and failure modes, including risks of power concentration or intrusive enforcement.

Their framework asks whether danger is close, coexistence is feasible, restraint is feasible, decisive advantage is attainable, and suppression is feasible. Different assumptions lead to different preferred strategies.

Predd argues that decision makers disagree not only about the answers but also about the amount of uncertainty. Monitoring how beliefs change can help reveal sources of instability.

Boudreaux emphasizes that the questions are not entirely outside human control: policy choices might affect the feasibility of restraint, coexistence or other options.

Chessen says researchers should explain evidence about possible danger to policymakers who have many responsibilities and may not be technical specialists. He presents proximity of danger as an important unresolved question.

Labenz questions whether a decisive strategic advantage is achievable and raises concerns about restrictions that could harden rivalry. The discussion presents competing judgments rather than a settled answer.

Predd says an advantage must be defined: a continuing technical lead is different from preventing rivals from reaching a particular threshold. The scope and durability of either are uncertain.

Boudreaux notes that advantage might differ across economic, military and technological domains, and that the speed of AI development could change the strategic window.

Chessen discusses cooperation and moratoriums, noting that both depend on whether states agree to constraints and comply with them. He describes this as a challenge in a divided international environment.

The group examines analogies with nuclear history while recognizing that AI has different development, commercial and diffusion characteristics. Analogies can inform choices without determining them.

Chessen describes a coercive suppression scenario and its potential failure mode: confrontation could harden into a prolonged standoff. This is an analysis of a scenario, not a recommendation to use force.

Predd divides the proposed work into investments to build capabilities and operational choices to be made over different time horizons.

Boudreaux calls for alignment research, evaluations and containment, while keeping people and institutions involved in consequential decisions. These are proposed safeguards, not proven guarantees.

He proposes evaluating human agency in actual deployments: an AI system might affect decisions by shaping an agenda, persuading people or excluding affected participants.

Predd argues that existing security institutions would also need to adapt to AI-related challenges, including cyber, biological and conventional security concerns.

Chessen describes an AI security architecture as technical and institutional capabilities developed by both government and the private sector. The aim is to support several possible strategies as evidence changes.

The host asks about cost. Predd acknowledges that the proposal does not yet have a detailed price tag and says the researchers have not completed budget estimates.

Boudreaux offers a decision criterion beyond monetary cost: whether a policy closes off future choices, and whether that loss is reversible in time.

They discuss signaling and deterrence alongside cooperation, including how one side’s actions can be interpreted by rivals. The host questions whether restrictions can strengthen a rival’s resolve.

The discussion returns to human agency: people may want institutions and spaces where they can make decisions and interact without automatically delegating everything to AI.

On red lines, the participants consider which high-stakes decisions should retain human oversight and how deployment rules might be established before capabilities arrive.

For indicators, they suggest tracking model capabilities, the pace of improvement, safety incidents and whether evaluations and security measures keep up. These indicators would inform decisions rather than remove uncertainty.