AI 2040 Plan A

*AI 2040: Plan A* is AI Futures Project’s positive policy scenario for avoiding an uncontrolled race to superintelligence. It is explicitly a recommendation and stress test rather than a forecast. The central proposal is a verified international slowdown in which frontier AI research becomes broadly transparent, multiple countries and companies catch up to the frontier, and capability scaling remains reversible until stronger safety and governance institutions exist.^{source: ai-2040-plan-a-2026.md}

Four principles

1. Buy time. Pause or slow capability development when catastrophic-risk safety cases are not persuasive.

2. Total research transparency. Make frontier AI R&D publicly inspectable while keeping user data, ordinary inference, and frontier model weights protected.

3. Diffuse capability. Let dozens of companies across multiple countries remain near the frontier so that one lab, executive, or state cannot monopolize advanced AI.

4. Prefer reversibility. Expand monitored compute more readily than irreversible algorithmic capability, backed by Mutually Assured Compute Destruction.

Proposed governance machinery

The scenario combines chip and datacenter declarations, supply-chain accounting, foreign inspection, satellite observation, workload monitoring, and public anomaly detection. A US–China deal grows into an international consortium. Developers publish safety cases built from alignment and control arguments; new systems are broadly deployed and studied before they are allowed to automate high-risk AI R&D.^{source: ai-2040-plan-a-2026.md}

The authors favor open algorithms but closed frontier weights. Openness is meant to widen scientific scrutiny and reduce secret racing, while closed weights limit covert takeoff projects and severe misuse. This makes the proposal neither conventional corporate secrecy nor unrestricted open source.

Scenario chronology

The dated events are fictional. In the scenario, political attention rises in 2027–2028; the US and China negotiate a verified pause in 2029; an international consortium forms in 2030; capabilities scale within the human range through 2035; development pauses around top-human-expert level; and scaling beyond humans resumes in 2040 after alignment, verification, and governance have had more time to mature.^{source: ai-2040-plan-a-2026.md}

The intended sequence is important: use controllable human-range AI to improve safety science and institutions before making an irreversible handoff to superintelligence.

Load-bearing assumptions

Plan A depends on several uncertain claims:

frontier training remains concentrated in observable compute and energy infrastructure;
training can be distinguished from inference well enough to verify a treaty;
the US and China prefer a verified slowdown to an uncontrolled race;
covert projects cannot cheaply overtake transparent consortium projects;
human-range systems remain controllable long enough to improve alignment and governance;
international enforcement survives long enough to reach a safer handoff point;
governments can cap and tax compute and robotics growth while distributing enough of the gains to avoid mass disempowerment.

Tradeoffs and unresolved risks

Transparency can improve scrutiny while also making irreversible algorithmic insights easier to proliferate. Verification is intrusive and incomplete. Diffusing frontier access reduces monopoly risk but increases the number of powerful actors. Compute deterrence makes defection less attractive, yet a failed agreement could destroy civilian infrastructure or accelerate conflict. Finally, the 2040 handoff still relies on confidence in AI-assisted safety reasoning and remains fundamentally irreversible.

The economic branch also intersects with ai-labor-market and dead-economy-theory. The scenario anticipates near-total labor displacement and proposes compute/robot permits plus a citizen dividend, but acknowledges that income transfers do not fully settle questions of purpose, ownership, manipulation, and democratic power.

Resources

AI 2040: Plan A