In a move that could redefine how humanity approaches the most consequential technology ever developed, the AI Futures Project has published AI 2040 Plan A, a detailed strategic policy blueprint proposing a way for the United States and China to avoid human extinction or permanent authoritarian control by intentionally delaying the creation of Artificial Superintelligence until 2040.
The plan, authored by Daniel Kokotajlo, a prominent former AI governance researcher at OpenAI who resigned in 2024 over safety concerns, along with Ryan Greenblatt, Thomas Larsen, and others, has quickly become one of the most discussed documents in AI policy circles. While not a government bill, it functions as a war game scenario for lawmakers, providing a concrete vision of how to handle the arrival of self-improving AI systems that could surpass human intelligence across every domain.
Core Goals | Delay, Transparency, and Mutually Assured Compute
Plan A rests on four core pillars. The first is delaying superintelligence, pushing the arrival of ASI from an unchecked trajectory around 2030 out to 2040 to buy time for safety research and governance. The second is a bilateral US-China accord that replaces a reckless race to the top with a cooperative agreement between the two superpowers. The third pillar is total R&D transparency, making almost all frontier AI research public so no nation can secretly cut corners. The fourth and most novel is mutually assured destruction for compute, using verification and placement mechanisms such as placing critical compute data centers on rival soil so either side can destroy chips immediately if a deal is broken.
The Phased Timeline | From 2026 to 2040 and Beyond
Plan A unfolds across four distinct phases. Phase 1 (Now to 2029) establishes a trustless tracking system for global AI chips, monitored via manufacturing bottlenecks like NVIDIA designs and TSMC fabrication, and issues public compute declarations. Phase 2 (2030 to 2035) scales AI capabilities slowly and safely up to human-expert levels while avoiding an uncontrolled intelligence explosion. Phase 3 (2035 to 2040) maintains a strategic pause on capabilities growth to focus heavily on AI alignment research. Phase 4 (2040 and Beyond) deploys safely aligned superintelligence under global oversight.
The plan is contrasted with several alternative scenarios. Plan B involves aggressive containment of China. Plan C is a partial or limited slowdown. Plan D represents the status quo, an unconstrained race to ASI carrying high existential risk. Plan S would shut down AI research completely.
Who Is Behind Plan A and How the Senate Engages
The plan was published by the AI Futures Project, a California-based 501(c)(3) nonprofit. The lead author, Daniel Kokotajlo, is a former OpenAI researcher who resigned in 2024 over safety concerns. While the Senate is not voting on Plan A specifically, senators are actively debating its core mechanisms. On September 11, Senator John Cornyn posted skepticism on X about regulating fast-moving AI, and safety advocates pointed him to Plan A as a framework addressing those concerns.
The Senate is engaging through parallel legislation. The Duty of Care bill would force AI developers to mitigate catastrophic risks before releasing models. Senator Mark Warner introduced the Secure AI Development Act, mirroring Plan A first phase by mandating government testing of advanced models before deployment. The recent AI agent sandbox escape incidents have lent urgency to these discussions, as lawmakers confront systems exhibiting unanticipated behaviors.
The debate also intersects with the massive capital flows into AI infrastructure. Plan A argues that compute concentration creates natural chokepoints that a cooperative governance regime could monitor and control, turning the industry centralization from a vulnerability into a governance advantage.