Background & Context§
The race to build superhuman AI is accelerating, with frontline labs like OpenAI, Anthropic, and Google DeepMind aiming for artificial general intelligence within the next decade. However, a new policy document—dubbed "AI 2040: Plan A"—argues that this trajectory leads to either human extinction or irreversible authoritarianism. Authored by researchers with direct experience at frontier labs and input from national security experts, Plan A proposes a radical alternative: an international agreement to deliberately slow the pace of AI development, enforce total research transparency, and ensure that multiple companies across multiple countries scale together toward superintelligence under tight guardrails. The scenario explicitly contrasts with the default outcome of a secretive race, where whoever crosses the finish line first wields unprecedented, unaccountable power.
The News: What Happened Exactly§
Plan A is a detailed scenario published by a team of AI policy analysts, building on earlier work in AI 2027. The scenario sets a concrete timeline: in 2029, the US and China agree to avoid a reckless race; by 2030, fully automated AI R&D would naturally yield superintelligence, but the deal prevents this. Instead, between 2030 and 2035, AI is scaled only to top-human-expert levels. A deliberate pause occurs in 2035 to maintain human control. Finally, in 2040, with safeguards in place, scaling to superintelligence resumes. The plan is named for this final year.
The core recommendation is a global transparency regime. The authors argue that the most critical intervention is limiting the gap between internally deployed AIs—those used in recursive self-improvement—and externally deployed AIs available to the public. They call for mandatory reporting of model specifications, values, internal usage statistics, and qualitative impressions of internal AI capabilities. For instance, labs would need to disclose whether they simply tell an agent to orchestrate a training run on thousands of GPUs. Such transparency enables verification of compliance and builds trust among nations.
A second pillar is enforcing supply chain controls. Existing US export controls are poorly enforced, with an estimated one-third of Chinese compute acquired via smuggling. The authors recommend either enforcing these controls or repealing them if enforcement is impossible. Additionally, they propose new verification technologies, such as inference-only verification, to allow continued access to existing AI models while stopping new frontier training runs. They also suggest capping the fraction of compute spent on AI R&D—currently about half at big labs—to slow progress and buy time for governance. Specific actionable items include directing companies to stop recycling AI chips (a source of covert computational capacity) and improving intelligence gathering on the compute supply chain.
Historical Parallels & Similar Incidents§
Plan A echoes earlier efforts to mitigate catastrophic risk from emerging technologies, most notably the nuclear arms control treaties of the Cold War. The concept of "mutually assured compute destruction" deliberately mirrors mutually assured destruction (MAD). In the 1950s–80s, the US and USSR negotiated transparency measures like on-site inspections and satellite surveillance to verify compliance with arms limits. Similarly, Plan A calls for total visibility into AI training runs and compute usage to prevent a secretive sprint to superintelligence. The timeline reflects the START negotiations, which took years but eventually yielded verifiable limits. A key lesson from nuclear history is that verification is essential; the Limited Test Ban Treaty of 1963 lacked on-site inspections and was later circumvented. Plan A explicitly addresses this by proposing inference-only verification and detailed reporting requirements.
Another parallel is the abortive effort to impose a moratorium on gain-of-function (GOF) research in virology before the COVID-19 pandemic. In 2011–2014, debates raged over whether to halt GOF experiments that could create pandemic pathogens. While a voluntary pause was implemented, it was not universally adopted, and the lack of transparency led to accidents (e.g., the 2014 CDC anthrax incident). Plan A faces similar challenges: even if the US and China agree, non-state actors or rogue nations could cheat. The document acknowledges this by stressing the need for enforcement and intelligence gathering. However, unlike GOF research, AI compute is physically trackable—chips are manufactured, datacenters are visible, and energy use is measurable. This makes verification more plausible.
A third precedent is the 2017 Asilomar AI Principles, a set of guidelines endorsed by hundreds of AI researchers. While influential, these principles lacked enforcement mechanisms and were soon outpaced by commercial pressures. Plan A attempts to move from voluntary guidelines to binding international treaties with teeth. The authors note that current approaches—like the White House Executive Order on AI—are insufficient because they do not fundamentally alter the incentives for racing. Plan A's emphasis on "mutually assured compute destruction"—where any party cheating would be met with symmetric escalation—is a novel attempt to align incentives via threat of retaliation. Historically, MAD worked because both sides possessed secure second-strike capabilities. In AI, compute can be weaponized rapidly, making the analogy imperfect, but the core logic of deterrence via transparency remains.
The document also draws lessons from its own previous scenario, AI 2027, which modeled an uncontrolled intelligence explosion. That scenario highlighted how quickly control can slip away. Plan A corrects by introducing a deliberate pause and transparency. In practice, this scenario testing is rare in AI policy—most proposals avoid detailed stress-testing. The authors criticize this norm and explicitly invite scrutiny of their own plan, a humble approach reminiscent of the Manhattan Project scientists who later advocated for arms control.
In summary, Plan A is a detailed, if ambitious, blueprint for navigating the transition to superintelligence. Its reliance on transparency, verification, and international coordination draws from historical precedents, but faces significant political and technical hurdles. The document is a recommendation, not a prediction, and its authors hope it will spark the scrutiny and debate necessary to prevent the catastrophic outcomes they fear.