AI 2040: A Deliberate Delay to Superintelligence as an Alternative to the Race

The AI Futures Project—the group behind the widely-circulated AI 2027 forecast—has published ‘Plan A,’ a scenario laying out how humanity might reach superintelligence without triggering extinction or an extreme concentration of power. Rather than a prediction, it’s framed as a policy recommendation stress-tested through a concrete narrative. The core proposal is an international agreement, anchored by a 2029 US-China deal, that trades secrecy and speed for full transparency in AI R&D. With all frontier research made public and dozens of firms across multiple countries able to reach the frontier, nations could verify each other’s progress and enforce shared guardrails—a regime the authors liken to mutually assured destruction applied to compute.

The timeline deliberately slows things down: absent intervention, the authors expect fully automated AI R&D by 2030 and superintelligence by year’s end. Under Plan A, developers instead scale only within the human range through 2035, pause at roughly top-human-expert capability to preserve human control, and unpause to build superintelligence in 2040. The authors argue the current industry approach—assuming alignment can be solved ‘on the fly’ while the strongest labs pull up the ladder behind them—carries an unacceptably high risk of either losing control of superhuman systems or handing a tiny group unprecedented power over an army of superintelligences.

What distinguishes the piece is its methodology. The authors contend most AI policy proposals collapse under ‘scenario scrutiny’—the discipline of writing out a detailed, plausible story in which the plan actually succeeds—and note this exercise is rare because it exposes uncomfortable weaknesses. By applying it to their own recommendation first, and contrasting Plan A against four alternative national responses (B, C, D, and S), they aim to raise the standard of debate beyond vague optimism that hard choices can be deferred.